« iepriekšējāTurpināt »
This problem was anticipated some now 60 engineers developing or sup- (1) Better integration of databases. years ago and for the last 1.5 years the porting such software. SDRC's solid Right now many programs are in system I was shown has been in use. It modeler is the front end of many of islands, requiring painful data is similar to commercial software avail- these studies, but Toyota has its own conversion. This is especially true able in the United States from at least pre/post-processor called CADETT. of FEM, where Toyota hopes some two companies. The programmer moves (Everyone I visit says it takes too long Al methods might be used to speed the robot's tip with the mouse or by to prepare data for FEM analyses. Even up meshing. indicating target points on the car body those that have “automatic” meshers and indicating offset distances from say so.) Two Crays and two Fujitsu (2) New ways to design so that the the targets to the tip. After a path pat- vector processors support this work.
percentage of automation of final tern has been programmed, it can be NASTRAN, MARC, PAMCRASH, assembly can be increased. Curduplicated and stepped along the sur- ABAQUS, and ADAMS are used along rently it is around 5% and they face so that it is “sprayed” completely with a home-grown solid modeler called would like to see 20% or 30%. The robot's motions can then be simu- SURFES.
(VW estimates that it has had 25% lated and two checks made: range of Engine component analyses include since 1984.) One approach is to motion in each joint and interference stress in pistons and connecting rods, use more modules or preassembetween robot and car. The required temperature distributions, vibration in bled units rather than single parts. motion time can also be calculated. pipes and oilpans, and residual stress Nissan brought up the same issue. Figure 9 shows the computer screen in machined aluminum engine blocks. during programming of spraying the Simulations using in-house software (3) Expert systems to help designers underside of an open car hood. The include piston slap, torsional vibration, lay out the
man factors of car line of sight is from the outside of the oil film thickness, volumetric efficiency, interior design. These include where transparent hood to the tip of the robot and valve train dynamics. Piston tem
pedals should be located, what is a behind the hood. A simple path con- perature distribution calculations
good view angle or range to dashnecting target points at the edges of the required a fuel consumption model as board items or out the window, hood is shown.
well as effects of coolant and lube oil and how much force is needed to Curiously, even though the car body on heat dissipation.
move handles, wheels, and buttons. is represented by the CAD shape data In a video I saw very accurate rollfrom the styling system, and die faces over simulations, side by side with actual (4) Artificial reality to aid interior can be machined by NC programs that cars doing the same thing, plus behav- design. are written directly from those data, ior of active suspensions and skid conthe paint robot cannot be similarly trol systems, engine block-transmission
trol systems, engine block-transmission Concluding Remarks programmed. The demonstrator, a case torsional vibrations, and acoustics Toyota programmer, said he will of a car interior.
Toyota emphasized several points implement this obvious capability soon. All of the demos I saw were either that I also heard at other companies.
Another curious fact is that there is on video or in the software develop First, CAD planning activities and no feedback to the programmer con- ment and training facility so I could not implementation departments are a cerning whether the robot path's veloc- tell how extensively they are used.
tell how extensively they are used. permanent feature of operations. These ity profile and distance from tip to part However, it is obvious that most of departments are staffed at least in part will provide a good paint finish. Possi- these capabilities are in daily use and by former engineers. They are very bly the software picks an appropriate many have been for 3 to 5 years or conscious of the needs of users when profile once the target points are given.
they design new CAD tools: try to
recreate the designer's previous workCAD and CAE in Engine Future Needs and Plans
ing environment and methods and Design and Body Engineering
provide tools that the designers can
As mentioned above, I did not get a understand and therefore trust. These Typical applications are FEM studies comprehensive view here like I got at ingredients are essential in getting the of various types. These date from 1975 Nissan. The needs expressed here are tools accepted and used regularly. and have grown steadily until there are probably a fraction of what is on their
minds these days. These include
Figure 9. The computer screen during programming of spraying the underside of an open car hood.
he feels that conventional CAD puts is a capsule description using qualitaUNIVERSITY OF TOKYO too much emphasis on pictures. Fasci- tive terms like accelerating, plus, con
nation with the pictures obscures the nected to, and so on. 15 August 1991
fact that no real knowledge has been A moment is described in terms of
captured in the design. The challenge two objects, one of which exerts a force Background
is to find a way to capture “deep knowl- on the other from a distance; if the
edge,” by which he and other Al-based moment is applied by one object, the Prof. Tomiyama and Prof. Yoshikawa researchers describe knowledge about other rotates. Coils and magnets can share a laboratory devoted to model complex relationships among physical exert forces. Shafts connected to bearing engineering knowledge and design phenomena, for example. Deep knowl- ings can rotate. When current passes processes. While Yoshikawa is engaged edge is different from "facts," such as through a coil, it acts like a magnet. It as Vice President of the university, “every screw goes into a hole." also gets hot. Things that get too hot Tomiyama is in charge in the labora- To capture basic knowledge, can melt. An inference engine browses tory. His background is mechanical Tomiyama is using Qualitative Physics through a model constructed of such engineering, but his research approach (QP) and Qualitative Reasoning (QR). chunks and finds out if the conditions is strongly linked to artificial intelli- The goal is to represent things symbol- for rotation are satisfied: there is an gence (AI). It is an interesting combi- ically so that facts can be reasoned about. object containing a coil that can rotate, nation with many possibilities. This approach has the promise of creat- there is another object that contains a The research has two related ing “intelligent CAD,” but in doing so
ing “intelligent CAD," but in doing so magnet, there is a source of current to branches: design and maintenance. Each appears to sacrifice many aspects of
appears to sacrifice many aspects of the coil, the resulting force can exert a should have a theoretical basis, and he existing quantitative models. Tomiyama moment, etc. aims at providing the base via repre- disagrees, saying that QP augments Qualitative physics sits at the top sentations of deep knowledge about quantitative models because it can be (currently) of a hierarchy below which "how things work."
the basis for generating symbolic models. are “confluence-based (C-B) QP” and
Long term, the goal is to create a “Qualitative Simulation" (QS). QS Design
meta-model (literally a model of models) comprises a qualitative description of
that will contain the usual engineering the relationships between variables in This research has two branches itself: models as subsets. Products would also equations. Typical output consists of understanding the design process and be modeled in this way. Other conclusions like “if voltage A increases, understanding designed objects. The researchers' ideas about product models current B will decrease.” It has been design process to him means under- are too geometry oriented for him. He applied by other researchers for at least standing how people think, including would rather model where the geom- 15 years to fault diagnosis of chemical cognitive process modeling. His route etry comes from. (This means that his plants and is used in the self-maintenance to a design theory is similar to that of goal extends beyond the needs of design copier described below. C-BQP takes researchers in the United States who, to the needs of engineering in general. a topological description of a designed like him, use protocol analysis to deter. The utilitarian needs of design may not object and seeks to derive a qualitative mine what the designer is doing. (Pro- require the wheels of basic engineering equation description, upon which QS tocol analysis comprises recording, often to be reinvented every day.)
can operate. QP allegedly can generate with video, what a designer does during A demonstration system (see below) the topology, but I think the designer an experiment and then analyzing the suggested how such deep knowledge currently must input nearly all the inforrecord--the protocol.) Unlike some U.S. might be able to find out behaviors or mation from which the topology is researchers, Tomiyama tries to link the consequences of a design that the derived. This topology comprises both protocols to a computer simulation of designer had not thought of. This would the physical arrangement of objects and the process. However, he admits that be a powerful and useful capability. To that of the physical principles underlythe protocols do not really tell him begin this effort, he put five students to
begin this effort, he put five students to ing the object's behavior. anything about what the designer is work last year defining all the engi- A quantitative version ofall this has thinking or how the process actually neering knowledge chunks they could existed for about 30 years in the form of proceeds.
think of. The result was 3,000 chunks. "bond graphs." Bond graphs model the Modeling of designed objects com- These describe phenomena in hierarchi- energy flows between fairly general prises “CAD without pictures,” a term cal classes. Under “movement” one finds discrete elements like energy storers, he uses jokingly but which is quite rotational, straight, oscillating, accel- transformers, and dissipaters. These important. Like others, including myself, erating, and so on. For each one there elements can actually be electrical, thermal, fluid, or mechanical in under- From the meta-model the designer publication (Ref 19) discusses gearlying nature. The designer supplies the can elicit simulations of all effects or he driven devices as well, but the opportutopology, which consists of mapping can request "aspect" views, such as nities for functional redundancy are the energy flows. The equations can be evolution of rotational states or ther- likely to be more limited in such cases. derived algorithmically and are quanti- mal states. Behavior of the rotational The central idea is self-diagnosis tative. The origin of this method is model is started from given initial and generation of a repair strategy using dynamical analogies, discovered when conditions on pole angle and velocity. QP. A copier in the laboratory has it was shown that the differential equa- Some states permit rotation to start been augmented by a monitoring comtions for many engineering systems had and continue. The computer can deduce puter, a diagnosis computer, four senthe same form.
that cyclic state behavior exists. Other sors, and three controllers. One of these Two objections can be raised to the initial states are neutral; the magnets measures copy density. A QP model of bond graph approach. First, it cannot cannot exert any net torque and rota- the elements that affect copy density be reasoned about by existing Al sys- tion does not occur. Instead, coil heat- has been built. Degradation of the tems. Second, it cannot easily handle ing occurs. In the thermal model, states halogen lamp was simulated and the logical state changes, such as that once progress from cool to hot to melted. system proceeded to determine that a wire melts, the coil will not conduct Operation of the motor is henceforth raising the lamp voltage was a feasible electricity any more. This capability is impossible regardless of changes in the fix. important if one wants to capture the initial state.
The method requires the user to put failure modes of a design, especially The dead state from which failure in the QP model, from which the comthe unanticipated ones. In addition, ensued is symmetric, a condition that is puter can derive a QS. The computer some systems undergo logical state impossible for a three-pole motor. Dr. first reasons backwards from the changes that are not failures but are Kiriyama admitted that his system does observed sensor readings to determine merely different normal behaviors. The not know the concept of symmetry and possible causes, listed as voltages, resiscopier in our office senses the position in any case could not generate the three- tors, and so on. It then assumes that the of the input paper guides, decides which pole design following detection of the causes cannot be changed and identiway the original is oriented, and selects two-pole design's shortcoming. He said fies variables that it can manipulate, as the correct paper tray.
that QP is not good at such generaliza- well as the effect these variables would Tomiyama feels that these are not tions.
have given what is apparently broken. drawbacks but merely reflect the fact
Among the available choices, some will that bond graphs are meant only to Maintenance Theory
produce side effects in addition to the model energy flows and fail to repre
desired one. The candidate with the sent other aspects of physics that The goal here is to understand how fewest side effects is chosen (see designers must take into account. to design things so that they can fix Figure 11).
I was shown a demo of a QP system themselves, at least to a limited degree. The diagnosis computer recomfor modeling simple discrete physical There are two advantages to such prod. mends a small voltage increase for the systems--the same kind that bond graphs ucts. First, time and effort are saved. lamp, which is implemented automatmodel (Ref 18). The item modeled is a Second, the product may be able to fail ically through a controller. The test is simple one-, two-, or three-pole motor soft. That is, degraded behavior may be repeated, as is the recommended fix, comprising a shaft on bearings carrying available as a backup even if the failure until it cannot further improve the copy the coils (poles), wiring of the coils to a has eliminated top quality behavior. density. Either the process halts here commutator, and two permanent mag- Two issues are involved here: how or the diagnosis computer shifts to a nets. Facts about magnetic attraction, to design things that have backup modes second variable, the main static charger moments, rotation, state change of the and how to equip products to diagnose voltage, and tries to gain further improvecommutator, and rotary acceleration themselves and bring these backup ments. This is actually a simple axis-bycaused by torque are assembled by the modes into play. The latter has been axis search. designer into a model of the motor, a explored in a self-maintenance copier. Note that all of this could have been portion of which appears in Figure 10. Electrical variables are manipulated to accomplished with a quantitative linear The computer knows many side effects compensate for the degradation of model composed of influence coeffiof these knowledge chunks and con- electrically driven components. A cients relating controllable and observstructs its own extended meta-model.
able variables. This method is well known
and widely used in control, simulation, Multi-axis solutions can also be readily permit traceability of symptoms to causes and optimization. The objection is again found, such as simultaneously raising that are very subtle. He cites Three
. that logical state changes are not easy one voltage a lot while lowering another Mile Island and Chernobyl as examples to represent, except by altering the one a little. I do not know if QP can do of systems that failed for lack of suitcoefficient matrix. However, I do not this.
able reasoning during their design. see why that is difficult. Everything Tomiyama feels that linear algebra (However, the fact that Chernobyl-type that is called "reasoning” above can be approaches cannot handle noisy sen- reactors are inherently unstable at low accomplished using linear algebra or sory data, represent portions of a sys- power is well known and was known circuit theory implementations of it. tem that are completely broken, or
tem that are completely broken, or when it was built.)