Latest / Mechanical Engineering Made Simple / Why Your Vibration Data Lies to You

Transcript

Why Your Vibration Data Lies to You · Mechanical Engineering Made Simple

0:00
Welcome back to Mechanical Engineering Made Simple, the
0:02
podcast that makes you sound intelligent without needing APHD
0:04
or therapy. Or a very expensive legal team,
0:07
depending on how bad your design fails.
0:09
Right. Yeah, exactly.
0:11
So I want you to imagine just for a second that you are the
0:14
lead design engineer on this like highly critical aerospace
0:18
project. OK, no pressure, right?
0:21
You spent, I don't know, three years and maybe $5 million
0:25
designing this absolutely beautiful intricate titanium
0:29
mounting bracket for a satellite component.
0:32
You've run the finite element analysis models hundreds of
0:35
times. The color gradients on your
0:37
computer screen or just they're perfect.
0:39
A sea of comforting blue. Yeah, zero stress
0:41
concentrations. The math is pristine.
0:44
So you finally send the physical prototype down to the shop floor
0:46
to be run on a shaker table. You go down to watch the test,
0:49
fully expecting a victory lap, and within what, 45 seconds your
0:53
$5,000,000 bracket chatters cracks and completely shears in
0:56
half your summit drops. You failed, but here is the
1:01
kicker and the whole reason we are doing this deep dive today.
1:04
Yeah, your bracket was perfectly fine.
1:07
It was the test that lied to you.
1:08
Yeah. And honestly, that is it's the
1:11
ultimate nightmare scenario for any design engineer.
1:15
And the scary part is, it happens far more often than
1:18
anyone in the industry actually wants to admit.
1:20
Yeah, there is this massive, incredibly treacherous gulf
1:24
between the idealized mathematical environment
1:28
upstairs in the engineering department and the messy,
1:30
chaotic physical reality down on the shop floor.
1:33
OK, let's unpack this because if you are listening to this right
1:36
now, you are probably someone who wants to avoid those costly,
1:39
career damaging mistakes. We're looking at a massive stack
1:42
of definitive source material on shock and vibration analysis
1:45
today. Our mission is basically to map
1:47
out this disconnect because on one side we have the folks who
1:50
run the math and write the specs, and on the other side we
1:53
have the technicians who actually have to, you know, melt
1:56
the sensors, glue the strain gauges, and run the drop tables.
1:59
And I guess the foundational truth we just have to accept
2:02
right up front is that the physical world is inherently
2:05
hostile to pristine CAD models. It is entirely hostile.
2:09
I mean, the whole premise of analyzing a structure's
2:11
vibration assumes that you can, you know, accurately measure it
2:15
in the first place. The design engineer just assumes
2:18
the data coming back from the shop floor is pure.
2:20
Right, it's just numbers on a spreadsheet to them.
2:22
Exactly. But the very act of capturing
2:25
that physical data introduces all these physical variables
2:28
that completely poison the results if you don't account for
2:31
them. OK, so let's start right there
2:33
at the actual point of physical contact O Before I can look at a
2:36
gorgeous grah of resonance frequencies on my monitor, a
2:40
technician has to physically attach a transducer, an
2:43
accelerometer to a piece of metal.
2:46
But you know, my brain tells me a sensor is just a passive
2:49
observer. It's just sitting there
2:51
listening to the vibration. But the sources are telling me
2:53
that is a fundamental misunderstanding of the physics.
2:57
It is a profound misunderstanding.
2:59
The sensor, and crucially, the mounting bracket used to hold
3:02
the sensor, are not passive observers at all.
3:06
They are. They're dynamic systems in their
3:08
own right. Every physical object has mass,
3:11
it has stiffness, and it has its own natural frequency.
3:14
OK, so when the shop technician attaches a mounting bracket and
3:17
a transducer to your titanium satellite part, they are no
3:20
longer testing your part. Wait, what are they testing?
3:22
They are testing a brand new combined mechanical system
3:25
consisting of the part plus the bracket plus the sensor.
3:29
Oh wow. So by trying to measure the
3:31
structure, I have essentially built a completely new
3:34
structure. But I mean, how much of a
3:37
difference does a tiny little bracket actually make?
3:39
We're talking about a piece of hardware that fits in the palm
3:41
of your hand. It can completely rewrite your
3:43
data. What's fascinating here is how
3:45
something so small just dominates the math.
3:49
The sources highlight this beautiful, terrifying example.
3:52
Let's say the technician decides the standard steel mounting
3:55
bracket is too heavy. Or maybe, I don't know, they
3:58
just don't have 1:00 nearby. Sure, yeah, shop floor reality.
4:01
Right, so they swap it out for a bracket made of phenolic
4:04
plastic. It has the exact same
4:05
dimensions, the same geometry, but just by changing the
4:09
material from steel to plastic, you actually have the resonance
4:14
frequency of the mounting system.
4:15
Wait, cut the frequency in half, literally 50% just by swapping
4:19
the metal for plastic? Explain the physics of that to
4:22
me. Why does the material change the
4:23
frequency so drastically? Think about a weight bouncing on
4:26
the end of a spring. The speed at which it bounces,
4:29
its natural frequency, depends on two things.
4:32
It depends on the mass of the weight and the stiffness of the
4:34
spring, right? OK.
4:36
A bracket is effectively a very short, very stiff spring
4:40
connecting the sensor to the structure.
4:43
Steel has a very high modulus of elasticity, right?
4:46
It is incredibly stiff. Yeah.
4:48
Phenolic plastic is much softer. It is a weaker spring.
4:52
Yeah. If you put the exact same sensor
4:54
on a weaker spring, it is going to bounce slower.
4:56
The frequency drops. Oh geez.
4:58
So the data acquisition system is going to record a massive
5:01
resonance at, say, I don't know, 5000 Hertz?
5:04
The engineer sitting upstairs is going to panic thinking their
5:07
titanium part has a fatal flaw at 5000 Hertz, but that flaw
5:10
doesn't even exist. It doesn't exist at all, it's
5:12
just the plastic bracket jiggling.
5:14
That's wild. And you know, it goes further
5:16
than that. Let's say they stick with the
5:18
steel bracket, but they change how they attach it to the
5:21
structure. The design might call for a
5:23
mechanical screw mount, but the technician decides they don't
5:27
want to drill a hole in the pristine prototype.
5:30
Because it's a $5,000,000 part. Exactly, so they use a high
5:33
strength epoxy resin adhesive bond instead.
5:37
OK, let me play devil's advocate here for a second.
5:39
Yeah, glue is strong. Modern epoxies are practically
5:43
indestructible. Surely that doesn't change the
5:46
dynamic profile of the test. Oh, it changes it significantly.
5:50
The epoxy bond typically creates a much larger, more uniform area
5:54
of contact between the bracket and the structure compared to
5:57
just a single screw which you know only contacts at the
6:01
threads in the base. OK, I see where you're going.
6:03
Yeah, the epoxy massively increases the stiffness of the
6:06
connection. So going back to our spring
6:08
analogy, you just made the spring wave stiffer.
6:11
The resonance frequency shoots up, the data is altered all over
6:14
again. That feels incredibly
6:16
counterintuitive. The tool you are using to
6:18
measure the object fundamentally alters the object.
6:21
It kind of reminds me of trying to measure the speed of a tiny
6:25
delicate paper boat floating down a stream, and your method
6:29
of measuring it is to drop a massive 5 LB GPS tracking brick
6:33
right into the center of the boat.
6:35
Yes, that is a perfect analogy for the concept of mass loading.
6:39
And mass loading is the absolute bane of vibration testing on
6:43
lightweight structures. If you take a thin aluminum
6:45
panel and you attach a heavy transducer to it, you are the
6:49
GPS brick in the paper boat. You're sinking it.
6:52
You've added localized mass to the panel without adding any
6:56
stiffness. What happens to our weight on
6:58
the spring when we add more weight?
7:00
The bouncing slows down the natural frequency drops.
7:03
Exactly. The engineer gets the report
7:05
back and the data shows the panel resonating at a
7:08
dangerously low frequency. They might spend weeks
7:11
redesigning the panel, you know, adding ribs, adding weight,
7:14
trying to stiffen it up to push that frequency higher, but the
7:18
panel is perfectly fine to begin with.
7:20
They're literally redesigning the part to survive the weight
7:22
of the sensor. Exactly.
7:24
The tail is completely wagging the dog here.
7:26
Yeah. OK, so we we have the bracket
7:28
lying to us, we have the weight of the sensor lying to us, but
7:31
there is a detail in the source material that sounds straight
7:33
out of a science fiction novel and it has to do with the
7:36
cables. Well, the cables, yes.
7:37
The wires connecting the sensor back to the computer
7:40
Triboelectricity. What in the world is a
7:43
triboelectric effect and how is a piece of wire ruining my
7:47
$5,000,000 test? Triboelectricity is essentially
7:51
cable noise generation, and it is a silent killer of accurate
7:54
data. Inside a typical measurement
7:57
cable, like a standard coaxial cable, you have a central
8:00
conductor wire surrounded by a layer of dielectric insulation
8:04
surrounded by a braided metal shield.
8:07
These layers are made of completely dissimilar materials.
8:10
Like copper and Teflon, right? Yes.
8:12
Now imagine the test starts. The shaker table is roaring, the
8:15
structure is vibrating violently, but the technician
8:18
left the cable completely loose, just draping from the sensor
8:20
down to the floor. As the structure shakes, that
8:23
cable whips around. It flexes, it bends, twists.
8:27
And because it's bending, the layers inside the cable are
8:30
physically rubbing against each other.
8:31
Yes, exactly. The friction between the
8:34
dielectric insulation and the braided shielding physically
8:37
strips electrons from one material and deposits them on
8:39
the other. It generates a false static
8:42
electrical charge. It is the exact same physics as
8:45
rubbing a balloon on your hair to create static electricity,
8:48
just happening on a microscopic level inside a wire.
8:51
But wait, the DD acquisition system, The DAQ only reads
8:55
voltage, right? It doesn't know where the
8:57
voltage came in. So the GAQ sees this electrical
9:00
charge from the friction and just assumes the piezoelectric
9:03
crystal inside the sensor generated it.
9:05
It reads it as a massive spike in structural vibration.
9:09
The DAQ is entirely blind to the source.
9:12
A voltage spike from a shattered titanium bracket looks exactly
9:15
the same as a voltage spike from a wiggling piece of Teflon.
9:18
It all just looks like raw data. The engineer sees this
9:21
incredible overload on the graph and assumes the part failed.
9:25
This is terrifying. Yeah, we're talking about highly
9:27
advanced mathematical models being completely defeated by a
9:31
floppy cable. How does a shop floor ensure
9:34
this doesn't happen? Do we just like mathematically
9:36
filter it out? You cannot reliably filter it
9:38
out because the triboelectric noise often occurs in the exact
9:42
same frequency bands as the actual physical vibration you
9:45
were trying to measure. Oh wow, so you're just stuck
9:47
with it? Unless you fix it physically,
9:50
the fix relies entirely on physical hygiene on the shop
9:53
floor. First, To combat the bracket and
9:55
mounting issues, the technician must tap the installed
9:58
transducer with a small tool before the test even begins.
10:01
Like literally just hitting it with a tiny hammer.
10:04
A very small tap, yes. This excites the local
10:07
resonances. If you look at the real time
10:10
analyzer and see resonance spikes appearing that you know
10:13
do not belong to the transducers internal frequency, which you
10:16
know is listed on its calibration sheet, you instantly
10:19
know you have a mounting compliance issue.
10:21
You have to remount it. And what about the cables?
10:24
How do you stop the static electricity?
10:26
You manually move the cable while watching the screen.
10:29
Just grab it and wiggle it back and forth.
10:32
If moving the cable generates noise spikes on your monitor,
10:35
the cable is compromised or you are experiencing a severe
10:38
triboelectric effect. So you just toss it.
10:40
You either replace that cable entirely or you implement
10:43
rigorous tie downs. A cable should never be twisted
10:46
when connecting it to the transducer because that preloads
10:49
internal stress, and it must be securely taped or anchored to
10:52
the structure so that the cable moves perfectly in phase with
10:56
the structure, eliminating the internal flexing.
10:59
So if the cable moves with the metal, it doesn't bend, so it
11:02
doesn't rub. It is a beautifully simple
11:05
physical fix for what looks like an impenetrable mathematical
11:10
anomaly. It really is.
11:11
But so far we are just talking about temporary sensors.
11:14
You bolt them on, you run the test, you take them off.
11:17
What happens when the engineer needs permanent data?
11:20
What happens when they specify an electrical resistance strain
11:22
gauge that has to be physically glued to an airplane wing that
11:25
is going to fly through freezing altitudes and scorching deserts?
11:29
That introduces an entirely new tier of physical sabotage?
11:32
Doesn't. It.
11:33
Oh absolutely. The strain gauge is arguably the
11:35
greatest friction point between the design engineer and the shop
11:38
floor. Engineers treat strain gauges
11:41
like magic data portals. Magic data portals.
11:43
I love that. It's true, they provide
11:46
incredible direct insight into the localized stress a part is
11:50
experiencing. But technicians know that a
11:53
strain gauge is essentially an incredibly fragile temperamental
11:57
science experiment that has to be conducted on a dirty
12:00
workbench. OK, before we talk about how
12:02
they fail, I want to make sure I truly understand how they work.
12:05
My understanding is that a strain gauge is basically an
12:08
impossibly thin piece of metal foil, usually some specific
12:12
alloy attached to a very thin plastic backing.
12:16
You glue the backing to your structure.
12:18
When the structure bends or stretches, the foil stretches
12:21
with it. That is the mechanical premise,
12:23
but the data actually comes from the electrical properties of
12:25
that foil. As you stretch a piece of wire,
12:28
two things happen. It gets longer and it's
12:30
cross-sectional area gets thinner.
12:31
Like stretching a rubber band. Exactly like that.
12:34
And both of those geometric changes make it more difficult
12:37
for electrons to flow through the metal.
12:39
Therefore it's electrical resistance increases.
12:42
If the structure compresses, the wire gets shorter and fatter and
12:45
the resistance decreases. OK, I follow you.
12:47
So we run a small current through the foil, measure that
12:51
microscopic change in resistance, and use it to
12:53
calculate the exact physical strain of the part.
12:56
It is an elegant concept, but the sources highlight that this
13:00
entire chain of logic rests on one microscopic vulnerability.
13:04
Right? The glue.
13:06
Yes, the chemical cleanliness and integrity of the bond.
13:10
If you have a sensor that is measuring physical deformation
13:12
at the micro inch level, the transfer of that deformation
13:16
from the heavy metal structure through the layer of adhesive,
13:19
through the plastic backing and into the foil must be absolute.
13:23
It must be perfect. But the shop floor is not a
13:26
semiconductor clean room. What happens if a technician who
13:29
you know has been working on hydraulic lines all morning
13:32
accidentally leaves a microscopic smudge of oil from
13:34
their fingerprint on the metal before applying the epoxy?
13:37
The adhesive will not bond properly at that specific
13:40
microscopic location. When the massive metal structure
13:43
stretches, the foil sitting directly above that fingerprint
13:46
smudge does not stretch, it simply floats over the gap.
13:49
Wow, just from a fingerprint? Just a fingerprint.
13:53
The resistance change is stunted.
13:55
The engineer's computer receives a reading that the stress is
13:58
well within safe limits, when in reality the structure might be
14:02
yielding and moments away from catastrophic failure.
14:05
The math is completely blind to the fingerprint.
14:08
That is the scary part. The DAU doesn't flash a warning
14:11
light that says bad clue. It just gives you a lower
14:14
number. But let's say the technician is
14:17
flawless. They use chemical solvents, they
14:19
prep the surface perfectly. The bond is molecularly
14:23
pristine. The sources still indicate we're
14:25
walking into a massive analytical trap called apparent
14:29
strain. I need you to explain this to me
14:31
because it sounds like the material itself starts lying.
14:34
It does start lying. Apparent strain is a brilliant
14:37
illustration of why an engineer cannot just look at a
14:39
spreadsheet. They must understand the
14:41
physical environment of the real world.
14:43
OK. Give me an example.
14:44
Let's imagine you perfectly bond a strain gauge to an aluminum
14:48
beam. You leave it sitting on a bench
14:50
in a laboratory. The air conditioning kicks off
14:52
over the weekend and the temperature in the room rises by
14:55
30°. OK, no forces are being applied
14:58
to the beam. It's just sitting there getting
15:00
warm. Correct, but if you look at your
15:02
data acquisition system, it will show a massive amount of
15:05
mechanical strain. It will look like an invisible
15:08
giant is trying to pull the beam apart.
15:10
Wait, because the metal changes electrical resistance when it
15:12
gets hot. That is half the problem.
15:15
The thermal coefficient of resistance of the foil changes,
15:17
yes, but more importantly, the aluminum beam undergoes thermal
15:21
expansion. As it gets hot.
15:23
The beam physically grows. Because the strain gauge is
15:26
perfectly bonded to the beam, the foil is forced to stretch
15:29
along with the expanding aluminum.
15:31
So the foil is physically stretching and its resistance is
15:35
changing due to heat, all without anyone touching it.
15:39
The DAQ measures this massive change in resistance and blindly
15:42
converts it into mechanical stress data.
15:44
It cannot distinguish between a hydraulic press is stretching me
15:48
and I am sitting in a hot room. This is a parent strain.
15:51
It is a completely phantom load. So if I am an engineer designing
15:55
A structural test for a jet engine casing where temperature
15:58
is going to fluctuate wildly, how on earth do I design around
16:02
that? If the heat is creating fake
16:04
strain, my data is completely useless.
16:07
You solve it through meticulous material selection of the foil
16:10
alloy, and this is where engineers often make critical
16:13
errors by specifying the wrong gauge entirely.
16:16
There are two primary alloys used in strain gauges, and they
16:19
serve entirely different thermodynamic purposes,
16:22
Constantan and ISO elastic. OK, let's break those down.
16:25
Why would I use Constantan? Constantan is an alloy
16:28
specifically formulated to have a relatively flat thermal
16:31
response curve. It is self temperature
16:34
compensating to a degree. If you are doing a static load
16:38
test, meaning you are piling weights onto a wing and leaving
16:40
them there for three weeks to see if it slowly bends.
16:42
The temperature in the hangar is going to change between day and
16:45
night, right? You must specify a constant tan
16:48
gauge because you need the alloy to resist changing its
16:51
properties due to those ambient temperature fluctuations.
16:54
That makes perfect sense, but what is ISO elastic?
16:57
Why not just use Constantine for everything if it ignores
17:00
temperature? Because Constantine has a
17:02
relatively low gauge factor, it doesn't produce a massive change
17:06
in resistance for a tiny amount of strain.
17:08
It is not incredibly sensitive. ISO elastic alloys, on the other
17:12
hand, are highly sensitive. They produce a huge, easily
17:16
readable signal, but the trade off is that isoelastic is wildly
17:20
sensitive to temperature. The slightest thermal shift will
17:23
create massive apparent strain. So why does it even exist?
17:26
It sounds like a total liability.
17:27
It is designed exclusively for dynamic shock events.
17:31
Let's say you are testing a ballistic impact or a massive
17:33
pyro shock event. The entire mechanical event
17:36
occurs in two milliseconds. You need the extreme sensitivity
17:40
of the isoelastic gauge to capture that incredibly rapid,
17:44
fleeting spike in strain. And because it's so fast, the
17:46
temperature doesn't matter. Exactly because the event only
17:49
lasts 2 milliseconds, the ambient temperature does not
17:52
have time to change. The thermal instability of the
17:54
alloy is totally irrelevant because the time scale is too
17:57
fast for heat transfer to occur. Let me stop you right there.
18:00
I see the trap. If a design engineer just looks
18:02
at a catalog and says oh this isolastic gauge has a much
18:06
stronger signal output, I'll specify that one for my 3 month
18:11
structural fatigue test. Oh, it's a disaster.
18:14
They will end up painstakingly measuring the cyclic performance
18:17
of the laboratories HVAC system rather than the fatigue of their
18:21
aerospace part. That's hilarious and terrible.
18:24
The data will just be this rolling wave of apparent strain
18:27
that completely obscures the tiny mechanical fatigue cycles.
18:30
And the technician on the floor probably knows this.
18:33
The technician sees the work order specifying an ISO elastic
18:36
gauge for a static test, and they know the data will be
18:38
garbage. But the specification is the
18:41
law. Yep, they just follow the
18:43
blueprint. And that brings us to another
18:45
geometric trap hidden inside the strain gauge.
18:48
Transverse sensitivity. I have looked at these things
18:50
under a microscope. The foil isn't just a straight
18:53
line, it is patterned into a grid.
18:56
It winds back and forth like a hairpin Rd. ascending a steep
18:58
mountain. Why do we do that and how does
19:01
it ruin the data? Well, the foil's incredibly
19:03
thin, but we need the electrical path to be very long.
19:06
To get a measurable starting resistance, say 120 or 350 ohms,
19:11
you need a lot of wire. If you just used a straight
19:14
wire, the strain gauge would be a foot long.
19:17
To compress all that wire into a footprint the size of a postage
19:20
stamp, we etch the foil into a Serpentine grid, a series of
19:25
long vertical lines connected at the top and bottom by tiny
19:28
horizontal loops. Right, the hairpin turns.
19:30
Exactly. You mount the gauge so that the
19:32
long vertical lines are pointing in the direction of the string
19:35
you want to measure. When the part stretches long
19:37
line, those long lines stretch and you get your data.
19:41
But here's the fatal flaw. The tiny horizontal loops at the
19:44
ends of the grid. Because those segments of foil
19:46
are traveling sideways transversely, they will stretch
19:49
and change resistance at the metal beneath them stretches
19:51
sideways. I think I see the problem.
19:53
If I put a heavy load on a block of rubber, it squishes down, but
19:57
it also bulges out the sides. Metal does the same thing, just
20:00
on a microscopic scale, right? Poisson's ratio.
20:02
Yes. When you stretch a piece of
20:05
metal lengthwise, it simultaneously contracts across
20:08
its width. The transverse loops on the
20:10
strain gauge pick up that lateral contraction and alter
20:14
the total electrical resistance of the grid.
20:17
The gauge accidentally reads strain from a direction you did
20:20
not intend to measure. So a portion of my signal is
20:22
contaminated by transverse motion.
20:24
If the engineer doesn't know the exact transverse sensitivity
20:27
coefficient of that specific gauge batch, they might be
20:30
overestimating or underestimating the primary
20:33
longitudinal stress. We're building a massive tower
20:36
of assumptions here. And that is the reality of the
20:39
physical disconnect. The math is only as solid as the
20:42
cleanliness of the technician's hands, the thermal foresight of
20:44
the engineer, and the geometry of a microscopic loop of foil.
20:48
OK, let's say a miracle occurs. The shop floor executes Florida
20:51
State, The brackets are stiff, the cables are anchored, the
20:54
strain gauges are perfectly bonded with the exact right
20:56
alloy, and the room temperature is locked in tight.
21:00
The physical setup is Immaculate.
21:02
The raw data, pure unadulterated voltage, now flow through the
21:05
wires into the engineer's computer.
21:07
The physical pain is over. And the digital pain immediately
21:11
begins. Awesome.
21:13
Welcome to the data processing minefield.
21:15
This is where an entirely new set of brutal misconceptions
21:19
lives, because taking continuous analog voltage and forcing it
21:22
into a digital spreadsheet requires mathematical
21:25
compromises. I want to start with what the
21:28
sources call out as the apex predator of condition monitoring
21:31
mistakes. It is a misconcetion so common
21:34
that it literally dictates how entire factories are run.
21:38
The overall level misconception break this down for me.
21:41
This is a pervasive systemic issue when monitoring the health
21:44
of operating machinery. Humans have a deep seated
21:47
psychological desire to simplify complex data.
21:50
If you have a massive industrial compressor running in a factory,
21:53
management does not want to look at a swirling 3 dimensional
21:56
waterfall plot of frequency Spectra.
21:58
They just want a green light or red light.
22:00
Exactly. They want a single, easily
22:02
digestible number on a dashboard.
22:04
They want to boil the entire health of a complex machine down
22:07
to an overall vibration vibration level.
22:09
Right. If the overall vibration goes
22:10
above a certain threshold, the alarm sounds.
22:13
It seems incredibly logical. More vibration means more
22:17
damage. It seems logical, but it is a
22:19
catastrophic trap. The sources give a fantastic
22:22
example of why this fails. Relying on an overall root mean
22:26
square measurement of vibration often results in totally
22:30
incorrect predictions, almost always wildly overestimating the
22:34
amount of lead time you have before machine tears itself
22:36
apart. Wait, why does it overestimate
22:38
the time? Why doesn't the overall level
22:40
spike when a bearing starts to fail?
22:42
Because of the massive disparity in energy between low frequency
22:46
motion and high frequency motion, let me give you an
22:48
analogy. Imagine you are trying to listen
22:51
for the faint high pitched whine of a single mosquito, but you
22:55
are standing inside a room with a roaring low frequency jet
22:59
engine. OK, the jet engine completely
23:01
dominates the acoustic energy in the room.
23:03
Yes, if you have a decibel meter, an overall sound level
23:07
meter, it will read a massive number, say 120 decibels,
23:10
generated entirely by the jet engine.
23:13
Now let's say the Mosquito gets 10 times louder.
23:15
Will your overall decibel meter change?
23:17
Not a bit. The mosquito's energy is a
23:19
rounding error compared to the roar of the engine.
23:22
That is exactly what happens in machinery.
23:24
The heavy rotating mass of the shaft, the unbalance of the
23:27
rotors, the normal pulsing of the fluids.
23:30
These generate a massive low frequency wave that dominates
23:33
the overall vibration level. A failing rolling element
23:37
bearing, however, produces tiny sharp high frequency impacts as
23:42
the cracked steel balls roll over a pit in the Raceway.
23:45
I see the energy contained in that high frequency click is
23:48
microscopic compared to the low frequency hum of the machine.
23:52
So the bearing could be grinding itself into metallic dust,
23:55
sending out thousands of tiny high frequency shock waves, and
23:58
the overall vibration average won't even twitch.
24:01
The overall level will not shift enough to trigger an alarm until
24:04
the bearing completely eases, the shaft drops and the machine
24:08
physically breaks itself apart. You get no early warning.
24:11
By the time the overall level spikes, it is no longer a
24:14
warning, it is a post mortem. Wow.
24:16
So how does an engineer actually catch the mosquito?
24:19
How do we find that bearing defect before the machine
24:22
destroys itself? You must abandoned the overall
24:24
level entirely and rely on specific frequency Spectra.
24:28
The data acquisition system takes the complex time waveform
24:32
and runs it through a fast Fourier transform and FFT.
24:36
This algorithm breaks the chaotic wave down into its
24:38
individual frequency components. It separates the jet engine from
24:42
the Mosquito. It creates a graph where the
24:44
horizontal axis is frequency and the vertical axis is amplitude.
24:48
Correct. You then compare that specific
24:51
spectrum against a baseline reference spectrum that was
24:53
taken when the machine was brand new and healthy.
24:55
You are no longer looking for an overall increase, you are
24:58
looking for highly specific localized tells fingerprints of
25:02
failure. What does a fingerprint of
25:04
failure look like on a frequency graph?
25:06
Let's say you see a sharp spike in vibration appearing at
25:09
exactly 2 times the running speed or the RPM of the shaft.
25:12
That is a classic signature of mechanical looseness.
25:15
Like the bolts are rattling. Yeah, maybe the mounting bolts
25:18
on the base plate are backing out.
25:20
The unbalance of the shaft hits the loose footing twice per
25:23
rotation, generating a spike at two times the RPM.
25:26
If you see discrete spikes at exactly 123 and four times the
25:31
RPM of a belt drive, you are likely looking at a faulty, worn
25:34
belt slapping the sheaves. You diagnose the machine by
25:37
looking at the specific frequencies where the energy is
25:40
hiding. That is incredibly elegant.
25:42
But to get those beautiful diagnostic frequency Spectra,
25:46
the analog voltage signal from the sensor has to be digitized
25:48
to the DAQ has to be chopped up into discrete numbers, and the
25:52
sources detailed 2 massive signal processing pain points
25:56
that occurred during this analog to digital conversion, aliasing
25:59
and leakage. Both absolute nightmares.
26:01
I. Want to spend some time on
26:02
aliasing because the math behind it feels almost deceptive.
26:05
Aliasing is perhaps the most insidious error in digital
26:08
signal processing, because if it happens, the computer doesn't
26:11
crash. It doesn't give you an error
26:13
code, it simply lies to you with absolute mathematical
26:16
confidence. Aliasing occurs if you're
26:19
sampling rate. The speed at which the computer
26:22
takes snapshots of the incoming voltage is not fast enough to
26:25
capture the true physical frequency of the vibration.
26:29
The sources mentioned the Nyquist frequency limit, and
26:31
here's where I want to challenge the standard textbook
26:34
explanation. The text says F sub A = 1 / 2
26:39
times delta T Delta T being the time between digital samples.
26:43
That formula means nothing to be physically.
26:45
I need to visualize why the math breaks down.
26:47
Let's visualize it without the formula.
26:49
To accurately capture the shape of any wave, the absolute
26:52
minimum requirement is that you must take a sample at least
26:55
twice per cycle. You need to catch the peak of
26:57
the wave and the valley of the wave.
26:59
If a wave is vibrating at 100 Hertz, 100 cycles per second,
27:02
your computer must take at least 200 samples per second to see
27:06
it. That is the Nyquist limit.
27:08
What happens if I don't? What happens if a high frequency
27:11
vibration sneaks into my sensor? Say vibration at 400 Hertz, but
27:16
my computer is only taking snapshots at 100 times a second.
27:19
The digital system cannot accurately represent that fast
27:22
400 Hertz wave. It doesn't have enough
27:24
snapshots. But the computer has to draw a
27:27
line connecting the snapshots it did take.
27:30
Because the high frequency wave is oscillating so rapidly
27:33
between the slow snapshots, the computer draws a smooth line
27:36
connecting the dots that happens to form a completely false low
27:40
frequency wave. The high frequencies physically
27:43
fold back into the data. I'm trying to think of a
27:45
physical analogy for this. It's like standing in a
27:49
windowless room trying to figure out how fast ocean waves are
27:52
crashing against a pier, but you're only allowed to open a
27:55
small window and look outside for a split second once every 10
27:58
seconds. That is a phenomenal analogy.
28:00
Let's play that out. If the real ocean waves are
28:02
crashing fast, say a new wave hits the pier every single
28:05
second, but I only open my window every 10 seconds, I might
28:09
accidentally look out right as a wave is hitting the pier. 10
28:12
seconds later I look out again and I happened to catch another
28:14
wave hitting the pier. I close the window and write
28:17
down. The ocean is very calm.
28:19
One wave hits the pier every 10 seconds, and you have
28:22
constructed a perfectly logical, mathematically sound conclusion
28:27
based on your data. But your data is a
28:29
hallucination. The high frequency waves have
28:32
aliased into a false low frequency observation.
28:35
And the most common visual example of this is the wagon
28:38
wheel effect in movies, right? Yeah.
28:40
You are watching a movie of a car accelerating.
28:42
The wheel is spinning faster and faster.
28:44
Suddenly the spoke seemed to freeze and then they start
28:47
spinning backward even though the car is going forward.
28:49
Yes, exactly. The movie camera is taking a
28:51
snapshot 24 times a second. As the wheel spins faster, it
28:56
eventually rotates so far between frames that the next
29:00
spoke moves into a position slightly behind where the
29:03
previous spoke used to be. The camera connects the dots and
29:06
your brain perceives a slow backward rotation.
29:09
That backward rotation is an aliased frequency.
29:12
So if a design engineer gets aliased vibration data, they
29:15
might see a massive low frequency resonance spike on
29:19
their graph. They might spend months
29:21
redesigning the heavy structural supports of their machine to
29:25
eliminate a low frequency hum. A hum that does not physically
29:28
exist. The actual problem was a high
29:30
frequency vibration from a gear mesh that alias downward.
29:33
They are trying to solve the ghost in the machine.
29:35
That's insane. So how do you prevent it?
29:37
Do you just buy a faster computer?
29:39
You cannot solve it digitally. Once a signal is aliased in the
29:42
digital domain, it is mathematically impossible to
29:44
separate the fake low frequencies from the real low
29:47
frequencies. It is baked in.
29:49
The only way to prevent it is in the analog physical world.
29:52
So what do you do? You must pass the raw analog
29:54
voltage through a physical hardware low pass filter before
29:58
it hits the digital converter. The filter physically scrubs
30:01
away any frequencies higher than your Nyquist limit, ensuring
30:05
they cannot enter the computer and cause havoc.
30:07
You literally have to put a bouncer at the door of the
30:09
computer to keep the fast waves out.
30:11
I love that. Now what about the second signal
30:15
processing trap leakage? This one sounds messy.
30:18
Leakage is an artifact of the Fast Fourier Transform algorithm
30:21
itself. The mathematics of the FFT
30:24
assume that the chunk of time you are analyzed in the sample
30:27
record is a perfectly periodic signal that loops infinitely.
30:31
Meaning the ways you recorded perfectly repeats forever.
30:33
Yes. For the math to work perfectly,
30:36
your sample record must capture a complete whole integer number
30:39
of cycles of the vibration wave. It must capture exactly 5 full
30:43
waves, or exactly 50 full waves. It has to start at zero and end
30:46
at 0. But in the real world, you were
30:48
just hitting the record button for an arbitrary amount of time.
30:51
You might accidentally capture 5 1/2 waves.
30:53
You chop the wave off mid cycle. And when you chop a wave off mid
30:57
cycle, the FFT algorithm panics. The sharp discontinuity at the
31:02
end of the recording violates the mathematical assumption of
31:05
continuity. As a result, the algorithm
31:07
cannot pinpoint the exact frequency.
31:10
The energy from that sharp, precise frequency literally
31:13
leaks out into the adjacent frequency bins on the graph.
31:16
So a beautiful needle sharp resonance spike at exactly 120
31:21
Hertz suddenly smears out. It looks like a broad, sloping
31:24
mound of vibration energy covering everything from one
31:27
head 10 to 130 Hertz. Exactly.
31:30
The engineer looks at the leaked data and thinks they have a
31:32
broad sloppy resonance problem affecting a wide band of
31:35
frequencies, when in reality it is a highly tuned, very specific
31:39
single frequency issue. They might specify massive
31:42
amounts of damping material to try to absorb that broad amount
31:45
of energy, adding completely unnecessary weight and cost to
31:48
the part. To fix it, you use a windowing
31:49
function, right? The software artificially forces
31:52
the edges of your chopped recording down to 0 to smooth
31:55
out the transition. Yes, a handing window is the
31:57
most common. It gently tapers the beginning
32:00
and end of the recording to 0, suppressing the leakage effect
32:03
at the cost of slightly reducing the overall amplitude accuracy.
32:08
It is a necessary mathematical compromise.
32:10
OK, let's step back and survey the battlefield.
32:13
The design engineer has navigated the mass loading, the
32:16
temperature shifts, the aliasing and the leakage.
32:19
They have pure, accurate data. They design the part based on
32:22
this flawless data. Now the finished prototype goes
32:25
back down to the shop floor for qualification testing.
32:28
The final exam. It is time to put the prototype
32:30
on a massive shaker table or a drop table and subject it to the
32:34
ultimate physical stress to see if it survives before they put
32:37
it in a rocket. And this is where the
32:38
theoretical in the physical world often clash most
32:41
violently, because perfectly simulating a complex real World
32:44
Service environment inside a laboratory is incredibly
32:47
difficult, and the methodologies engineers use to bridge that gap
32:50
are plagued by outdated, deeply destructive myths.
32:54
The largest, most pervasive myth detailed in the sources is the
32:57
equivalent damage concept. This is the idea that you can
33:00
simulate a complex multi layered random vibration environment by
33:04
just taking the part, putting it on a shaker table and hitting it
33:08
with a really hard pure sign wave for a short period of time.
33:12
It is a fundamentally flawed, archaic concept that refuses to
33:16
die because it is cheap and easy to program into a Shanker table.
33:20
Let's contrast the two environments.
33:21
In the real world, equipment experiences random vibration.
33:25
Think of a truck driving over a redded dirt road, or the
33:29
aerodynamic buffeting on the skin of an aircraft.
33:32
Random vibration contains energy at all frequencies
33:35
simultaneously, constantly shifting an amplitude.
33:37
It's broadband noise, like static on a television.
33:40
Exactly. When you expose a structure to
33:42
random vibration, it excites multiple structural modes
33:45
simultaneously. The wings flap, the fuselage
33:48
twists, the panels bow, all at the same time.
33:51
These different modes interact with each other.
33:53
Stresses combine in complex, unpredictable ways.
33:56
And a sine wave test is the exact opposite of that.
33:58
A sine wave test is a pure single frequency sweeping slowly
34:02
up the spectrum. It is like hitting a tuning
34:04
fork. The equivalent damage myth
34:07
assumes that you can mathematically equate the
34:09
fatigue damage caused by thousands of hours of complex
34:12
random vibration to a short burst of intense single
34:16
frequency sign vibration. We know it's going to experience
34:19
10 years of random rattling in the truck, so let's just shake
34:22
it incredibly hard at 50 Hertz for two hours and call it even.
34:26
Why does that damage model fail? Because fatigue is not linear
34:30
and structural interaction is critical, when the shaker hits
34:33
the structure with a hard sine wave at its primary resonance
34:36
frequency, it excites that single mode violently far beyond
34:40
what it would ever experience in the real world.
34:42
This usually results in a severe over test.
34:45
The part tears itself apart in the lab from a single axis
34:47
stress that would never occur in isolation in the field.
34:50
So you throw away a perfectly good design because the test was
34:53
artificially harsh. Or conversely, you severely
34:56
under test the equipment. A sine wave sweep only spends a
34:59
few seconds at any given frequency.
35:02
If the part has a critical vulnerability that requires time
35:05
to fatigue, the sine wave sweeps right past it before the crack
35:09
can propagate. The weak design passes the test,
35:11
goes into the field, and fails catastrophically 3 months later.
35:16
Whenever physically possible, you must simulate random
35:18
vibration with random vibration testing.
35:21
You cannot take shortcuts with physics.
35:23
The random vibration is just shaking.
35:26
It gets even crazier when we talk about shock testing sudden
35:28
violent impacts. Let's breakdown a massive real
35:31
world mistake detailed in the sources, attempting to use a
35:35
standard drop table to simulate a pyroshock.
35:37
Oh, this is a classic, deeply misunderstood area of testing.
35:41
It stems from engineers treating the word shock as a universal
35:45
constant. A shock is not just a shock.
35:47
The energy profile matters immensely.
35:49
Let's define the 2 events. A standard drop table test is
35:52
used to simulate a velocity shock.
35:54
You bolt the equipment to a heavy carriage, raise it in the
35:56
air, and drop it onto a specialized anvil.
35:58
It is the exact physical simulation of a warehouse worker
36:02
accidentally dropping a television out of the back of
36:05
delivery truck. The TV is falling at high speed
36:08
and then in a fraction of a millisecond it hits the
36:10
concrete. The velocity goes from 30 miles
36:12
an hour to absolute 0 instantly. A massive sudden velocity
36:16
change. Exactly.
36:18
That massive velocity change generates A tremendous amount of
36:21
low frequency energy. The heavy structural frame of
36:25
the TV has to absorb all of that kinetic energy, which causes the
36:29
frame to yield, bend OR snap. Now consider a pyro shock.
36:33
An explosive event. Yes, a pyro shock occurs when an
36:36
explosive device is detonated directly on or near the
36:39
structure. A common aerospace example is an
36:41
explosive bolt firing to separate 2 stages of a rocket in
36:44
space. When the explosive detonates, it
36:47
creates an incredibly intense localized shockwave of extreme
36:50
high frequency energy that ripples through the metallic
36:53
skin of the rocket. But, and this is the crucial
36:55
difference to me, there is almost no net velocity change.
36:58
The rocket doesn't suddenly stop moving.
37:00
It is still cruising through space at Mach 10.
37:03
The structural frame doesn't experience a massive shift in
37:05
momentum. Precisely the heavy structural
37:08
members of the rocket, the thick beams and tresses are physically
37:12
too massive to respond to high frequency energy.
37:16
They act as low pass filters. The shockwave simply passes
37:19
through them without causing them to significantly flex or
37:21
bend. A pyro shock rarely damages
37:24
heavy structural components. So what does a pyro shock
37:27
destroy? It is deadly to delicate high
37:30
frequency components. The intense high frequency
37:33
energy resonates the tiny elements inside microchips, it
37:36
chatters delicate electrical relays, and it fractures
37:39
microscopic solder joints on circuit boards.
37:41
OK, so let's set the trap. A design engineer looks at a
37:44
specification sheet. The sheet says the new satellite
37:47
component must survive a pyro shock event, generating a peak
37:50
acceleration of 2000 GS. The engineer sends the component
37:53
down to the shop floor and says the spec says 2000 GS.
37:57
Put it on the drop table and drop it until it hits 2000 GS.
38:00
What happens? It is an absolute disaster.
38:03
The technician raises the drop table carriage to the ceiling
38:07
and drops it. To generate A2000G peak on a
38:10
drop table, you have to induce an astronomical velocity change.
38:14
When the carriage hits the anvil, it unleashes A colossal
38:17
tidal wave of low frequency energy into the component.
38:21
The television hitting the concrete.
38:23
Yes, and because it is low frequency energy, the heavy
38:26
structural members of the satellite component respond
38:28
violently. The brackets snap, the thick
38:30
titanium beams warp and shear in half.
38:33
The entire assembly is reduced to twisted metal.
38:36
The engineer comes down, sees the wreckage, and assumes their
38:39
design was incredibly weak. They spend months beefing up the
38:42
titanium, adding weight to a satellite where every oz costs
38:45
$10,000 to launch. And the tragic irony is that the
38:48
original lightweight structure would have survived the actual
38:51
explosive pyro shock in space without a single scratch.
38:55
Forcing a drop table to simulate a pyro shock forces a massive
38:59
low frequency structural failure that would never, ever happen in
39:02
the real world. The test environment becomes the
39:04
destructive agent, not the service environment.
39:07
It's like testing a submarine's ability to survive water
39:09
pressure by dropping a bank safe on it.
39:12
You are applying the wrong kind of force.
39:15
And this friction between the equipment and the test machine
39:18
leads us to another fascinating phenomenon detailed in the text
39:21
Equipment structure interactions.
39:24
This is where the physical reality of the test bench fights
39:27
back against the digital controller.
39:29
If we connect this to the bigger picture, we have to realize that
39:32
a shaker table is not an infinite, unstoppable force of
39:35
nature. It is a machine with physical
39:38
limits. When the shop floor mounts a
39:40
massive, heavy piece of equipment to a shaker table, the
39:44
equipment itself becomes an active participant in the
39:47
physics of the table. The sources refer to this as the
39:49
dynamic absorber problem. Yes.
39:51
Let's say you are testing a massive 500 LB electrical
39:54
generator. The generator has its own
39:57
internal resonant frequencies. The shaker table begins its
40:00
sweep, increasing in frequency. When the shaker hits one of the
40:04
generators internal resonant frequencies, the internal
40:07
components of the generator begin to oscillate wildly.
40:10
The mass inside the generator starts moving in all position to
40:13
the table. Exactly.
40:15
The resonating mass of the generator acts as a dynamic
40:17
vibration absorber. It essentially sucks the kinetic
40:20
energy right out of the shaker table.
40:22
The generator is physically fighting to hold the table
40:25
still. The part is screaming at the
40:27
test machine. No, I am vibrating now you stop
40:30
moving. That is precisely the physical
40:32
interaction. But remember, the Shaker table
40:34
is being driven by a digital control loop.
40:36
The computer has a strict mathematical profile.
40:39
It must follow maintain an acceleration of five GS at all
40:43
costs. When the generator absorbs the
40:46
energy, the accelerometers on the table detect a sudden
40:49
massive drop in motion. The table is no longer hitting
40:53
five GS. Because the part is eating all
40:55
the energy. So what does the control
40:56
computer do? The control system panics.
40:59
It sees the dip in acceleration, assumes it is not pushing hard
41:02
enough, and commands the hydraulic amplifiers to pump
41:05
massive, extreme amounts of force into the table to push
41:09
through the resistance and force the table back up to the
41:11
specified motion curve. And what happens to our £500
41:14
generator? It gets absolutely hammered, the
41:16
input faint spikes to destructive levels, and the
41:19
generator is subjected to a severe over test at the exact
41:23
frequency where it is most vulnerable.
41:25
Let's compare that to reality. If that generator was bolted to
41:28
the wing of an airplane and it started resonating and absorbing
41:31
energy, the wing wouldn't suddenly summon infinite
41:35
hydraulic power to fight back, right?
41:36
Exactly. The airplane wing is flexible.
41:39
It yields when the generator resonates.
41:42
The wing allows the local vibration level to dip.
41:44
The structure and the equipment negotiate the energy transfer,
41:48
but the shaker tables control loop refuses to negotiate.
41:51
It acts like an infinitely stiff, infinitely powerful wall.
41:55
It Billy's apart to death. So what is the fix?
41:58
How do we program the Shaker table to act more like a
42:00
flexible airplane wing and less like an unyielding tyrant?
42:04
The solution is a technique called dual control testing.
42:07
Instead of giving the shaker table a single absolute mandate,
42:11
maintain this acceleration motion at all costs.
42:14
You place intelligent limits on both the input motion and the
42:17
input force. You give the control loop 2
42:20
sensors, an accelerometer to measure motion, and a force
42:23
transducer to measure the physical effort being exerted.
42:25
A dual feedback loop. Yes, you tell the system try to
42:29
achieve the requested 5 GS of motion, but if the physical
42:32
force required to achieve that motion exceeds A realistic
42:35
precalculated limit, pull back do not exceed the force limit.
42:39
This allows the table to temporarily dip below the
42:42
required motion when the equipment resonates perfectly,
42:44
mimicking the natural yielding of a real world flexible
42:47
foundation. It prevents the catastrophic
42:49
over test. It introduces empathy into the
42:52
digital control loop. It respects the physics of both
42:55
the part and the environment. I love that.
42:58
Now let's pivot and look at this from the other side of the
43:01
chasm. How can the design engineer
43:04
sitting at their CAD workstation design the system right the
43:07
first time so the shop floor isn't forced to run impossible
43:10
tests or build assemblies that are doomed from the moment they
43:13
are bolted together? Let's talk about designing for
43:16
survival. The most common pitfall
43:18
originating in the design department is the psychological
43:20
trap of over specification. It stems from a very human,
43:24
well-intentioned desire to be safe, to build in a margin of
43:29
error. The classic engineering safety
43:31
factor. If the bridge needs to hold 10
43:33
tons, build the old 30 tons. Exactly.
43:36
An engineer looks at the flight data and sees that the service
43:38
environment will expose the electronic equipment to a
43:41
maximum of 1G of vibration. They want to ensure it survives,
43:44
so they write the formal specification demanding that the
43:47
equipment be designed and tested to withstand 2 GS.
43:50
Double the requirement, double the safety.
43:52
It sounds perfectly logical. It sounds logical, but it
43:55
completely ruins shock and vibration isolation systems.
43:59
Isolators like rubber elastomer mounts, wire rope coils, or
44:03
pneumatic springs are highly tuned, highly non linear dynamic
44:07
systems. You do not just buy generic
44:09
rubber. You select the specific
44:10
stiffness of the isolator based entirely on the exact load and
44:14
the exact deflection it needs to handle.
44:16
Let's follow the math. If the engineer artificially
44:19
double S the specification to 2G's, they must mathematically
44:22
assume the forces involved are twice as high.
44:25
Yes, and to handle those imaginary heavier forces without
44:29
the equipment crashing into the surrounding bulkheads, the
44:32
engineer is forced to specify an isolator that is physically much
44:35
stiffer. So they buy a very hard rigid
44:38
piece of rubber instead of a soft compliant piece of rubber.
44:40
Correct. Now the equipment is built with
44:42
these ultra stiff isolators and it is put into the real world.
44:46
But the real world only delivers 1G of vibration.
44:49
The actual force is half of what the isolator was designed to
44:52
handle. So what happens to the stiff
44:54
rubber? It barely moves.
44:56
It requires 2 GS of force to compress it, so the 1G
45:00
environment cannot overcome its stiffness.
45:02
Because it does not deflect, it does not isolate the kinetic
45:05
energy. It acts like a solid metal
45:08
block, transmitting the high frequency shock and vibration
45:11
directly into the fragile electronics it was supposed to
45:13
protect. By trying to be overly safe on
45:16
paper, the engineer practically guaranteed a structural failure
45:20
in reality. The desire for a safety margin
45:23
actually destroyed the safety mechanism.
45:25
The sources provide a very specific hard geometric rule of
45:28
thumb for this when dealing with elastomeric or rubber
45:31
compression isolators. It isn't just about the
45:34
stiffness, it's about the shape of the rubber changing under
45:36
load. Yes, when you compress a block
45:38
of rubber, the volume doesn't change the material just
45:41
displaces sideways. It bulges.
45:43
The text dictates that to ensure the rubber actually isolates
45:46
dynamically and doesn't just turn into a rigid rock under
45:48
sudden impacts. The ratio of the static
45:50
deflection, how much it swishes just from the weight of the
45:53
equipment to the original uncompressed thickness of the
45:55
rubber should never exceed .15. South a maximum of 15%
45:59
compression. Why is that specific number so
46:02
critical? If you compress the rubber past
46:05
that 15% mark, it bulges too much outward, the internal
46:09
geometric stresses change, and the rubber dramatically stiffens
46:12
in its dynamic response. It might feel soft when you push
46:15
it slowly, but when a high speed shockwave hits it, that highly
46:19
bulged rubber acts almost like steel.
46:22
It loses its isolating magic. It bottoms out dynamically, even
46:26
if it hasn't bottomed out physically.
46:28
That is a brilliant piece of structural insight.
46:31
And Speaking of geometric rules, there is another rule in the
46:33
text that feels like a cheat code for CAD designers.
46:37
Is the center of gravity rule for isolator placement.
46:39
This is perhaps the most vital geometric rule for ensuring an
46:42
isolation system works predictably.
46:44
When you mount a delicate piece of equipment, say a laser
46:47
targeting system, on a set of four isolators, you want the
46:50
entire system to move linearly. When a shock hits, you want the
46:54
box to bounce straight up and down, or shift perfectly
46:56
straight side to side. You want translation.
46:59
Right, like a car riding smoothly over a bump, the whole
47:02
chassis goes up and down evenly. Exactly what you desperately
47:06
want to avoid are rocking modes. A rocking mode occurs when the
47:10
equipment twists, rotates, and pitches on the mounts.
47:14
One side goes up while the other side goes down.
47:16
Why are rocking modes so bad? Because they are incredibly
47:20
complex to damp out, they couple lateral and vertical vibrations
47:24
together chaotically, and they place huge, uneven tearing
47:28
stresses on the isolators themselves, drastically
47:31
shortening their lifespan. So how does the engineer sitting
47:34
at their computer screen prevent the box from rocking?
47:37
By mathematically aligning the geometry of the mounts with the
47:40
center of gravity, the rule is strictly defined.
47:43
The vertical distance between the horizontal plane of the
47:45
isolators in the center of gravity of the equipment must be
47:48
equal to or less than 1/3 of the minimum horizontal spacing
47:52
between the isolators. Let me visualize that to make
47:55
sure the mass translates to reality.
47:57
Imagine I am designing a square metal box and I'm putting a
48:00
rubber mount on each of the four bottom corners.
48:03
If those four rubber mounts are spaced exactly 12 inches apart
48:05
from each other horizontally. Then you take that 12 inches
48:08
divided by three and you get 4 inches.
48:11
The center of gravity of your metal box, the heavy core where
48:14
all the mass is concentrated, cannot be higher than 4 inches
48:18
above the plane of those rubber mounts.
48:20
If the center of gravity is 5 or 6 inches high, what happens?
48:23
The box becomes dynamically unstable.
48:26
It acts like an inverted pendulum.
48:28
The mass is too high relative to the base.
48:30
When a horizontal vibration hits the base, the high center of
48:33
gravity creates a massive moment arm and the box immediately
48:37
wants to tip over and rock. And the beauty of this is that
48:40
the engineer can fix this problem entirely in CAD before a
48:44
single piece of metal is cut. If the center of gravity is too
48:47
high, they don't need complex active damping systems or crazy
48:51
spring rates. They just redesign the brackets
48:53
to lower the attachment points where they widen the stance of
48:56
the mounts from 12 inches to 18 inches.
48:59
Exactly. By simply dragging A mounting
49:01
bracket 3 inches wider on a computer screen, the engineer
49:05
completely eliminates A dynamic, chaotic nightmare that the shop
49:09
floor would have spent weeks and thousands of dollars trying to
49:12
damp out physically. It is the perfect example of
49:15
bridging the gap using the theoretical geometry to solve a
49:19
physical reality. It's amazing how a few pixels on
49:22
a screen translates to surviving a real world shock test.
49:26
As we work toward the end of this incredibly dense material,
49:29
I want to pivot slightly. We've been looking at this from
49:32
a high level systemic perspective, but the sources
49:35
also contain a wealth of granular practical knowledge.
49:38
I want to weave in some of the most common fundamental
49:40
questions that practitioners, both engineers and technicians,
49:43
are constantly asking when they try to implement these rules.
49:46
Let's start with the foundational measurement problem
49:48
that I kept tripping over while reading.
49:49
Let's Hear it. The shop floor technician is
49:51
staring at a massive vibrating industrial fan.
49:54
They have a data acquisition system that is capable of
49:58
converting the raw voltage into three different mathematical
50:01
parameters, displacement, velocity, or acceleration.
50:05
Displacement is how far it moves, Velocity is how fast it
50:08
moves. Acceleration is the rate of
50:10
change of that speed. Which parameter should they
50:13
actually tell the computer to measure and display?
50:15
Does it even matter since they are all mathematically related?
50:18
It matters immensely from an instrumentation and analysis
50:20
standpoint, and the choice is dictated by what we call the
50:23
flattest spectrum rule. The flattest spectrum rule.
50:26
Explain that concept. Let's assume you are measuring a
50:29
complex machine that is generating vibration across a
50:31
huge range frequencies from 10 Hertz up to 10,000 Hertz.
50:36
You want your plotted data graph to be relatively uniform,
50:39
relatively flat across that entire frequency band.
50:42
Why do I want it flat? Why do I want huge spikes on one
50:45
end and tiny microscopic ripples on the other?
50:47
Because of the dynamic range limitations of your sensors and
50:51
your analog to digital converters, if your data has
50:54
massive towering low frequency peaks and incredibly tiny high
50:58
frequency peaks, the computer has to scale its graph to fit
51:02
the massive peaks. When it scales up to fit the
51:04
giants, the tiny high frequency details shrink down into the
51:08
mathematical noise floor and become completely invisible.
51:12
A flat spectrum and shows that an anomaly at 20 Hertz is just
51:15
as visible and requires the same instrument resolution as an
51:18
anomaly at 8000 Hertz. It balances the picture.
51:22
OK, that makes perfect logical sense.
51:24
OK, back to the fan. Which parameter gives me the
51:26
flat picture? For the vast majority of
51:28
rotating machinery, pumps, fans, motors, gearboxes, velocity is
51:32
the parameter that provides the flattest overall spectrum.
51:35
Displacement heavily biases low frequencies.
51:38
A displacement graph of a machine will show massive
51:40
mountains at the low end and practically nothing at the high
51:42
end. Acceleration does the opposite.
51:44
An acceleration graph will look like a flat line at the low end
51:47
and explode into massive spikes at the high end.
51:49
But velocity sits comfortably in the middle, pulling down the low
51:52
frequency mountains and lifting up the high frequency spikes.
51:56
Precisely. Velocity gives you a roughly
51:58
uniform landscape, meaning an increase in vibration at any
52:01
frequency has an equal chance of being spotted by the technician.
52:05
It is the gold standard for general machinery health
52:08
monitoring. That leads me directly into my
52:10
next question, which involves a massive physical roadblock.
52:13
The technician has their velocity sensor.
52:15
They place it on the bearing housing of a critical milling
52:18
machine. But that milling machine isn't
52:20
floating in space. It is bolted tightly to a
52:23
concrete factory floor, and 30 feet away a massive stamping
52:27
press is slamming down every 3 seconds, sending shock waves to
52:31
that shared floor. If the sensors on the milling
52:33
machine, how on earth do you know if the vibration you were
52:36
recording is actually a failing bearing in the mill, or just the
52:39
stamping press vibrating the floor?
52:41
How do you separate the machine from its environment?
52:43
That is one of the most complex challenges in physical
52:46
acoustics, separating the source vibration from the transmission
52:49
path vibration, because in a normal time waveform signal,
52:53
just watching the squiggly line on an oscilloscope, those two
52:57
elements are deeply convolved. Convolved, meaning they're
53:00
multiplied and tangled together. Mathematically, they're baked
53:03
into a single cake, and you cannot easily separate the flour
53:06
from the sugar. Exactly.
53:07
The vibration of the mill and the vibration of the floor
53:10
interact, amplify and cancel each other out in the time
53:13
domain. But there is a highly specific,
53:16
very powerful mathematical tool designed to untangle this knot,
53:20
Cepstrom analysis. Cepstrum.
53:23
CEPSTRUM sounds like an anagram of spectrum.
53:27
It is an anagram of spectrum, and that is a deliberate joke by
53:30
the mathematicians who invented it.
53:32
The Cepstrom is literally the inverse Fourier transform of the
53:35
logarithm of the power spectrum. It is a spectrum of a spectrum.
53:38
A spectrum of a spectrum. OK, my brain is starting to
53:41
melt. I don't need to do the math by
53:42
hand. But physically, what does taking
53:44
the Cepstrom actually achieve? Why does it untangle the cake?
53:47
The magic lies in the logarithm step in the normal time signal.
53:51
As we discussed, the source effects and the path effects are
53:54
convolved. They are multiplied together,
53:56
but when you take the logarithm of the spectrum, mathematical
53:58
properties dictate that multiplied terms become added
54:01
terms. Oh oh wow.
54:04
So instead of a tangled multiplication problem, it
54:06
becomes a simple addition problem?
54:08
A+B. Exactly, and when you perform
54:10
the final transform to create the cepstrum graph, the
54:13
different physical characteristics separate into
54:16
entirely different physical regions on the plot.
54:19
The effects of the source, the sharp repeating harmonics of the
54:23
milling machines gears cluster into highly specific localized
54:27
peaks called romonics. Another anagram romonics instead
54:31
of harmonics. Yes.
54:32
Meanwhile, the effects of the transmission path, the broad
54:35
resonant echo of the concrete floor, stretch out into
54:38
completely different wide regions of the graph.
54:41
You can literally draw a vertical line down the middle of
54:44
the Cepstrom plot. You look at one side to analyze
54:46
the gears in the machine, and you look at the other side to
54:48
analyze the resonance of the factory floor.
54:50
The math forcibly separates the machine from the environment.
54:53
That is essentially computational magic.
54:55
It allows the engineer sitting 1000 miles away to
54:58
mathematically isolate a single piece of metal in a chaotic
55:02
factory. OK, final roadblock.
55:05
Let's move away from monitoring and back to testing.
55:08
I'm looking at two different test specifications in the
55:10
sources, a durability test and a functional test.
55:14
To a layman, a test is a test. Put it on the shaker and shake
55:17
it. What is the actual physical
55:19
difference between these two paradigms?
55:20
It entirely comes down to what specific category of physical
55:23
failure the engineer is hunting for.
55:25
The testing philosophies are diametrically opposed.
55:28
A durability test is designed to find hard failures.
55:31
Define a hard failure. Physical irreversible breakage,
55:34
Metal fatigue, A structural beam cracking in half a mounting
55:38
bracket shearing its bolts off the chassis, permanently
55:41
warping. These are structural failures
55:43
that physically destroy the integrity of the part.
55:46
And the crucial element of a hard failure is that it is time
55:48
dependent, right? You don't usually bend a piece
55:51
of steel back and forth once and have it break.
55:53
It takes thousands or millions of cycles for the microscopic
55:57
crack to propagate. Exactly.
55:59
Fatigue is a function of time and cycle accumulation.
56:03
Because of this, durability tests are highly accelerated.
56:06
An engineer cannot afford to put a car suspension on a shaker
56:09
table and run it for 10 years to see if it lasts 10 years.
56:12
So we use the inverse power law of fatigue.
56:15
We dramatically increase the amplitude of the vibration,
56:18
shaking it much violently than it would ever experience in
56:21
reality, in order to compress 10 years of normal fatigue
56:25
accumulation into a few 100 hours of testing in the lab.
56:28
So during a durability test it is shaking violently.
56:32
If the digital screen on the dashboard glitches out and
56:35
resets during that violent shaking, does the engineer fail
56:38
the test? No, they do not care.
56:40
The durability test is purely about the physical structure.
56:43
The electronics are not expected to function normally under those
56:45
artificially accelerated exaggerated loads.
56:48
As long as the dashboard doesn't physically snap off its mounts,
56:51
the test is a success. So.
56:53
That brings us to the functional test.
56:54
The functional test hunts for soft failures.
56:57
These are transient, temporary failures that disrupt the
57:00
operation of the equipment without physically breaking it.
57:03
A delicate electrical relay chattering open for a
57:06
millisecond. A laser optic momentarily
57:08
vibrating out of alignment. A software glitch caused by
57:12
signal noise. And unlike metal fatigue, a soft
57:15
failure isn't accumulating overtime.
57:17
It doesn't take 10 years for a relay to chatter.
57:19
If the vibration is high enough, it chatters instantly.
57:22
Precisely. Soft failures are not time
57:24
dependent, they are amplitude dependent.
57:26
Therefore, you cannot accelerate A functional test.
57:29
If you artificially increase the vibration amplitude, you will
57:32
cause relays to chatter that would never ever chatter in the
57:35
real world. You will induce fake failures.
57:37
So how do you run it? You run the functional test at
57:40
the exact maximum vibration level that the equipment is
57:43
expected to encounter in its normal service environment, not
57:46
a fraction of AG higher. And the mandate is absolute.
57:50
The equipment must operate Florida State with full
57:53
electrical and mechanical functionality during the
57:55
entirety of the vibration exposure.
57:57
The duration of the test doesn't really matter, it just has to
57:59
work perfectly while it's shaking.
58:01
That distinction accelerating to break the metal versus real time
58:05
levels to test the electronics is so critical for avoiding the
58:08
twin demons of over testing and under testing.
58:12
We have covered an immense amount of ground today,
58:14
traveling from the microscopic smudges under a strain gauge to
58:17
the colossal low frequency tidal waves of a drop table.
58:21
As we wrap up this deep dive, I want to leave you with a final
58:23
thought to Mull over as you walk out onto the shop floor or sit
58:27
down at your CAD terminal. We've talked extensively about
58:30
the friction between the theoretical model and the
58:32
physical reality. Consider how the very act of
58:35
observing and isolating our physical world fundamentally
58:38
alters it. When we talk about this concept
58:40
in quantum mechanics with the observer effect, the idea that
58:43
observing an electron changes its path.
58:46
But we rarely acknowledge that we experience the exact same
58:48
phenomenon on a greasy workbench.
58:50
When you stick a heavy accelerometer to a thin beam,
58:53
you're not measuring the beam, you're measuring a new system
58:56
you just created. When you bolt a stiff isolator
58:59
to a flexible airplane floor to protect a computer, the floor
59:02
and the isolator engage in a physical argument over who have
59:04
to absorb the energy. We are never just measuring or
59:07
isolating a pure theoretical system in a vacuum, We are
59:11
always inevitably interacting with it.
59:14
The perfect design, the flawless frequency, the pure sine wave.
59:17
These only exist on the mathematical paper.
59:19
The real physical world is a relentless, chaotic,
59:22
collaborative negotiation between the engineer's pristine
59:24
math and the shop technicians physical reality.
59:27
It is a profound realization, and it raises an ultimate
59:30
question for every design. Are you designing for the purity
59:33
of the paper, or are you designing for the chaos of the
59:36
negotiation? Because only one of those
59:38
designs will survive the Shaker table.
59:41
Victory belongs to those who refuse to lose.
59:44
Thanks for riding with me, design junkies.
59:47
Tay hungry. If this eisode gave you even one
59:49
solid take away, drop a comment. It takes you a few seconds, but
59:52
it tells me this workbench is actually helping and worth
59:54
keeing alive. We don't make a dime off this,
59:56
and that's by design. As long as it's helping you get
59:59
sharper, that's enough.