Process control in metalworking
hasn't changed much during the past 100 years. At first
the process was a kind of brute force method: Make some
parts and measure them, make some more and measure again,
throw out the bad ones, measure an entire batch of parts
if you thought they were all suspect, and so on. Then came
Walter Shewhart and statistical process control. SPC provided
a means that aided in process monitoring and identifying
special causes of variation. Yet, even today, process control
in metalworking is still more art than science. Material
inconsistency, variation of tool geometry, temperature oscillation
on the cutting edge and the sheer physics involved when
a tool comes in contact with a work piece create a process
that is difficult to predict and control. This article discusses
some of the difficulties of process control in metalworking
and then describes how another technique, intelligent process-adaptive
control technology, can be used in discrete-part manufacturing
to control a process to achieve zero defects.
The interaction between a machine tool and the part being
worked is extremely complex and introduces variables that
are unique and problematic to this industry. First, the
process is a combination of a continuously deteriorating
tool and dynamically changing variation--random and nonrandom.
This deterioration is both unidirectional (the tool is constantly
shrinking) and nonlinear (the wear rate varies during the
life of the tool).
Typically, a tool-wear curve consists of small, steep
slopes of irregular duration within a single tool lifespan:
Cutting-edge round-off and initial flank wear are low-rate
processes, but, in the latter stages of tool wear, flank
wear and cratering might be a high-rate process. Tool wear
in an ideal metal-cutting process progresses in three phases:
round-off of cutting edge, low-rate tool wear and transition
to high-rate tool wear. The duration of each phase isn't
repeatable even in the most stable process conditions. Multiple
sources of variation on the cutting or forming edge change
the tool wear curve every time an insert is indexed or a
form tool is sharpened.
Next, most metalworking processes are designed to use
a tool for creating more than one dimension, e.g., one tool
cutting two diameters. It's difficult to avoid a displacement
between the locations of each dimension relative to specification
limits. The difference in cutting speed for associated dimensions
(the cutting speed of the larger diameter is faster than
that for the smaller diameter) leads to different tool-wear
rates on the same tool. Roundness, taper and surface finish
are all related to a tool. The objective of any process
control technology is to determine the moment when a tool
should be compensated for, or changed, in order to provide
compliance with quality requirements, including all dimensions
and tolerances created by this tool.
Predictive modeling provides an accurate and easy-to-use
method to control such processes. MICRONITE, an expert system
based on the iPACT concept and developed by High Tech Research
Inc. of Deerfield, Illinois, uses predictive modeling as
part of its intelligent process-adaptive control technology.
There are four major components of variation that iPACT
addresses. The following causes of variation identified
for each component can be found in any metalworking process:
Primary variation--Includes measurement error (e.g.,
accuracy, precision and repeatability), shape variation
(e.g., taper and roundness) and machine precision (e.g.,
piece-to-piece variation, inter-spindle and inter-fixture
variation). The extent of this variation determines whether
the tolerance is wide-open, open, close or extremely close.
As an example, a close tolerance on a multispindle screw
machine becomes an open tolerance on a CNC lathe. Quantifying
primary variation helps uncover causes and reduce variations
that force frequent and unnecessary adjustments and tool
Process-dependent variation--Reflects tool wear
and the instability of a dynamic system, including nonrandom
variation due to tool-wear trend and random variation due
to the instability of process variables.
Cumulative product variation--Depends on the alignment
of process runs. These include variation due to the displacement
of the locations of sample averages between sequential processes
and variation caused by indiscriminate process interruption
Special causes of variation as defined by SPC--Variation
related to machine, material and people
Knowledge of the dynamic nature of discrete processes
helps us understand different ways to achieve a state of
process control. Adaptability is where it all begins. An
intelligent system should be capable of controlling a multitude
of relatively stable and unstable processes at once. The
system should cope with a range of quality requirements,
from compliance-to-specification to Six Sigma. This means
that unique sampling design and automatic execution of predictive
modeling is required for every operation and characteristic.
MICRONITE uses three types of predictive models: nonlinear
trend control, control of probability of defects and control
of extent of tool wear. Because sampling time is critical,
adaptive intervals for every tool are updated after each
data entry. The system also allows predetermined levels
of product variation. Real-time control by a CpK model (without
using X-bar and R charts) guarantees a customer-required
A workstation is located near a machine, a group of machines
or a cell. Upper and lower specification limits are entered
for each dimension, as is the cycle time for each cutting
operation. Based on the tool-wear model, the number of parts
run and other parameters, the software prompts the user
to measure a unit or sample. The software then tells the
operator to continue, input a compensation adjustment into
the machine or change the tool. It will also tell the operator
how long before the next measurement must be taken. The
software adapts the model as new data is collected and analyzed.
Here's an oversimplified description of what actually
occurs. First, a process (i.e., tool wear) curve is segmented
in real time by one of iPACT's rule-based models. After
every data entry (either sample or unit), a model predicts
the risk of defects until the next inspection, the rate
of tool wear related to the location of data averages, and
sample variation. If the system determines that a risk of
defects has increased, a compensation adjustment will be
recommended; if tool-wear rate has dangerously increased,
a tool change is advised (See figure).
A trend control model will stop a process at the point
of accelerating tool wear and increased risk of defects.
A model controlling the extent of tool wear will stop a
process before severe tool deterioration occurs. Stable
processes with relatively low tool-wear rates are controlled
by separate estimates of the probability of defects on the
lower and upper specification limits. Sampling time is critical
and is adjusted dynamically after each measurement; a stable
process with little variation would require a longer sampling
interval than a process with large variation and high-rate
When a tool is indexed, sharpened or changed, you can't
expect duplication of a process curve; all tool-wear processes
are nonrepetitive. This means that all parameters, such
as sample variation, tool-wear slopes, rate of tool-wear
acceleration and a number of mini-processes, will probably
change. Therefore, a new model is needed in order to control
a new process. Whenever the user indicates a process change,
iPACT will start a new modeling cycle. Only when mature
and long-running processes repeat will iPACT use historical
data and statistical process control for process control
Predictive trending and adaptive sampling intervals ensure
that a machine is operated to its maximum capacity. Tools
achieve maximum usage before change-out, and sampling is
performed only when needed. Both of these decrease waste
and increase machine uptime and operator efficiency.
Let's look at how predictive modeling works with different
tools and process types.
An ideal condition for process control exists when a finishing
tool creates only one critical characteristic and a process
can be easily adjusted to nominal specifications. If metal
cutting were only this, SPC would be an obvious solution.
However, most metal-removal and -forming operations are
designed to create multiple characteristics with a single
finishing tool. These multiple tool-bonded processes are
divided between the following:
Type one--Single-point and simple-shape tool (e.g.,
insert, mill or reamer). Individual characteristics are
cut sequentially using the same tool. A process control
solution must be able to control for all misaligned tool-bonded
characteristics and control primary variation.
Type two--Step tool (e.g., form tool, step drill
or grinding wheel). Individual characteristics are cut simultaneously.
A process control solution must control for all misaligned
tool-bonded characteristics, different tool-wear rates and
Type one is considered a single process with multiple
outputs (i.e., one tool cutting several dimensions). Type
two is considered multiple processes with multiple outputs
(i.e., a step-wise tool generating related dimensions and
tolerances). Tool-bonded processes of type one can be controlled
by a key characteristic if misalignment and primary variation
aren't significant. It's here that SPC can be used effectively.
Typically, however, a close-tolerance metalworking operation
faces the problem of misalignment between related characteristics
and a substantial primary variation. All of these may be
key characteristics that need to be controlled. In this
case, iPACT uses multiple models to control each critical
product characteristic, one model per characteristic.
Step tools, or type two, are even more problematic, having
two or more critical characteristics with different dynamic
patterns in terms of tool-wear rates, variation and the
location of sample averages. If more than one process is
running on a complex tool-cutting edge, then iPACT uses
a multiprocess model. This forces the optimization of internal
specifications for tool-bonded characteristics and helps
develop solutions for maximum tool efficiency.
Although we've been considering the use of predictive
modeling at the part level, it's most effective when applied
across the entire plant. This is achieved when all workstations
are networked and data is shared upstream.
The basic pattern of data organization is divided into
Production plant--At this level, the software
provides an overview of all jobs and operations and alerts
Multioperation job--Here, data continuity is provided
for all operations.
Metal-cutting or -forming machines--The equipment's
capability to hold tolerance is verified by in-process capability
control and off-line studies.
Operation--Control of individual and tool-bonded
characteristics, along with roughing and finishing tools,
is aimed at increased efficiency and limited machine attendance.
A group of tool-bonded characteristics--Reducing
misalignment and variation leads to an extended time between
tool adjustments and changes.
Individual characteristics--Compliance with specification,
a sampling discipline and preventing tool failure is accomplished.
We've only briefly discussed this exciting new technology,
which may be the only current viable solution for in-process
metal-cutting and -forming control; in many instances, it
does what SPC can't. Considering that in 1998, metal-cutting
and -removal operations accounted for more than $400 billion
of domestic expenditure, billions of dollars per year could
be saved through waste reduction and increased machine and
operator efficiencies using this technology.
Stephen Birman, Ph.D., is president of High Tech Research
Inc. He has 20 years of experience developing expert-based
control systems for discrete part manufacturing. He developed
the concept of intelligent process adaptive control technology
and led a team of engineers to create MICRONITE, a commercial
implementation of the iPACT technology. Letters to the editor
regarding this article should be e-mailed to email@example.com.