Note: Descriptions are shown in the official language in which they were submitted.
-
2 1 62433
TOOL CONI)ITION MON~OI~ING SYSTl~M
BACKGROUND OF THE INVENTION
The present invention relates to a tool monitoring system for monitoring the
condition of an electric motor dnven tool pe~rorll~ing a cyclical operation.
Tool condition monitoring is one of the major concerns in modern machining
operations, especially in machining operations for mass production. Failure to
10 detect tool failure and wear leads to poor product quality and can even damage
machine tools. On the other hand, a false detection of tool failure or wear may
cause an unnecessary interruption of an entire production. Both can result in
significant monetary loss.
Known tool monitoring systems include systems for "on-line tool condition
15 monitoring." In on-line tool condition monitoring, the tool is monitored for defects
after each cut or cycle. These tool monitoring systems typically use optical sensors
or laser optical sensors which measure the geometry of the tool after each cut.
However, on-line tool condition monitoring can only detect catastrophic failure of
a tool after a cut and cannot monitor the gradual wear of a tool or predict the tool's
20 failure. Further, these systems are vulnerable to chips, coolant, and environmen~l
noises.
Other known methods for tool condition monitoring attempt to predict tool
condition based on various sensor signals such as cutting force, acoustic emission,
and vibration. However, sensors for monitoring cutting force are too expensive to
25 use with multiple stations and multiple spindles. Acoustic emission and vibration
sensors require additional wiring and are vulnerable to various noises.
2 1 62433
Some monitoring systems monitor power consumption (or motor current) of
the tool. As the tool wears (or if it fails) its power consumption changes.
However, the power signals are complicated and the power signals to provide a
reliable, accurate indication of it has proven difficult to use. The power signal does
5 contain some "noise" due to factors other than tool condition. Typically, these
systems sets a range of signal that a monitored signal should fall within. When the
monitored signal is outside this range, a worn tool or failure is indicated.
One major problem with monitoring the power consumption of the motor is
that occasional spikes are experienced in a machine tool even under normal
10 condition. The spikes can falsely indicate that the tool is worn. However, if the
threshold is increased to prevent false signals, a worn tool may go undetected.
SUMMARY OF THE INVENTION
The present invention provides a real-time tool monitoring system for
15 monitoring the condition of an electric motor driven tool performing a cyclical
operation.
In the inventive tool monitoring system, an accurate dynamic threshold is
generated by monitoring the actual power consumption of a machine tool of the
selected tool type while the machine tool performs a plurality of m~chining cycles.
20 The power consumption signal of the machine tool is decomposed into its time-
frequency components an,d reconstructed based upon certain selected components in
order to reduce the effects of noise. In the present invention it is also recognized
that the power consumption signal of a machine tool in normal condition will
2 1 62433
include a number of spikes in each machine tool cycle. Accordingly, the tool
monitoring system monitors the number of times the power consumption signal
crosses a selected threshold, rather than indicating an alarm after a single crossing
of a larger threshold.
The tool monitoring system of the present invention generally operates in two
modes: learning mode and monitoring mode. In the learning mode, the tool
monitoring system measures the power consumption signals of a certain number of
samples (say 20, 50 or 100 samples) of the selected tool type as the tool performs
the selected cyclical task. The tool is known to be a new tool or in a normal
condition. The tool monitoring system then uses a mathematical technique known
as wavelet packet transform to break the power consumption signal into components.
The system selects the components that contain the bulk of the information about the
overall signal, when using wavelet transforms, the selected components are the
"feature wavelet packets." The selected components contain sufficient information
about the original signal but not unnecessary or unwanted components such as noise.
The feature wavelet pacl~ets of the power consumption signals of the learning cycle
are then calculated.
The tool monitoring system then uses these feature wavelet packet to develop
thresholds. In one method, the system calculates the inverse wavelet packet
transform of the feature wavelet packets to reconstruct the power consumption signal
of each learning sample. A power threshold, having an upper limit and a lower
limit, is then generated based upon the reconstructed power consumption signals of
the learning cycle. The power threshold is a function of time relative to the
-3-
2 1 62433
machine tool cycle. The power threshold is not the extremes of the signal, but
rather some statistic functions of the signals. The signals of learning cycle will
cross the threshold some number of times. The tool monitoring system then counts
the number of crossings of the power threshold by the learning cycle power
S consumption signals and calculates their statistical properties.
In monitoring mode, the tool monitoring system continuously measures the
power consumption signal of the tool performing the cyclical task. The tool
monitoring system counts the number of crossings of the power threshold by the
monitoring power consumption signal and compares the number of crossings to the
10 statistical data regarding the number of crossings gathered from the learning mode.
The tool monitoring system generates an alarm when the number of crossings of the
power threshold by the power consumption signal increases to some prede~ermined
amount relative to the number of crossings experienced in the learning mode.
BRIEF DESCRIPTION OF THE DRAWINGS
The above, as well as other advantages of the present invention, will become
readily apparent to those skilled in the art from the following detailed description
of a preferred embodiment when considered in the light of the accompanying
drawings in which:
Figure 1 illustrates a tool monitoring system according to the present
invention, monitoring the'power consumption of a machine tool machining a series
of workpieces.
2 1 62~33
Figure 2 illustrates the power signal from one cycle of the machine tool as
received by the tool monitoring system of Figure 1.
Figure 3 is a flow chart of the tool monitoring system of Figure I in its
learning mode.
Figure 4 is the wavelet packet transform of the signal of Figure 2.
Figure 5 is a reconstructed power consumption signal of Figure 2,
reconstructed from the feature wavelet pacl~ets selected from Figure 4.
Figure 6 is a power threshold based upon several reconstructed power
consumption signals of the learning cycles.
Figure 7 is a flow chart of the tool monitoring system of Figure 1 in monitor
mode.
DETAILED DESCRIPTION OF THE PRE~FERRED EMBODIMENlr
Figure 1 shows a tool monitoring system 10 according to the present
invention including a current transducer 12 connected to an analog-to-digit~l
converter 14 and a CPU 16. The tool monitoring system 10 is shown moniloling
a machine tool 18 having an electric motor 20 driving a tool 22. For purposes ofillustration, the machine tool 18 is shown machining a series of workpieces 24 being
moved along a conveyor system 26. As will become apparent, the tool monitoring
system 10 of the present invention can be used with any selected tool type using an
electric motor and performing a repetitive, cyclical task.
In operation, the motor 20 and tool 22 are repeatedly loaded to cut each
worl~iece 24, and then the conveyor system 26 positions another workpiece 24 to
2 1 62433
the machine tool 18. The current transducer 12 continuously indicates the power
consumphon of the motor 20 by sending a power consumption signal to the analog-
to-digital converter 14, which converts the power consumption signal into a format
readable by the CPU 16. The analog-to-digital converter 14 sends a digital signal
S representing the amplitude of the power consumption signal at a series of current
time segments. The digitized power consumption signal is stored in the CPU 16 and
associated with its particular time segment, relative to the machine tool cycle.Figure 2 shows one cycle of the power consumption signal 28 of the machine
tool 18 of Figure 1, as received by the CPU 16. The machining operation is in the
form of a cycle starting from tool engagement and ending with tool withdrawal. At
the beginning of the cycle, the tool 22 is not engaging the workpiece 24 and thepower consumption signal 28 is at idling power 30. During the initial
engagement 32 of the tool 22 with the workpiece 24, the power consumption
signal 28 rises. When the tool 22 is fully engaged in the workpiece 24, the power
consumption signal 28 reaches full engagement consumption 34. At full
engagement 34, the power consumption signal 28 reaches a level and remains
relatively unchanged, though there are fluctuations caused by various noise, such as
cutting a hard spot in the workpiece 24. Due to this fl~lct l~tinn, it has been difficult
to use a power signal to accurately predict tool condition. A high "spike" in the
- 20 signal from an unworn tool might be sometimes interpreted as a worn tool. The
present invention overcomes this problem. After completion of m~chining the toolis withdrawn. During withdrawal 36 the power consumption signal 28 decreases
steadily and finally returns to idling power 38.
2 1 62433
The inventive tool monitoring system 10 is generally based upon the
observation that the machine tool 18 will consume more power to perform the same
work when it reaches a worn condition. As will be explained in detail below, the
tool monitoring system 10 according to the present invention generally operates in
5 two modes: a learning mode and a monitoring mode. In learning mode, the tool
monitoring system 10 preferably receives data from several sample cycles of
machine tools 18 of the selected tool type. Information related to the power
consumption during each cycle run by each machine tool 18 is stored to develop
expected signal ranges, or thresholds. Then in monitoring mode, the tool
10 monitoring system 10 compares the power consumption signal of a machine tool 18
with data gathered in the learning mode and signals an alarm when the tool
monitoring system 10 detects that the tocl 22 is worn. The determination is made
by comparing the signal to the expected learning cycle, signal ranges, or thresholds.
Since the thresholds are developed by samples, they are more accurate than prior
15 art "selected" thresholds.
Figure 3 shows a flow chart for the learning mode 40 of the tool monitoring
system 10 of Figure 1. In learning mode 40, numerous le~rning cycles of a
plurality of tool 22 of the selected tool type are run in 42. The tool 22 is selected
to be a new tool or in normal condition. The power consumption signals 28 of the
20 learning cycle are ~igiti7ed by the analog-to-digital converter 14 and stored in the
CPU 16 in 44.
The CPU 16 then selects feature components of the power consumption
signal 28 in 46. In one preferred embodiment, wavelet transforms are used to break
2 1 62433
the signal into components, as explained below. In 46, the samples of the learning
cycle are decomposed into different time-frequency components. The feature
wavelet packets are selected from the co~l~ponents to represent the main information
about the original power consumption signal 28, thereby the unwanted components
of the power consumption signal 2~, i.e. noise are filtered out from the signal.In 50, the CPU 16 reconstructs the power consumption signal 28 of each
learning cycle from the selected feature wavelet packets by the inverse of the
function used to break the original power signal into components. The reconstructed
power consumption signal 28 then contains sufficient information from the original
power consumption signal 28, but with reduced noise. Notably, while only some
of the learning cycles need be used to select the feature wavelet packets in step 46,
preferably all of the learning cycles are used to develop data at step 50. The more
cycles utilized, the more accurate the system.
In 52, the CPU 16 generates a power threshold based upon statistical data
calculated at 50 from the learning cycles. The power threshold is a function of time
over the machine tool cycle and includes an upper limit and a lower limit. the
upper and lower limits are not the extremes of the signal, but rather some statistic
function of the signal. The learning cycle signals will occasionally exceed these
thresholds.
In 54, the CPU 16 compares the power threshold to the power consumption
signals of the learning cycles. The CPU 16 col~pales each power consumption
signal to the power threshold at each time segment and counts the number of
crossings by each power consumption signal. The crossings of the lower limit of
-8-
2 ~ ~2433
the power threshold can be counted separately from the crossings of the upper limit
of the power threshold, or as a separate number.
In 56, the CPU 16 calculates the statistic properties of crossings of the power
consumption signals of the learning cycles. If the upper limit crossings are counted
5 separately from lower limit crossings, the two means would also be calculated
separately. The means and variances will be compared to the monitored signals.
Since the system compares expected numbers of crossings, rather than looking for
a single crossin~, the occurrence of a few "spikes" in a monitored signal will not
lead to a false indication that a tool is worn.
As mentioned, in the preferred embodiment, the power consumption
signal 28 is broken into its components using wavelet packet transforms. Wavelet
transform is a signal processing teehnique. The wavelet transform decomposes a
signal into various components at different time scales and frequency bands, all of
which form a surface in time-frequency plane. Both the time scale and the length
15 of the frequency band can be changed, hence the characteristics of a signal can be
magnified upon different resolutions.
In the ~ ed
20 embodiment of the present invention, the CPU 16 calculates a discrete wavelet
transform, specifically ~he wavelet packet transform. The wavelet transforrn is
defined as follows:
Where the wavelet bases is:
B
2 1 62433
Ws~t)]= J .~)lY!(t r)d~
Y~sr(t)=~
The wavelet transform can be considered as signal decomposition. It decomposes
a signal f(t) into a family of wavelet bases, and the weighting coefficients, W~f(t)],
represent the amplitudes at given location t and frequency s. The wavelet transform
is a time-frequency function which describes the information of f(t) in various time
S windows and frequency bands. It forms a three-dimensional figure against
time-frequency plane. As a result, wavelet transform is capable of capturing
non-stationary signals such as frequency variation and m~gnitllde undulation.
The properties of wavelet transforms are determined by wavelet base
functions. A number of wavelet base functions have been developed. When a
10 wavelet base function, ~( ), is specified, its family, y6,r(t), is called the wavelet
bases.
For digital signals, discrete wavelet transforms can be used. In discrete
wavelet transforms the frequency parameter, s, is taken as an integer power of two,
i.e., s=2j, j=l, 2, . . .; and the time parameter, t, is taken as a series of integer
15 k (i.e. t ~ k= 1,2, . . .); that is:
~,~(t)=2~ k), j, k=l, 2, . . .
One of the most commonly used discrete wavelet transform is binary orthogonal
wavelet transform. Let Aj[-] and Dj[-] be a pair of operators. At jth resolution,
-10-
2 1 62433
Aj[f(t)] is an approximation of the signal f(t) and Dj[f(t)] represents the information
loss, or the detail signal ~5]. It has been verified ~4]:
Aj[f(t)] =f(t)*~j(t)
Dj~f(t)] =f(t)*~j(t)
S where, ~j(t) is a smooth scaling orthogonal bases, ~j(t) is an orthogonal wavelet
bases, and "*" denotes convolution. Furthermore, ~j(t) and ~j(t) are correlated
through a pair of quadrature mirror filters h(t) and g(t) defined below:
(t)=h(t)*~j l(t)
~j(t) = g(t) *~j-l (t)
In one embodiment of the present invention, a pair of 4th-order filters are
used as defined below:
t=0 t= 1 t=2 t=3
h(t) 0.48296 0.83692 0.22414 -0.12941
g(t) 0.12941 0.22414 -0.83652 0.48296
From the above equations, the discrete binary wavelet transform is then
obtained:
Aj[f(t)] =h(t)*A~I[f(t)]
Djlf(t)] =g(t)*Aj l[f(t)~
20 or
Aotf(t)] =f(t)
Aj[f(t)] =~;"h(k-2t)Aj ,tf(t)]
Dj[f(t)~=~,,g(k-2t)A; Itf(t)]
where, t=l, 2, . . ., N, j=1, 2, . . ., J, and J=log2N.
21 62433
Since the wavelet transform is a complete representation of the signal, the
original signal f(t) can be reconstructed by means of inverse wavelet transform or
the reconstruction formula below:
Aj[f(t)] =2{~"h(k-2t)Aj+l[f(t)] +~,~g(k-2t)*Dj+,[f(t)]}
5 where, j=J-l, J-2, . . ., 1., O.
Let operators H and G be the convolution sum defined below:
H = ~,~h(k-2t)
G = ~,~g(k-2t)
Then,
Aj[f(t)]=HAj l[f(t)]
Dj[f(t)] =GAj ,[f(t)]
It is seen that the binary wavelet transforms uses H and G only on the
approximation Aj l[f(t)] but not on the detail signal Dj ,tf(t)]. If the operators H and
G are applied on both Aj l[f(t)] and Dj ,[f(t)], then, it results in the wavelet packet
15 transform. The wavelet packet transformation can be computed by the recursive
algorithm below:
Po(t)=~t)
P2~ (t)=HPj il(t)
P2ji=GPj_l(t)
where
21 62433
P~t)
is the ith packet on the jth resolution, t- 1,2, . . . ,2~ 1,2, . . .2j, j = 1, 2, . . .
J, J=log2N.
The original signal f(t) can be reconstructed by the inverse wavelet transform
below:
PJ~t) 2t~P-~.+l (t) GP~.+l(t)]
where, j=J-l, J-2, . . ., 1, 0; i=2j, 2j-~, . . ., 2, 1, and the operators H and G are
the conjugate of H and G:
H = ~kh(t-2k)
G = ~g(t-2k)
Returning to the learning cycle as shown in Figure 3, at step 46, the CPU
16 calculates all of the wavelet packets to select the feature wavelet packages. As
will be shown wavelet packet transforms result in breaking a complex signal into a
number of components, with only a few of the components carrying the bulk of the
signal information.
Figure 4 shows the wavelet transform 60 of the power consumption signal 28
of Figure 2. As can be seen from the figure, the wavelet transform 60 calculated
to a first resolution 62 is decomposed into a first packet 64 and a second packet 66.
The first packet includes~most of the information, while the second packet generally
includes background noise. At a second resolution 68, the wavelet transform 60
comprises four packets. At each resolution, the wavelet packets contain the
2 1 62433
complete information for the original signal. In one embodiment, the CPU 16
calculates the wavelet transform 60 to at least its fifth resolution 70, which
comprises 32 packets. From this calculation, one can determine that at the fifth
resolution, the first packet, Psl(t) represents the trapezoid trend and the 11th packet,
5 P5ll(t) represents the dynamic wave. These two wavelet packets may then be
selected as the feature wavelet packets. The feature wavelet packets can be selected
by taking the wavelet packets containing the most inforrnation, or the most energy.
Increasing the number of wavelet packets selected increases the accuracy to which
the components represent the original signal, however, the more packets that are
10 sele~ted the more complicated the calculation becomes. The number of wavelet
packets can be increased until the desired accuracy is achieved.
In step 50, the CPU 16 performs the inverse wavelet packet transform on the
feature wavelet packets, while setting the other packets to zero. Setting the other
packets to zero elimin~t~s noise from the signal. The reconstructed signal 76 of one
lS of the learning cycles, shown in Figure 5, therefore comprises the principal
components of the power consumption signal 28, without the unwanted components
such as various noises.
In step 52, the CPU 16 generates a power threshold 78, as shown in Figure
6, which is based upon statistical plOp~ ies from the reconstructed power
20 consumption signals 76 from the learning cycles. The mean and standard deviation
of the reconstructed power signals 76 are calculated at each time segment relative
to the machine tool cycle. Therefore, both the mean power consumption signal and
the standard deviation are functions of time over one machine tool cycle. The
-14-
2 1 62433
power threshold includes an upper limit 80 and a lower limit 82, which are both
functions of time over the machine tool cycle. The thresholds are selected to be
some function of the mean and standard deviation of the learning cycles signals. In
this embodiment upper limit 80 and lower limit 82 are preferably calculated as plus
5 and minus a certain number of standard deviations from the mean of the
reconstructed power consumption signals 76 of the learning cycles.
In step 54, the CPU 16 compares the power consumption signal 76 from the
learning cycles with the power threshold 78. One of the power consumption signals
76 from the learning cycles is shown in Figure 6. In practice, the power
10 consumption signals from numerous learning cycles would be compared with the
power threshold 78. Notably, the signal in Figure 6 is shown crossing the
thresholds. The CPU 16 counts the number of crossings by the power consumption
signals 28 of the learning cycles of the upper limit 80 and lower limit 82 of the
power threshold 78. The CPU 16 then de~ermines the mean number of crossings
15 by the learning cycle signals in step 56. The upper limit 80 crossings can be
counted separately from the lower limit crossings 82, or a composite number.
After creating a power threshold 78 and statistical data from crossings of the
power threshold in learning mode 40, the tool monitoring system 10 enters the
monitoring mode 84, shown in Figure 7. In monitor mode 84, the tool monitoring
20 system 10 is again connected to a machine tool 18 of the selected tool type as shown
in Figure 1. Preferably, the same CPU 16 is used in both the learning mode 40 and
monitor mode 84, however, it is recognized that the power threshold 78 data could
be downloaded to a different l~PU for the monitor mode. Moreover, it is preferred
-15-
2 1 62433
that the learning mode be performed at the actual work station where the CPU will
be monitoring. Using the actual workstation for the learning mode will insure that
any individual characteristics of the motor, tool mounts, etc. will be accounted for
in the thresholds.
In monitor mode 84, as shown in Figure 7, the CPU 16 keeps a counter,
"Crossings", which is initially set to zero in step 86. The CPU 16 monitors the
power consumption signal 28 of the machine tool 18 while the machine tool 18
performs its repetitive cyclical machining operations in 88.
In step 91, the CPU 16 compares the power consumption signal of the
machine tool 18 to a catastrophic threshold. The catastrophic threshold is chosen
to be much larger than the power threshold 78. It can be based upon the data
gathered in the learning mode 40 or can be determined beforehand. The
catastrophic threshold is selected to be so high, as to only be met when there is a
severe failure. If the power consumption signal 28 crosses the catastrophic
threshold at any time, the CPU 16 signals an alarm in step 92 which immediately
ceases the machining operation and disengages the motor 20 and spindle 22.
In step 94, the CPU 16 monitors whether the power consumption signal 28
fails to rise above the idling power 30 in the time segments of the cycle
corresponding to the initial engagement of the tool with the workpiece 24. This
indicates that the workpiece 24 is missing or that the tool 22 is broken. If so, the
CPU 16 signals an alar~ 92. This threshold can be based upon the data gathered
in the learning mode 40 or can be determined b~forehalld.
-16-
21 62433
The CPU 16 then determines whether the power consumption signal 28
crosses the power threshold 78 in 96. If the power consumption signal 28 crosses
the power threshold 78, the CPU increments the counter, Crossings, in step 98. In
step 100, during a cutting cycle, as soon as the counter indicates that the number
5 of crossings by the power consumption signal 28 of the power threshold 78 is more
than of crossings caIculated in the learning mode, the CPU 16 indicates an alarm
92. Since the crossings are counted during a cutting cycle, the number of crossings
can be compared to statistical data from the learning mode before the end of the
cycle, i.e. the number of crossings can be compared for selected fractions of the
10 cutting cycle.
In 102, if the CPU 16 detects that the machine tool 18 has not completed a
cycle, the CPU 16 retum to step 88 to monitor the next time segment of the power
consumption signal. If the end of a cycle is detected, then the number of crossings,
which is stored in "Crossings", is stored with a predetermined number of previous
15 numbers of crossings.
In step 106, the CPU evaluates the trend of the number of crossings by the
power consumption signal 28. If the number of crossings is increasing steadily over
the previous predetermined number of cycles, this would indicate that the machine
tool 18 is worn and h~(ling towards failure. The CPU 16 may intliç~t~ an alarm
20 92.
If the end of a cycle is detected in 102, and the trend of crossings does not
indicate tool wear in 106, the counter, Crossings, is reset in step 86 and the tool
monitoring system 10 monitors the next cycle.
2 1 62433
Note that all of the steps above require only one addition and comparison,
therefore they can be performed within the time interval of two monitoring sample
points, so that it can be used for real-time monitoring. Moreover, the present tool
monitoring system 10 accurately predicts tool failure by selecting a threshold and
5 dealing with the effects of various noises in several ways. The tool monitoring
system monitors only selected components of the power consumption signal to
elimin~te the effects of unwanted noise. In recognizing that the power consumption
signal of a tool in normal condition will have a number of spikes, the tool
monitoring system monitors the number of crossin~s of the chosen power threshold,
10 rather than indicating an alarm after a single crossing of a larger power threshold.
Further, by basing the threshold on statistical data from the power consumption
signal from learning cycles, the power threshold is an accurate function of time over
the machining cycle.
It should be understood that this invention can be broadly utilized in a
15 number of distinct fashions. As one example, the learning cycle could be performed
with the first several cycles of each new tool. Thus, the system could be placed in
a new tool, and can recalculate its thresholds with the first several cycles, which
will be known to be operations on a new tool. With such a system, there is also the
alternative for including in the memory of the controller an çst;m~te of what the
20 thresholds typically arrived at for the particular type of tool. Thus, when the tool
begins to set its own t~resholds based on its actual learning cycle, it wiIl be
modified from this predicted threshold. Moreover, these predicted thresholds are
2 1 62433
utilized during the first several cycles to predict catastrophic failure, as described
above.
With re~ard to the actual monitoring of the tool during its cyclical operation,
it is anticipated that the actual monitored signals will not need to be broken into
5 components. That is, the actual raw signals can simply be compared to the power
thresholds in a real-time fashion. This greatly simplifies the calculation and
required times to determine when a tool is worn. In such a system, some simple
adjustment of the thresholds as calculated from the learning cycles may be necessary
to include the effect of the removed information from the learning cycles. This
10 adjustment may be as simple as adding some fixed amount to the thresholds to
account for that which was removed to create the thresholds.
Alternatively, it may also be preferable to take only components of the signal
that is being monitored and compare those components to the thresholds. That is,
the actual monitoring of the tool may utilize the same method steps that were used
15 to create the thresholds in its learning cycle.
In accordance with the provisions of the patent statutes, the present invention
has been described in what is considered to represent its ~lefelled embodiment.
However, it should be noted that the invention can be practiced otherwise than as
specifically illustrated and described without departing from its spirit or scope.
-19-
