POLITECNICO DI MILANO
Automotive and motorsport engineering
Master of Science (MS) degree in Mechanical Engineering
Course of Advanced dynamics of mechanical systems - Prof. Corradi, Dr. La Paglia
A.Y. 2023 – 2024
Group members:
- Erfan Yadegari
- Vishavmeet Singh
- Vera Carvalho
- Pedro Pessoa
- Guilherme Assunção
1
Assignment 1
Experimental Modal Analysis
Contents
Part A: Cantilever Beam ............................................................... 3
1 – Matrix Formulation of the Problem ................................................... 3
2 – Natural Frequencies .......................................................................... 4
3 – Vibration Modes ............................................................................... 4
4 – Frequency response function ............................................................ 5
5 – Identification of modal parameters ................................................... 7
Part B: Experimental Modal Analysis of a light weight rail wheel . 9
1 – Test Result Visualization .................................................................... 9
2 – Mode Shapes .................................................................................. 10
3 – Parameters Prediction..................................................................... 11
2
The purpose of this assignment is to characterize the dynamic behaviour of a one
dimensional system. The assignment is divided into two parts, firstly, analysing a simple
cantilever beam, and secondly, analysing real world data. To conduct such work, a
MATLAB code was developed.
Part A: Cantilever Beam
For this first part, the system under scrutiny is the simple cantilever beam seen below:
Figure 1 - Cantilever Beam's Geometry
Table 1 - Cantilever Beam's Properties
1 – Matrix Formulation of the Problem
In order to translate this system into a matrix formulation, we must first look at the
boundary conditions. From the clamped end and the free end, we find the following
four equations:
𝑑𝑤 𝑑3 𝑤 𝑑2 𝑤
1) 𝑤(0, 𝑡) = 0 ; 2) | = 0 ; 3) 𝐸𝐽 𝑑𝑥 3 | = 0 ; 4) 𝐸𝐽 𝑑𝑥 2 | =0
𝑑𝑥 𝑥=0 𝑥=𝐿 𝑥=𝐿
Using the standing wave condition, the boundary condition equations can be derived
and manipulated into a matrix equation:
𝐻(𝜔)𝑧 = 0
From now on, everything was done on MATLAB.
3
2 – Natural Frequencies
The natural frequencies represent the frequencies at which the system will naturally
vibrate in the absence of external forces.
det[𝐻(𝜔)] = 0 → 𝜔𝑖
The eigenvalues of 𝐻(𝜔) are the natural frequencies of the system. To find the first four
we employed two methods, graphically and computationally (using the fzero function).
The results can be seen below:
Figure 2 - Determinant of H(𝜔)
Natural Frequencies [Hz]
𝜔1 4.5038
𝜔2 28.2249
𝜔3 79.0306
𝜔4 154.8684
Table 2 - Calculated Natural Frequencies
3 – Vibration Modes
The mode shapes represent the characteristic patterns of deformation exhibited by the
structure when it vibrates at its natural frequencies. These mode shapes are important
to understand how a structure responds to dynamic loads or disturbances.
𝜔𝑖 → [𝐻(𝜔𝑖 )]𝑧 (𝑖) = 0 → Φ𝑖 (𝑥)
4
Following the identification of the natural frequencies, their integration into the matrix
𝐻 unveils the system's mode shapes. For the solution to be non-trivial, coefficients
matrix needs to be singular, so the first variable is fixed to 1. The other variables are
discovered and are dependent on the first.
Figure 3 - Mode Shapes of the Cantilever Beam
4 – Frequency response function
The frequency response function (FRF) we can analyze how the system responds to
sinusoidal inputs at different frequencies and points.
Figure 4 - Input and Output graphical explanation
Φ𝑖 (𝑥𝑗 )Φ𝑖 (𝑥𝑘 )
𝑛
𝑚𝑖
𝐺𝑗𝑘 (Ω) = ∑
−Ω + 𝑗2𝜉𝑖 𝜔𝑖 Ω + 𝜔𝑖2
2
𝑖=1
5
With the identified mode shapes of the system, the next step is to script the FRF. In this
process, the value of the damping was fixed at 1% for all natural frequencies found. A
total of four combinations of input (𝑥𝑘 ) and output (𝑥𝑗 ) locations were chosen to
diversify the responses obtained.
Making use of the following equation and the MATLAB function trapz, that allows the
calculation of an integral by the trapezoidal method, the modal mass was calculated for
each vibration mode.
𝐿
𝑚𝑖 = ∫ 𝑚Φ𝑖2 (𝑥)𝑑𝑥
0
The responses to each FRF were plotted in a Bode diagram, allowing a closer look at the
particularities of each combination of inputs and outputs. As the values of 𝑥𝑗 and 𝑥𝑘
change, there can be verified multiple changes in the amplitude and the phase.
Figure 5 - FRF for each input and output pair
Observations:
• In all the plots we can see resonance at the natural frequencies, characterized by
the peaks in the amplitude plots. In last two amplitude plots it is visible two anti-
resonance points in the frequency range of 50 Hz.
• Across the entire set of plots, we have the maximum peaks of resonance in the
early stages.
6
5 – Identification of modal parameters
As the final section in this first part of the assignment, the responses of the four FRF
were taken as stand ins for real world experimental data to develop an identification
code for the modal parameters.
The numerical model for the FRF is based on the fact that, close to each natural
frequency, the real FRF can be approximated by:
𝑛 (𝑖) 𝐿
𝑁𝑈𝑀
𝐴𝑗𝑘 𝑅𝑗𝑘 𝐻
𝐺𝑗𝑘 (Ω) =∑ + + 𝑅𝑗𝑘
−Ω2 + 𝑗2𝜉𝑖 𝜔𝑖 Ω + 𝜔𝑖2 Ω2
𝑖=1
𝑁𝑈𝑀
Defining 𝐺𝑟𝐸𝑋𝑃 (Ω𝑠 ) as the generic element of the experimental FRFs matrix [𝐺𝑗𝑘 ],
corresponding to the 𝑟 column (FRF) evaluated in correspondence of the frequency Ω𝑠 ,
and 𝐺𝑟𝑁𝑈𝑀 (Ω𝑠 ) as the numerical FRF estimation around a certain 𝜔𝑖 , corresponding to
the 𝑟 column (FRF) evaluated in correspondence of the frequency Ω𝑠 , defined by:
𝑛 (𝑖)
𝑁𝑈𝑀 (Ω )
𝐴𝑟 𝑅𝑟𝐿 𝐻
𝐺𝑟 𝑠 = ∑ 2 2
+ 2 + 𝑅𝑟
−Ω 𝑠 + 𝑗2𝜉 𝜔 Ω
𝑖 𝑖 𝑠 + 𝜔 𝑖 Ω 𝑠
𝑖=1
The parameters can be estimated via the least square minimization, so by minimizing the
following error function near each natural frequency found (Ω𝑠 ∈ 𝜔𝑖 ± 0.5 𝐻𝑧 with a
spacing of 0.01 𝐻𝑧):
𝑁 𝑀
𝜖 = ∑ ∑ 𝑅𝑒(𝐺𝑟𝐸𝑋𝑃 (Ω𝑠 ) − 𝐺𝑟𝑁𝑈𝑀 (Ω𝑠 ))2 + 𝐼𝑚(𝐺𝑟𝐸𝑋𝑃 (Ω𝑠 ) − 𝐺𝑟𝑁𝑈𝑀 (Ω𝑠 ))2
𝑟=1 𝑠=1
Before implementing this minimization in MATLAB, it is essential to establish initial
(𝑖)
values for the unknown parameters 𝜔𝑖 , 𝜉𝑖 and 𝐴𝑟 (for ease, the initial values of the
residuals 𝑅𝑟𝐿 and 𝑅𝑟𝐻 are set to 0). This can be easily done using simple single mode
methods for the natural frequency and damping ratio and then just solving for 𝐴 if
𝐺𝑟𝐸𝑋𝑃 (Ω𝑠 ) = 𝐺𝑟𝑁𝑈𝑀 (Ω𝑠 ).
Since the error function 𝜖 non-linearly depends on the unknown parameters, the
minimization should be implemented in MATLAB using the function lsqnonlin. By doing
so, an estimate for the characteristic parameters of the system near the resonance peaks
was found. These results for 𝜔𝑖 and 𝜉𝑖 were plotted and compared to the original
experimental FRF.
7
Figure 6 - Experimental and Numerical Response
Natural Damping Natural Damping
Frequency Ratio Frequency Ratio
[Hz] [%] [Hz] [%]
Mode 1 Mode 2
Model 4.5 1.0 Model 28.2 1.0
Identified 4.5 1.0 Identified 28.2 1.0
Mode 3 Mode 4
Model 79.0 1.0 Model 154.9 1.0
Identified 79.0 1.0 Identified 154.8 1.0
Table 3 - Comparisson of the Experimental and Numerical parameters
It is clear that the identified parameters on every occasion were correctly identified, with
only one natural frequency being off (only by 0.1 Hz). This shows the potential of this
tool in real world scenarios. That leads the way to the second part of the assignment.
8
Part B: Experimental Modal Analysis of a light weight rail wheel
For the second part of this assignment, we had to discover the modal parameters of a
real system through Experimental Modal Analysis. The problem is related to railway
environmental noise, since the natural frequencies of the rail wheel can amplify it,
creating even more noise pollution. To study this, a wheel was suspended through elastic
supports, then was subject to a dynamometric impact hammer applying an axial load.
To measure the values of axial vibrations, 12 sensors (Piezoelectric accelerometers) were
placed, only on one half of the wheels with a spacing of 15 degrees between each one.
1 – Test Result Visualization
The results from this test were compiled into a DATA file (Data.mat). The file has the FRF
and coherence values for a certain range of frequencies ([0: 5000]𝐻𝑧) for the 12
different points, so the first step is to process the data and visualize it in 3 diagrams: one
for the amplitude of the FRF response, one for the phase of the FRF response and one
for the coherence, all varying with frequency.
Figure 7 - Bode and Coherence Diagrams for the rail wheel data
9
Observations:
• Coherence dropping to zero is normal in anti-resonance points.
• An anti-resonance followed by a resonance means that the sensor is not perfectly
in the nodal position because the denominator and the nominator don’t cancel
out each other, but they are very close.
• When the sensor is in a vibration node, the coherence is flat at 1.
2 – Mode Shapes
We can clearly see the four axial modes through the amplitude plot. Therefore, the next
step is to find out an estimation of the frequency of each peak, so the natural
frequencies, and the damping coefficient for the first two peaks. This allows us to then
(𝑖)
to calculate the value for each 𝐴𝑟 , as done in the first part of this assignment. The values
obtained for the first two natural frequencies (the peaks used for the parameters study
later on) were 667 Hz and 1625 Hz and the damping values were 0.008 and 0.00037.
Knowing the values of the natural frequencies, leads to plotting the mode shapes of
each. This was done in two ways. The first is the same representation as the oe seen in
the first part of the assignment, the second is through a polar plot, where the amplitude
is a radius and the phase the angle. These two representations can be seen here below.
Figure 8 - Mode Shapes of the Rail Wheel
10
Figure 9 - Polar Plots for the Mode Shapes of the Rail Wheel
3 – Parameters Prediction
Finally, to make a comparison between the experimental and numerical values of the
first two peaks, we used the prediction algorithm, based on the least square
minimization, previously developed. The results found are plotted and exposed here
below.
11
Figure 10 - Identification of the first mode, seventh sensor
Figure 11 - Identification of the second mode, fifth sensor
An analysis of the graphics shows that the identification for the second mode failed.
Looking at the parameters vector, the values for the Residual values did not change and
so the FRF fails to predict the values further away from the peak.
12
