2.

Design of
Experiments

1
Why Are Experiments Needed in Engineering: Basic purposes
To learn unknowns based on what’s known

• Discovery/Development:
uncover unexpected behaviors Material
or phenomena that were not discovery
initially considered and test
innovative ideas or products.

• Optimization: fine-tune
designs and processes for Manufacturing
maximum efficiency and optimization via
performance. mechanical tests

• Validation/Confirmation:
confirm whether theoretical
calculations and simulations Validation of
align with real-world behavior. physical
laws/models
Our course
2
Why Are Experiments Needed in Engineering: Practical aspects
• Ensure Safety, Reliability, and Real-World Validation:
• To identify potential failure modes, ensuring safety and
validating performance under real-world conditions with
variability, weather conditions, and uncertainty.
• Examples: Stress testing airplane wings to determine their
breaking points under extreme conditions.

• Cost-Effective Risk Reduction:

• Identify and address potential issues early in the design
process, saving time and money later.
• Example: Prototyping and testing a robot design before
mass production to avoid expensive redesigns.

• Regulatory Compliance:
• Many engineering solutions must meet specific legal and
industry standards, which require experimental validation.
• Example: Testing medical devices to meet Department of
Health certification requirements.
3
Steps in experiments
Example: alloy toughness optimization
1. Recognize the Problem:
Define the objective and scope of the experiment.
2. Select the Factors and Response Variable:
Determine the controlled factors and their levels/range of variation.
Identify the measurable outcome variables for the objective.
3. Design of the Experiment:
Select an appropriate experimental design (e.g., full factorial,
fractional factorial, etc.) to systematically study the factors and
their interactions. Defines how experiments are structured and
conducted (e.g., Full Factorial, Taguchi, RSM)
4. Perform the Experiment: Full factorial
Execute the experiments according to the design while controlling
constants and minimizing spurious factors.
5. Analyze the Data:
Use statistical tools to evaluate the main effects, interactions, and
significance of the factors on the response variable.
6. Draw Conclusions and Make Recommendations:
Summarize findings, identify optimal conditions, and provide
actionable insights for decision-making or future studies. 4
Design of Experiments (DoE) in Mechanical Engineering
1. Importance Key Components of DOE
2. Optimizes Engineering Processes – Helps in • Factors and Levels – Identifies
systematically planning experiments to improve independent variables (factors) and their
performance, efficiency, and reliability in possible levels affecting the system.
mechanical systems. • Response Variables – Defines measurable
3. Reduces Time and Costs – Efficiently outputs that indicate system performance.
determines the effects of multiple variables, • Experimental Design Types:
minimizing the number of required • Full Factorial Design
experiments.
• Fractional Factorial Design
4. Enhances Product Quality – Identifies key
• Taguchi Method
factors influencing performance, leading to
better design and manufacturing outcomes.
5. Supports Data-Driven Decision Making –
Uses statistical methods to analyze
experimental results, ensuring objective
conclusions.

5
Design of experimental factors*
Why? Experiments are Type Role in DoE Example
• expensive;
Factors intentionally kept
Wall insulation, room size,
• time consuming; Constants fixed to avoid influencing
number of windows.
the results.
• not perfectly repeatable.
Variables deliberately Heater power, ventilation
Controlled Factors manipulated to study their air flow rate, window &
Example: winter room temperature effects. curtain usage.
Outcomes or results
Response Variables measured during the Room temperature (°C).
experiment.
Outdoor temperature,
Unintended or uncontrolled
Uncontrolled Factors humidity fluctuations,
variables but may
(aka spurious factors ) heater/sensor fluctuation,
influence results.
human metabolism.
Combined effects of two or
Interaction between heater
more factors, which may
Interaction Factors power and ventilation air
differ from the sum of their
flow.
individual effects.

Important! 6
Design of experimental factors: example 1

A student wants to study how the cup material (ceramic, plastic, or metal)
affects how quickly hot coffee cools down. They pour coffee at 90°C into
different cups and measure the temperature drop over 10 minutes. The
experiment is conducted in an air-conditioned room.
Which of the following correctly identifies the response variable, controlled
factor, constant, and spurious factor?

A. C.
• Response Variable: Cup material • Response Variable: Initial coffee temperature
• Controlled Factor: Temperature drop • Controlled Factor: Room temperature
• Constant: Room temperature • Constant: Cup material
• Spurious Factor: Air humidity • Spurious Factor: Coffee stirring speed
B. D.
• Response Variable: Temperature drop • Response Variable: Room temperature
• Controlled Factor: Cup material • Controlled Factor: Coffee type
• Constant: Initial coffee temperature • Constant: Cup material
• Spurious Factor: Room temperature fluctuations • Spurious Factor: Coffee cooling rate

7
Design of experimental factors: example 2

A group of students is conducting an experiment to study the efficiency of a small

internal combustion engine by varying the fuel type. They record the engine's
power output (W) and fuel consumption (L/hr). The experiment is conducted in a
controlled lab environment with constant air pressure and temperature.
Which of the following correctly identifies the response variable, controlled
factor, constant, and spurious factor in this experiment?

A. C.
• Response Variable: Power output • Response Variable: Power output
• Controlled Factor: Fuel type • Controlled Factor: Air pressure
• Constant: Air pressure • Constant: Fuel type
• Spurious Factor: Ambient temperature • Spurious Factor: Fuel quality variations
B. D.
• Response Variable: Fuel type • Response Variable: Fuel consumption
• Controlled Factor: Power output • Controlled Factor: Engine temperature
• Constant: Fuel consumption • Constant: Power output
• Spurious Factor: Air temperature • Spurious Factor: Engine load variations

8
Types of experimental designs (DoE)
Best guess Example: winter comfort

• A trial-and-error approach where experiments are based on intuition or prior

knowledge to identify potential solutions.
• Advantages: Simple and fast for initial exploration.
• Disadvantages: May miss interactions between factors and is not
systematic.
One-Factor-at-a-Time (OFAT): Best guess:
Turn heater to medium power
• Only one factor (or variable) is changed at a time while keeping all others
constant to observe its effect on the outcome.
• Advantages: Easy to understand and implement.
• Disadvantages: Inefficient for complex systems, as it ignores interactions
between factors and requires a large number of experiments.

9
Types of experimental designs (DoE)

Full Factorial Design:

• All possible combinations of factors and their levels are tested to study the effect of each factor and their
interactions comprehensively.
• Advantages: Provides complete information about main effects and interactions. (more efficient than OFAT)
• Disadvantages: Requires a large number of experiments, especially as the number of factors increases.

Fractional Factorial Design:

• A subset of a full factorial design where only a fraction of the possible combinations is tested. This is useful for
studying the most critical factors and interactions with fewer experiments.
• Advantages: Reduces the number of experiments while still capturing key insights.
• Disadvantages: May miss higher-order or less significant interactions.

Taguchi Methods (not required):

• A robust design technique that uses orthogonal arrays to systematically study factors and their levels while
reducing variability and improving quality.
• Advantages: Efficient, focuses on minimizing variability and improving performance under real-world
conditions.
• Disadvantages: May not provide detailed information about high-order interactions.

10
Test matrix

Tabulate all variables and their levels to be tested

A two-factor experiment: response R as a function of X and Y.

(e.g., R is temperature, X is the air flow in m3/s, Y is the heater power in kW)

Determine the ranges and levels of X and Y for a test matrix as follows:

X Y Values
Values Y1 Y2 Y3 Y4 Y5
X1     
OR

R(x,y)
X2     
X3     
X4     
X5     

11
One-factor-at-a-time (OFAT) vs. full factorial
Y
X Y1 Y2 Y3 Y4 … Y10

X1 
. .
.     … 
. .
X10 

1. Choose a spot at random.

2. Holding all other factors constant ('y' in
this case), vary the one remaining factor
('x') across the allowable space, seeking
to find the maximum.
3. Choose another factor to vary ('y' is the
only one left in this experiment), hold all
others constant.
4. Repeat step 2,3 until all factor's levels
have been tested (fully fill the matrix with
…

some reptations)
12
One-factor-at-a-time (OFAT) vs. full factorial
• Assume we have two factors, each has two levels
Regular OFAT will need 6 runs But full factorial only needs 4 runs
Run Factor A Factor B Notes 1,2
1 Low Low Baseline (starting point) A B
Change A (A effect at B- B+
2 High Low
B=Low) 3,4
A- -, - -, +
Repeat Baseline (to
3 Low Low A+ +,- +,+
compare B effect)
Change B (B effect at
4 Low High 5,6
A=Low)
Blue arrows = a possible OFAT
Change A again (A effect
5 High High Table = full factorial
at B=High)
6 Low High Repeat B effect test A-, B+

A+, B-

A-, B-
13
One-factor-at-a-time (OFAT) vs. full factorial

Reasons for disfavoring OFAT include:

1. OFAT requires more runs for the same precision in effect
estimation
2. OFAT cannot estimate interactions
3. OFAT can miss optimal settings of factors.
Get worse with more variables and levels

• Factorial designs test multiple factors simultaneously,

allowing for interaction effects to be measured.
• Factorial designs grow exponentially, but OFAT grows
even faster (because of repeated measurements), For 3 factors, OFAT needs 16 runs
making factorial much more efficient as factors increase. while full factorial needs 2^3=8 runs

14
Two-Level Factorial Experiment: Microwave Popcorn
Objective:
Identify key factors affecting popcorn taste score and unpopped kernels

Factor Name Units Low Level (–) High Level (+) Standard Run # A B C Y1 Y2
A Brand Cost Cheap Costly 1 8 – – – 74 3.1
B Time Minutes 4 6 2 1 + – – 75 3.5
C Power Percent 75 100 3 2 – + – 71 1.6
Responses: 4 4 + + – 80 1.2
Taste Score Y1: 1 (Worst) to 100 (Best). 5 3 – – + 81 0.7
Unpopped Kernels Y2: by weight (lower is better).
6 5 + – + 77 0.7

Why Use a Factorial Design? 7 7 – + + 42 0.5

8 6 + + + 32 0.3
Efficient (8 runs instead of 16 for OFAT).
Tests all combinations of factors simultaneously. Basic questions
• Which factors had the biggest impact?
Reveals interactions (e.g., Does time affect taste
• Were there interaction effects (e.g., long time
differently for different brands?). + high power = burned popcorn)?
15
Two-Level Factorial Experiment: Microwave Popcorn
• Let’s begin the analysis by investigating the “main
effects” on the first response (Y1)—taste. (same can Standard Run # A B C Y1 Y2
be done to Y2) 1 8 – – – 74 3.1
• The effect of a factor in the two-level case can be 2 1 + – – 75 3.5
calculated as
3 2 – + – 71 1.6
4 4 + + – 80 1.2
5 3 – – + 81 0.7
For Factor A (Brand)
6 5 + – + 77 0.7
7 7 – + + 42 0.5
8 6 + + + 32 0.3
• Brand (A) has little effect on taste.
• Power (B) and Time (C) have significant effects on taste. Effect Y1 –1.0 –20.5 –17.0
• B (Power): -20.5
• C (Time): -17 Effect Y2 –0.05 –1.1 –1.8
• Power (B) has the strongest impact on taste, followed by
Time (C).
• Brand choice alone does not significantly influence taste.
16
Understanding Interaction Effects in Factorial Design

Interaction effects is calculated by multiplying the

parent terms
Two-factor interactions (AB, AC, BC) show how
two factors combine to influence the response.
• Example: Does cooking time (C) affect taste
differently depending on power (B)?
Three-factor interaction (ABC) examines how all
three factors interact together.
Total of 7 effects can be estimated from an 8-run
factorial design:
• 3 main effects (A, B, C)
• 3 two-factor interactions (AB, AC, BC)
• 1 three-factor interaction (ABC)

• Notice that the interaction effect of BC is even greater on an absolute scale than its parents B and C.
• In other words, the combination of time (B) and power (C) produces a big (negative) impact on taste.
What else can you find? Discuss with your classmates.
17
Reduced one-factor-at-a-time (OFAT) test matrix
Under the condition/assumption that factors are uncorrelated (do not interact), we can design
a reduced test matrix (reduced OFAT), where we vary one factor at a time (holding the other
factor constant). We can derive mathematically all the other untested conditions.
Example
Math example R(x,y) = 1- y2-x2
Y Values
X Values Y1 Y2 Y3 Y4 Y5
X1 

R(x,y)
X2 
X3     
X4 
X5 

• In special cases, the mathematical expression of a response

(R) as function of two non-interacting factors (X & Y)
Knowing a few points along
• Two possibilities: R = F(X)•G(Y) or R = F(X)+G(Y)
the red lines can essentially fit
the surface
• The knowledge of their interaction can be based on theory,
assumption, or prior tests.
• In the example, assume the effects of heater and air flow
non-interacting 18
Reduced one-factor-at-a-time (OFAT) test matrix
Example: the response from a partly filled test matrix suffices to determine all responses
that depend on two non-interacting factors (variables)

Y Values
X Values Y1 Y2 Y3 Y4 Y5
X1 
X2 
X3     
X4 
X5 
10
10
𝑅 𝑋, 𝑌 = ෍ 𝑘 ⋅ 𝑎1 𝑋 2 + 𝑎2 𝑋 + 𝑎3 + 𝑎4 𝑌 2 + 𝑎5 𝑌 + 𝑎6
𝑘
𝑘=1
10 10
10
= ෍ 𝑘 ⋅ 𝑎1 𝑋 2 + 𝑎2 𝑋 + 𝑎3 +෍ 𝑎4 𝑌 2 + 𝑎5 𝑌 + 𝑎6
𝑘
𝑘=1 𝑘=1
Add zeros
𝑅 𝑋1 , 𝑌4 = 𝑃(𝑋1 ) + 𝑄(𝑌4 ) = 𝑃 𝑋1 + 𝑄 𝑌3 − 𝑄 𝑌3 + 𝑃 𝑋3 − 𝑃 𝑋3 + 𝑄 𝑌4

= 𝑃 𝑋1 + 𝑄 𝑌3 + 𝑃 𝑋3 + 𝑄 𝑌4 − 𝑃 𝑋3 + 𝑄 𝑌3

= 𝑅 𝑋1 , 𝑌3 + 𝑅 𝑋3 , 𝑌4 − 𝑅 𝑋3 , 𝑌3 19
One-factor-at-a-time (OFAT) test matrix

How about 𝑅 𝑋, 𝑌 = 𝑎1 𝑋 2 + 𝑎2 𝑋 + 𝑎3 𝑎4 𝑌 2 + 𝑎5 𝑌 + 𝑎6 ?

We want to find out R(X4,Y2)

Y Values
X Values Y1 Y2 Y3 Y4 Y5
X1 
X2 
X3     
X4 
X5 

Discuss with your classmates and identify the matrix elements to use to find out the R(X4, Y2)

Hint: Divide and multiple the same R(Xn, Ym)

Which three values in the table shall we need to figure out R(X4,Y2)

20
More complex fractional factorial
Combinations are strategically chosen for efficiency.
A. Heater power B. Indoor air flow C. Window & Curtain
+: high +: high +: open
- : low - : low - : closed
Full factorial Fractional factorial
Expt. A B C Expt. A B C
1 + + + 1 + + +
2 + - + 2 + - -
3 + - - 3 - + -
4 + - + 4 - - +
5 - + +
• Assumption: B = A × C, meaning air flow depends on the interaction between
6 - - +
Heating Power (A) and Window & Curtain Position (C).
7 - - - • Aliasing: Assume the interaction effect A × C is aliased with the main effect of
B, so any observed effect of B could come from either B or the interaction A × C.
8 - - +
• Simplification: Assume A × C is much smaller than B (due to active ventilation),
allowing the DoE to treat them as the same and reduce experimental runs.
A good reference: 22
Design and Analysis of Experiments By Doug Montgomery Not required
Additional consideration – Statistics and Sequence
Statistical Considerations in DOE
• Randomization & Replication – Ensures unbiased results and improves reliability.
• Analysis of Variance (ANOVA) – Determines which factors significantly affect performance.
• Regression Analysis – Models relationships between variables for prediction and optimization.
• Signal-to-Noise Ratio – Used in robust design to improve performance under varying conditions.
(see Lecture 3)

• Tests are preferably completed in a

random order, if possible.
• In time-varying experiments, the data
must be taken in a time-sequenced
manner.
• In tests of destruction, such as material
strength tests, the data must be taken
Monitor the stress in the base of a bridge: Tensile test: sequentially since the sample changes
time varying Destructive during the test.

Analysis of variance - Wikipedia 23

Additional consideration – Scheduling

First, define a set of tasks.

A task is a well-defined activity for which several attributes must be determined:
• Qualifications of the person(s) performing the work
• Necessary conditions for performing the task
• Duration (calendar time) needed to complete the task
• Task goals
Then, illustrate the schedule with a chart (e.g., Gantt chart).

Week
Task 1 2 3 4 5 6 7 8 9 10 11 12
A
B
C
D
E
F
G
H 24
3. Sampling and
Analysis

25
Two types of data in mechanical engineering
A. Static or steady-state B. Transient signal (Time dependent)
(Time independent or weakly dependent)
Pressure Mechanical loading
Length ofinside
a bolta pipe

Stress
Strain
Time

Temperature
of a room Complex function of time

26
Uncertainty: Systematic and Random error

A measured value always come with a degree of uncertainty (e.g., a length of 2.00.1 m)

The form of 𝑥0 ± ∆𝑥0 means that we should expect the true value to fall in the range of
𝑥0 − ∆𝑥0 < 𝑥 < 𝑥0 + ∆𝑥0

Large random error

Two types of errors: Systematic and Random

• Systematic errors appear as shift/bias in

measurements; their mean value is displaced from
the true value. (It can typically be minimized with
careful calibration)

• Random errors fluctuate from one measurement to

the next, distributed around a mean value. They Large
are quantified as uncertainty. systematic
error

27
Data analysis: mean of the sample
Suppose x1, x2, …, xn is a set of n observations (a sample with size n) under the same condition.

What is the best approximate (x0) to the true value?

Given the sum of the squares of the deviations from x0

𝑛
2
𝑆 = ෍ 𝑥𝑖 − 𝑥0
𝑖=1

x0 should minimize the value of S (the principle of least mean square)

𝑛 𝑛
𝜕𝑆
= 0 = ෍[−2 𝑥𝑖 − 𝑥0 ] = −2 ෍ 𝑥𝑖 − 𝑛𝑥0
𝜕𝑥0
𝑖=1 𝑖=1

Therefore, the unbiased estimator of the true value is the arithmetic mean
𝑛

𝑥0 = ෍ 𝑥𝑖 /𝑛 = 𝑥ҧ
𝑖=1 28
Uncertainty analysis: two kinds of standard errors
How to get ∆𝑥0 from a finite number of measurements?

x1, x2, …, xn fluctuation

The sample standard deviation (sample error)
• How much samples scatter from
σ𝑛𝑖=1 𝑥𝑖 − 𝑥ҧ 2 the mean value
𝜎𝑥 = • Essentially the random error due to
𝑛−1
the variation of the measurements
1/(n-1) not 1/n to correct for bias, read: Point Estimators for Mean and Variance

෍ 𝑥𝑖 /𝑛 = 𝑥ҧ 𝑥𝑡𝑟𝑢𝑒 • How close the mean is to the “true”

𝑖=1 value (assuming no drift)
The standard error of the mean (SEM) • The more measurements, the
𝜎𝑥 σ𝑛𝑖=1 𝑥𝑖 − 𝑥ҧ 2 smaller the uncertainty
𝜎𝑥ҧ = =
𝑛 𝑛(𝑛 − 1)

Intuitively, to ensure high confidence that someone else’s measurement will lie within the
range you gave based on your measurement, the interval should be reported as
ഥ − 𝐬𝐜𝐥𝐚𝐞 ∗ 𝐒𝐄𝐌, 𝒙
[𝒙 ഥ + 𝐬𝐜𝐚𝐥𝐞 ∗ 𝐒𝐄𝐌]
29
Linear regression from the minimization of the square error
• Useful in both calibration and Output data y1, y2, …, yn measured under input conditions of x1, x2, …, xn.
data analysis For a linear correlation (y = ax + b) that best fits all the data,
the mean square error S must be minimized:
• Linear fitting most common 𝑛
1
• Visually simple 𝑆 = ෍ 𝑦𝑖 − 𝑎𝑥𝑖 + 𝑏 2
𝑛
𝑖=1
• Easy data processing Setting S/a=0 and S/b=0, we have

y = ax + b 𝑏 ෍ 𝑥𝑖 + 𝑎 ෍ 𝑥𝑖2 = ෍ 𝑥𝑖 𝑦𝑖

𝑛𝑏 + 𝑎 ෍ 𝑥𝑖 = ෍ 𝑦𝑖

Solving these equations simultaneously gives the slope and intercept:

𝑛 σ 𝑥𝑖 𝑦𝑖 − σ 𝑥𝑖 σ 𝑦𝑖
𝑎=
𝑛 σ 𝑥𝑖2 − σ 𝑥𝑖 2
σ 𝑦𝑖 σ 𝑥𝑖2 − σ 𝑥𝑖 𝑦𝑖 σ 𝑥𝑖
𝑏= = 𝑦ത − 𝑎𝑥ҧ
𝑛 σ 𝑥𝑖2 − σ 𝑥𝑖 2
Today’s computer code and calculators can easily find a, b and their SEM
Check out Simple linear regression - Wikipedia 30
Confidence interval and z-score

The vague measurement result we discussed previously 𝑥0 ± ∆𝑥0 is more rigorously expressed as

ഥ ± 𝒛𝝈𝒙ഥ
𝒙
where 𝜎𝑥ҧ is the standard error of the mean and z is the z-score corresponding to the desired confidence
level you want to report.

• This expression defines a confidence interval, which provides a range of values within which the true
population mean is likely to lie (depending on your desired percentage). The larger the z-score, the
larger the range, hence more likely the measured data is within the range.

• A 90% confidence interval means there is a 90% probability that the 𝒙

ഥ ± 𝒛𝝈ഥ𝒙 ​ contains the true mean.

• The z-score is a statistical measure that represents the number of standard deviations a data point (or
sample mean) is from the population mean.

• It comes from the z-test, assuming the data follows Gaussian distribution (strictly t-score should be
Student's t-test - Wikipedia

used when sample data size is taken into account, see Student's t-test - Wikipedia)
31
Confidence interval and z-score
The z-score connects uncertainty to confidence intervals by determining the width of the interval
based on the chosen confidence level. Larger confidence levels (e.g., 99%) require larger z-scores,
resulting in wider intervals to account for greater uncertainty.

Confidence Level z
90% 𝟏. 𝟔𝟓
95% 𝟏. 𝟗𝟔 (“2σ”)
99% 𝟐. 𝟓𝟖
99.7% 𝟑. 𝟎𝟎 (“3σ”)
99.9% 𝟑. 𝟐𝟗
99.99% 𝟑. 𝟖𝟗
99.999% 𝟒. 𝟒𝟐
99.9999% 𝟒. 𝟖𝟗

For example, at the confidence level of 95%, 95% of your repeated measurements should fall
ഥ − 𝟏. 𝟗𝟔𝝈ഥ𝒙 < 𝑥 < 𝒙
in the range of 𝒙 ഥ + 𝟏. 𝟗𝟔𝝈ഥ𝒙 while only 5% will be out of the range
32
A. Repeat measurements
A hypothetic measurement: 1000 times
Response = -3,-2,-1,0,1,2 or 3
(equal probability) 𝑥ҧ
Repeat 50 times

Standard Deviation
𝜎𝑥
2.0

1.5

1.0

With more measurements 0.5

• The mean (𝑥)ҧ approaches the true value (𝑥0 ). 200 400 600 800 1000
Measurement No.

Error of the mean

• 𝜎𝑥 stabilizes to the population std, reflecting 0.25

only the randomness in the measurement. 0.20

𝜎𝑥ҧ
0.15

• 𝜎𝑥ҧ decreases, giving us more confidence that 0.10

0.05
𝑥0 = 𝑥ҧ Why?
Measurement No.
200 400 600 800 1000 33
Error propagation: absolute error of a calculated value

Let F be a function of a few variables F=F(x, y, z, ...). If the values of these variables
are measured or found in literature with corresponding errors

x=x0 ±x0 ; y=y0 ±y0 ; z=z0 ±z0 ; …

If all errors are biased similarly and measured independently, the Absolute Error of
the function at F0=F(x0, y0, z0, ...) is

2 2 2 Subscript 0 means the data

𝜕𝐹 𝜕𝐹 𝜕𝐹
Δ𝐹0 = Δ𝑥 + Δ𝑦 + Δ𝑧 is around the value of the
𝜕𝑥 0 𝜕𝑦 0 𝜕𝑧 0 F0 calculated from the
measured x0, y0, z0
Kline-McClintock method

Use 𝒛𝝈𝒙ഥ for x if you want to include a CI for F, otherwise, 𝝈𝒙ഥ

34
Error propagation: relative error
Relative Error using the ratio expression (percentage change)
2 2 2
Δ𝐹0 𝑥0 𝜕𝐹 Δ𝑥0 𝑦0 𝜕𝐹 Δ𝑦0 𝑧0 𝜕𝐹 Δ𝑧0
= + +
𝐹0 𝐹0 𝜕𝑥 𝑥0 𝐹0 𝜕𝑦 𝑦0 𝐹0 𝜕𝑧 𝑧0
𝜕ln(𝐹) 𝑥 𝜕𝐹
Note = 𝜕ln(𝐹) Δ𝑥0
2
𝜕ln(𝐹) Δ𝑦0
2
𝜕ln(𝐹) Δ𝑧0
2
𝜕ln(𝑥) 𝐹 𝜕𝑥
= + +
𝜕ln(𝑥) 𝑥0 𝜕ln(𝑦) 𝑦0 𝜕ln(𝑧) 𝑧0
𝑦2
𝐹 =5∗ 𝑥∗
Example exp 𝑧

1
Take the ln of the function first ln 𝐹 = ln 5 + ln 𝑥 + 2 ln 𝑦 − 𝑧
2
then calculate. 𝜕ln(𝐹) 1 𝜕ln(𝐹) 𝜕ln(𝐹) 𝜕ln(𝐹)
= , = 2, = 1 = −𝑧
𝜕ln(𝑥) 2 𝜕ln(𝑦) 𝜕ln(𝑧) 𝜕(𝑧)
𝑧

Δ𝐹 𝜕ln(𝐹) Δ𝑥0 2 𝜕ln(𝐹) Δ𝑦0 2 𝜕ln(𝐹) Δ𝑧0 2 1 Δ𝑥0 2 Δ𝑦0 2 Δ𝑧0 2

= + + = + 2 + −𝑧
𝐹 𝜕ln(𝑥) 𝑥0 𝜕ln(𝑦) 𝑦0 𝜕ln(𝑧) 𝑧0 2 𝑥0 𝑦0 𝑧0
Assume x, y, z each have 5.0% error, 𝑧0 = 2, then
Δ𝐹0 1 2
= 5% + 2 ∗ 5% 2 + −2 ∗ 5% 2 = 2.87 ∗ 5.0% = 14% (two sig digits, why?) 35
𝐹0 2
Error propagation: special cases
Simple cases: Addition and subtraction

( x + y ) = (x )2 + (y )2
pessimistic estimate optimistic estimate realistic estimate

x x

y)
y y

x+
y

(
(x+y) (x+y) x
Simple cases: Multiplication and division
𝑥𝑦
𝐹= ln 𝐹 = ln 𝑥 + ln 𝑦 − ln(𝑧)
𝑧
2 2 2
Δ𝐹0 Δ𝑥0 Δ𝑦0 Δ𝑧0 Square of ln derivatives are all 1
= + +
𝐹0 𝑥 𝑦 𝑧
Relative error
36
Uncertainty analysis – An example

An experiment to determine the Young’s modulus of a plastic rod under a fixed load:
𝑚𝑔
𝐸=
𝐴𝜀

The weight (m) that exerts the load is measured by a balance, 10.1 ± 0.4 kg
The area (A) perpendicular to the load is determined with a ruler, 5.10 ± 0.02 cm2
Both at the confidence level of 99%
𝑔 = 9.81 m/s2

The engineering strain (ε) measured by a strain gauge for five time
0.0020, 0.0017, 0.0018, 0.0019, 0.0020 Kept one more significant
𝜀 ҧ = 0.00188 figure (than they should
𝜎𝜀ത = 0.000058, Δε = 𝜎𝜀ത ∗ 2.58 = 0.00015 have) to reduce rounding
errors (optional)
10.1 × 9.81 8 Pa = 0.10 GPa
𝛦= = 1.03 × 10
5.10 × 10−4 × 0.00188
CI z 𝑔𝛥𝑚 2 𝑚𝑔𝛥𝐴 2 𝑚𝑔𝛥𝜀 2
𝛥𝐸 = + − 𝐴2𝜀 + − 𝐴𝜀2 =0.01 GPa
𝐴𝜀
99% 𝟐. 𝟓𝟖
Finally: 𝛦 = 0.10 ± 0.01 GPa
37
Error reduction

Systematic Error (Bias)

Example: Eliminating Parallax Error
1. Calibrate instruments regularly.
• Error: A ruler reading is consistently
2. Minimize the influence of external factors: off because the observer views it
Perform calibration and measurements under from an angle.
consistent conditions, such as the same • Solution: Ensure eye level
temperature and humidity. alignment with the measurement

3. Ensure proper setup of measurement tools: scale to get an accurate reading.

For example, place sensors as close as possible

to the source of the signal to reduce
measurement distortion.

38
Error reduction

Random Errors
1. Increase reliability by repeating measurements and analyzing the
uncertainty.
2. Estimate uncertainty when uncertainty analysis is not possible:
• For stable measurements, use half of the least significant digit displayed
as the uncertainty estimate.
• For fluctuating measurements, use half the range of fluctuation displayed
as the estimate. Uncertainty:
3. Prevent instrument drift: Allow instruments to reach thermal equilibrium 0.5 if steady;
1 if fluctuating
with their environment by turning them on a few hours before use. btw 250 - 252
4. Discard invalid data caused by operator mistakes.
5. Develop skill and consistency through care and experience.

39
B. Time dependent signal: Fourier transform

Real signal as a function of time Diffusion equation

𝜕𝑇 𝜕2𝑇
=𝐾 2
𝜕𝑡 𝜕𝑥

A general solution
∞

𝑇 𝑘, 𝑥 = ෍ (𝑎𝑘 𝑠𝑖𝑛𝑘𝑥 + 𝑏𝑘 𝑐𝑜𝑠𝑘𝑥)

Joseph Fourier
𝑘=0
Fourier Transform

A function of frequency
Since temperature T can have any arbitrary profile,
can we decompose any arbitrary function into a series
of sin and cos as the basis?

40
The Use of DFT and Fourier Transform in Mechanical Engineering

The Fourier Transform (FT) and Discrete Fourier Transform (DFT) are powerful mathematical tools
used in mechanical engineering to analyze and process signals. They help engineers gain insights
into frequency components, identify patterns, and design better systems.

Applications:

• Vibration Analysis: Identifying

dominant frequencies in
vibrating systems to detect
faults and optimize designs.

vs.

41
The Use of DFT and Fourier Transform in Mechanical Engineering

The Fourier Transform (FT) and Discrete Fourier Transform (DFT) are powerful mathematical tools
used in mechanical engineering to analyze and process signals. They help engineers gain insights
into frequency components, identify patterns, and design better systems.

Applications:
Seemingly random data
• Vibration Analysis: Identifying dominant frequencies in
vibrating systems to detect faults and optimize designs.

• Main Component Analysis: Understanding the key

frequency components of a complex signal for efficient
system modeling.

42
The Use of DFT and Fourier Transform in Mechanical Engineering

The Fourier Transform (FT) and Discrete Fourier Transform (DFT) are powerful mathematical tools
used in mechanical engineering to analyze and process signals. They help engineers gain insights
into frequency components, identify patterns, and design better systems.
Raw time domain signal

Applications:

• Vibration Analysis: Identifying dominant frequencies in

vibrating systems to detect faults and optimize designs.

• Main Component Analysis: Understanding the key Transform then filter

frequency components of a complex signal for efficient
system modeling.

• Signal Filtering: Removing noise or unwanted Transform back

frequencies in sensor measurements.

43
Fourier series
1
Any bounded, periodic function (of a period T, and hence a single frequency 𝑓0 = 𝑇, angular frequency
2𝜋
𝜔0 = ) can be represented by the summation of an infinite series of sinusoidal functions
𝑇

∞ ∞
𝑎0 𝑎0
𝐹 𝑡 = + ෍ (𝑎𝑛 cos(2𝜋𝑛𝑓0 𝑡) + 𝑏𝑛 sin(2𝜋𝑛𝑓0 𝑡)) = + ෍ (𝑎𝑛 cos(𝑛𝜔0 𝑡) + 𝑏𝑛 sin(𝑛𝜔0 𝑡))
2 2
𝑛=1 𝑛=1

The coefficients can be computed by integrations within one period [-T/2, T/2]. Since the function is
periodic, any time range [t, t+T], like [0, T] can be used for integration

𝑇ൗ
2
1 𝑇Τ2 2
𝑎0 = ‫𝐹 𝑇׬‬ 𝑡 ⅆ𝑡, and 𝑎𝑛 = න 𝐹 𝑡 cos 𝑛𝜔𝑡 ⅆ𝑡
𝑇 −
2
𝑇
𝑇
−2

𝑇ൗ
2
2
𝑏𝑛 = න 𝐹 𝑡 sin(𝑛𝜔𝑡) ⅆ𝑡
𝑇
𝑇
−2
44
Fourier series
2𝜋
Example: a periodic square wave with amplitude of 2 (T = 60.0 s, 𝜔0 = 60.0 = 0.105 rad/s)

With only the fundamental harmonic at 𝜔0

1
1.5
1
Amplitude (V)

Amplitude (V)
0.5 0.5
0
0 -0.5 0 50 100 150 200 250
t (second)
0 50 100 150 200 250 -1
-0.5 t (second) -1.5
4
𝑓1 𝑡 = sin 𝜔0 𝑡
-1 𝜋

1.5
With two terms
1
4 sin 3 𝜔0 𝑡 sin 5 𝜔0 𝑡 sin 7 𝜔0 𝑡

Amplitude (V)
𝑓∞ 𝑡 = sin 𝜔0 𝑡 + + + +. . . 0.5
𝜋 3 5 7 0
-0.5 0 50 100 150 200 250
• In practical calculation, the summation is cutoff at -1
t (second)

a large number of n, not infinite. -1.5

4 sin 3 𝜔0 𝑡
• Let’s see how it approach the original function 𝑓2 𝑡 =
𝜋
sin 𝜔0 𝑡 +
3
45
Fourier series
We add increasingly higher harmonic sine waves into the series
1.5 1.5 With infinite terms
1 1 1
Amplitude (V)

Amplitude (V)
0.5 0.5

Amplitude (V)
0.5
0 0
0
-0.5 0 50 100 150 200 250 -0.5 0 50 100 150 200 250 0 50 100 150 200 250
t (second) t (second) -0.5 t (second)
-1 -1
-1.5 -1.5 -1

4 sin3𝜔0 𝑡 sin5𝜔0 𝑡 4 sin3𝜔0 𝑡 sin5𝜔0 𝑡 sin7𝜔0 𝑡

𝑓3 𝑡 = sin𝜔0 𝑡 + + 𝑓4 𝑡 = sin𝜔0 𝑡 + + + 𝑓∞ 𝑡 → 𝐹(𝑡)
𝜋 3 5 𝜋 3 5 7

Sawtooth, triangle, or any other periodic waveforms can be expanded into such
infinite series, just with different coefficients
46
Complex exponential form of Fourier series
Trigonometric form Complex exponential form
∞
𝑎0 ∞
𝐹 𝑡 = + ෍ (𝑎𝑛 cos(𝑛𝜔0 𝑡) + 𝑏𝑛 sin(𝑛𝜔0 𝑡))
2 𝐹(𝑡) = ෍ 𝑐𝑛 ⅇ 𝑖𝑛𝜔0𝑡
𝑛=1
𝑇ൗ
2 Euler’s identity 𝑛=−∞
2
𝑎𝑛 = න 𝐹 𝑡 cos(𝑛𝜔0 𝑡) ⅆ𝑡 ⅇ 𝑖𝜑 = 𝑐𝑜𝑠 𝜑 + 𝑖 𝑠𝑖𝑛 𝜑 𝑇ൗ
𝑇 1 2
𝑇
−2 𝑐𝑜𝑠 𝜑 = ⅇ 𝑖𝜑 + ⅇ −𝑖𝜑 ; 1
2 𝑐𝑛 = න 𝐹 𝑡 ⅇ −𝑖𝑛𝜔0𝑡 ⅆ𝑡
𝑇ൗ
2
𝑇
2 1 𝑖𝜑 𝑇
−2
𝑏𝑛 = න 𝐹 𝑡 sin(𝑛𝜔0 𝑡) ⅆ𝑡 𝑠𝑖𝑛 𝜑 = ⅇ − ⅇ −𝑖𝜑 ;
𝑇 2
𝑇
−2 𝑖 = −1

• Real and imaginary components with +𝑛𝜔0

and –𝑛𝜔0 combined to form the original
trigonometric form
• More efficient and concise form for computing
(no need to separately calculate a0)
47
From Fourier Series to Fourier Transform

• The Fourier Series decomposes periodic signals with a single frequency ω0 into discrete frequency
components nω0​.
• What if 𝑓 𝑡 is not periodic?
• For non-periodic signals, it essentially means the period T→∞, which means the spacing between
2𝜋
frequency components Δ𝜔 = 𝜔0 = →0, which can be treated as infinitesimal variation ⅆ𝜔
𝑇
2𝜋
• Intuitively, the frequency components 𝑛𝜔0 = 𝑛 = 𝜔 become a continuous variable
𝑇

• Thus, for non-periodic signals, instead of summing discrete frequency components, we integrate
over all possible frequencies.
Define 𝐹෨ 𝜔
∞ ∞
𝑖𝑛2𝜋 1 ∞ ′ 𝑖𝑛2𝜋
𝐹 𝑡 = lim ෍ 𝑐𝑛 ⅇ 𝑇 𝑡 = lim ෍ න 𝐹 𝑡′ ⅇ −𝑖𝑛𝜔0 𝑡 ⅆ𝑡′ ⅇ 𝑇 𝑡
𝑇→∞ 𝑇→∞ 𝑇 −∞
𝑛=−∞ 𝑛=−∞

𝑖𝑛2𝜋
1 2𝜋 𝑡 1 ∞
= lim σ∞𝑛=−∞ 𝐹෨ 𝜔 ⅇ 𝑇 = ‫׬‬−∞
𝐹෨ 𝜔 ⅇ 𝑖𝜔𝑡 ⅆ𝜔
2𝜋 𝑇→∞ 𝑇 2𝜋
By the definition of Cauchy integral

Derivation not required 48

Fourier Transform
• The Fourier Transform represents any function F(t) as a continuous superposition of
sinusoidal components.
• F(t) is not necessarily periodic, if it is, the Fourier transform will produce its Fourier
series coefficients (sum of weighted Delta functions at discrete frequencies)

Periodic
Fourier transform
∞
𝐹෨ 𝜔 = න 𝐹(𝑡) ⅇ −𝑖𝜔𝑡 ⅆ𝑡
−∞

Aperiodic
Inverse Fourier transform
1 ∞
𝐹 𝑡 = න 𝐹෨ 𝜔 ⅇ 𝑖𝜔𝑡 ⅆ𝜔
2𝜋 −∞

Fourier transform can be written as an operator F

Square pulse
Aperiodic

49
Discrete data from sampling

Analog signal
Continuous data

Sampling
Analog to digital (A/D)
Digital signal • Noise immunity in transmission
Discrete data
• Error detection & correction
• Data processing by computers

How frequent you sample and how much

Digital signals are:𝑥 0 , 𝑥 1 , 𝑥 2 , … 𝑥[𝑁 − 1] you sample?
Sampled at 𝑡0 , 𝑡1 , 𝑡2 , … 𝑡𝑁−1 = 0, ∆𝑇, 2∆𝑇, … (𝑁 − 1)∆𝑇
Sampling rate and data length as new
factors in analysis 50
Discrete Fourier transform (DFT)
Fourier transform for a periodic (T=2π/ω, one period), DFT for a discrete signal 𝑥[𝑛] of length N
continuous function is Fourier series
With sequence number n (unit = 1) as variable
∞ 𝑁−1
2𝜋
−𝑖𝑛 𝑁 𝑘
𝐹(𝑡) = ෍ 𝑐𝑛 ⅇ −𝑖𝑛𝜔0𝑡 𝑋[𝑘] = ෍ 𝑥[𝑛]ⅇ 𝑘 = 0,1,2 … , 𝑁 − 1
𝑛=0
𝑛=−∞
𝑇ൗ Compare Since F is only discrete, 𝑐𝑛 becomes
2
1 𝑁−1
𝑐𝑛 = න 𝐹 𝑡 ⅇ 𝑖𝑛𝜔0𝑡 ⅆ𝑡 1 2𝜋
𝑖𝑛 𝑁 𝑘
𝑇 𝑥[𝑛] = ෍ 𝑋 𝑘 ⅇ
𝑇 𝑁
−2 𝑘=0

DFT: derived by sampling the Fourier Series representation of a periodic signal at discrete time
• 𝑋[𝑘] represents the frequency spectrum of the signal.
• k is the frequency index (dimensionless frequency, unit = 1)
2𝜋
• Generally use digital frequency ( 𝑁 𝑘, radian per sample)
DFT forcefully assumes that the input signal is periodic, inherent in its mathematical definition.
𝑋 𝑘 = 𝑋 𝑘 + 𝑚𝑁
For page 51-59, the equations are not
Try to prove it yourself 𝑥 𝑛 = 𝑥[𝑛 + 𝑚𝑁] required to be memorized. But you
need to understand the coefficients, the
trends and the phenomena.
Discrete Fourier transform (DFT)
Represent the sampling in Time Domain and Frequency domain (sec and Hz as unit):
Assume sampling time interval of ΔT for a signal 𝑥 𝑛 = 𝑥[𝑡𝑛 ] with length N, the forward and
inverse DFT are
𝑁−1 𝑁−1
Forward 2𝜋𝑘
−𝑖 𝑛∆𝑇
𝑋 𝑓𝑘 = 𝑋 𝑘 = ෍ 𝑥[𝑛]ⅇ 𝑁∆𝑇 = ෍ 𝑥 𝑡𝑛 ⅇ −𝑖2𝜋𝑓𝑘 𝑡𝑛
𝑛=0 𝑛=0 Same thing as last
𝑁−1 𝑁−1 page, with physical
Reverse 1 𝑖
2𝜋𝑘
𝑛∆𝑇 1
𝑥 𝑡𝑛 = 𝑥[𝑛] = ෍ 𝑋 𝑘 ⅇ 𝑁∆𝑇 = ෍ 𝑋 𝑓𝑘 ⅇ 𝑖2𝜋𝑓𝑘 𝑡𝑛 time/frequency as
𝑁 𝑁 variables
𝑘=0 𝑘=0

in sec
in Hz

In practice, the FFT output of any real valued signal is symmetric with respect to y axis, and the
frequencies are presented in terms of positive and negative frequency components:

1 −1 −1
The frequency bins range from 0 to (positive frequencies) and from to (negative frequencies).
2∆𝑇 2∆𝑇 𝑁∆𝑇 52
Discrete Fourier transform (DFT): algorithm

DFT is usually performed through fast Fourier DFT for a discrete, aperiodic function X(k) of N samples
transform (FFT), with the Cooley-Tukey algorithm, at a time interval of ΔT
to reduce the number of operation to (N/2)Log2N. 𝑁−1
2𝜋
Cooley, James W.; Tukey, John W. (1965). "An algorithm for the machine −𝑖𝑛 𝑁 𝑘
calculation of complex Fourier series". Math. Comput. 19 (90): 297–301. 𝑋[𝑘] = ෍ 𝑥[𝑛]ⅇ
𝑛=0

𝑁−1
1 2𝜋
𝑖𝑛 𝑁 𝑘
𝑥[𝑛] = ෍ 𝑋 𝑘 ⅇ
𝑁
𝑘=0

The transform takes N2 operations, very time-

consuming even with computers

Not required in exam but FFT

will be used in exp. 1
Check out the video: The Fast Fourier Transform (FFT):
Most Ingenious Algorithm Ever?
53
DFT spectrum resolution and frequency band width

t0=0 tN-1=(N-1)ΔΤ
Sampling frequency decides the band width n=0 n=N-1
in DFT
1
𝑓𝑠 = ∆𝑇 (max frequency in the DFT spectrum)

Frequency resolution (gap between

neighbour frequencies), is inverse of the
total measurement time 𝑓𝑠 𝑓𝑠
1 1 𝑓 𝑓𝑘 = − 𝑓𝑘 =
2
∆𝑓 = = = 𝑠 𝑓𝑘 = 0
2
𝑁∆𝑇 𝑇𝑡𝑜𝑡𝑎𝑙 𝑁 𝑁 𝑁
𝑘 𝑘 𝑛=− 𝑛=
2 𝑛=0
Sampled frequencies are 𝑓𝑘 = = 𝑓 2
𝑁𝛥𝑇 𝑁 𝑠

The power spectrum represents the power

contained in each frequency bin and is
proportional to the square of the amplitude
𝑋 𝑓𝑘 2
spectrum. It is defined as: 𝑃 𝑓𝑘 = 𝑁

Python based demonstration 54

DFT spectrum withlow resolution

Frequency resolution (gap between

neighbour frequencies), is inverse of
the total measurement time
1 1 𝑓
∆𝑓 = 𝑁∆𝑇 = 𝑇 = 𝑁𝑠
𝑡𝑜𝑡𝑎𝑙
𝑘 𝑘
Sampled frequencies are 𝑓𝑘 = 𝑁𝛥𝑇 = 𝑁 𝑓𝑠

When your total sampling time is short, the

frequency resolution becomes low

• What are the differences between

the two spectra?

55
Aliasing

Signal frequency fixed at 𝒇𝑴 =0.6 Hz = 1/(10s)

Sampling frequency varied from 0.5 Hz to 5 Hz

1
𝑓𝑠 = ∆𝑇 (the highest frequency in FFT spectra)

Aliasing
• A sampled signal can be fitted to a function of a
lower and higher frequency than its true frequency

• The alias frequency appears at fM±fs, which is

shown as 0.6 ± 0.5 Hz =0.1 and 1.1 Hz in the
example

56
Shannon Sampling theorem
Can we recover the real-world analog signal?
Shannon sampling theorem:
1) Yes, any band-limited function can be perfectly recovered from equidistant sampling.
2) 2) Sampling frequency must be at least twice the maximum frequency in the signal so that
information will not be lost in the sampling process, i.e.
𝒇𝑺 > 𝟐𝒇𝑴 (or ΔT < TM/2)
• Sample the signal more frequently (don’t under-sample it).
• Remove the other frequency components not of your interest. (anti-aliasing band filter)

ΔT = 2s Shannon Nyquist Sampling Theorem

𝑓𝑠 = 0.5 𝐻𝑧 𝑓𝑀 = 0.6 𝐻𝑧
Anti- aliasing filter needed but
Power not always easy
𝒇𝑴
𝑓𝑠 𝑓𝑠
DFT
Filter Filter

Frequency (Hz)
0.1 0.6 1.1 57
DFT spectrum with aliasing for a realistic signal

Sampling frequency decides the

band width in DFT
1
𝑓𝑠 = ∆𝑇 (max frequency in the DFT
spectrum)

When our sampling frequency is

low, the signal appears to be
highly different and incorrect

• What are the differences between

the two spectra?

58
Leakage
When the signal frequency is exactly interger multiple Otherwise, leakage
of the frequency bins 𝑓𝑀 = 𝑚Δ𝑓 (m is an integer), the Reconstructed after DFT
signal can be well captured in the DFT spectra Power
Power 𝑓M Original
𝑓𝑀

Frequency (Hz)
Δf
1 Frequency (Hz)
∆𝑓 =
𝑁∆𝑇 When 𝑓M < Δ𝑓, signal cannot be reconstructed
• To suppress leakage, we can intentionally choose
1 𝑚 Power
𝑓𝑀 = 𝑚Δ𝑓 i.e., 𝑇 = 𝑡 ⇒ 𝑡𝑡𝑜𝑡 = 𝑚𝑇𝑀 , use the
𝑀 𝑡𝑜𝑡
𝑓M
measurement time just integer number of the signal
period (if 𝑓𝑀 is a single frequency).
• Or, sample for long time and collect enough
datapoints, so Δf becomes small.
• Make sure 𝚫𝒇 ≪ 𝒇𝒎 (the minimum frequency of Frequency (Hz)
interest). Δf 59
Recap of Fourier transform

Transform the signal as a function of time to • Fourier series, a summation of sin &
a function of frequency cos functions, for any periodic function

• Complex exponential makes Fourier

series more succinct

• Fourier Transform for aperiodic signal

• DFT for discrete signal

Signal

• FFT to speed up the transform

Power Spectrum
Peaks indicate major
Phase shift components for analysis and
frequency domain signal
filtering
60