12. Experimental Vibration and Modal Testing

Introduction and Context

Analytical vibration models are indispensable for predicting natural frequencies, mode shapes, resonance amplitudes, and dynamic stresses. Nevertheless, every model contains assumptions concerning geometry, material properties, joints, damping, boundary conditions, and excitation. Experimental vibration testing provides the evidence required to validate those assumptions and to discover dynamic behaviour that a model may omit. It is therefore not merely a means of confirming calculations; it is an engineering method for identifying, diagnosing, and improving mechanical systems.

Experimental vibration work ranges from a single accelerometer measurement on a rotating machine to a complete modal survey of an aircraft, vehicle body, or precision manufacturing platform. The quality of the result depends on the entire measurement chain: sensor selection, mounting, signal conditioning, data acquisition, sampling, excitation, spectral estimation, modal identification, and interpretation. Errors introduced at any stage can produce convincing but physically incorrect results.

This chapter develops a practical and analytical framework for experimental vibration and modal testing. It addresses measurement planning, sensor and acquisition-system selection, sampling and filtering, Fourier analysis, calibration and uncertainty, impact and shaker testing, operational modal analysis, and finite-element model updating.

Learning Objectives

After studying this chapter, the reader should be able to:

1. Select appropriate vibration sensors, excitation devices, and data-acquisition hardware for a test objective.

2. Explain sampling, aliasing, leakage, windowing, filtering, and frequency-resolution requirements.

3. Interpret FFT spectra, frequency-response functions, coherence, and modal parameters.

4. Design and conduct impact, stepped-sine, random, and operational modal tests.

5. Evaluate calibration, uncertainty, noise, sensor loading, and measurement quality.

6. Distinguish experimental modal analysis from operational modal analysis.

7. Diagnose common testing errors and formulate effective corrective actions.

8. Use experimental data to validate and update analytical or finite-element models.

From a Physical Structure to a Measured Signal

A vibration measurement is a representation of a physical quantity, not the quantity itself. An accelerometer converts acceleration into an electrical output; a signal conditioner supplies excitation or amplification; an analogue-to-digital converter samples the voltage; and software estimates spectral or modal quantities. Each conversion has a sensitivity, bandwidth, noise floor, and possible distortion.

The principal sensors are accelerometers, velocity transducers, displacement probes, strain gauges, force transducers, microphones, and non-contact optical or electromagnetic sensors. Piezoelectric accelerometers offer wide frequency range, high stiffness, and low noise, but usually require a charge or voltage amplifier and cannot measure static acceleration. MEMS accelerometers can measure low-frequency and quasi-static motion, consume little power, and support distributed monitoring, although their noise, bias drift, and bandwidth may limit precision. Seismic velocity transducers are useful at low frequencies, while eddy-current probes and laser vibrometers measure displacement or velocity without adding physical mass to the structure.

Sensor selection should begin with the measured variable and frequency range rather than with instrument availability. Important criteria include sensitivity, amplitude range, bandwidth, noise density, transverse sensitivity, environmental resistance, electrical compatibility, mounting method, and mass. An accelerometer attached to a lightweight panel may alter the very mode being measured. This mass-loading effect is particularly significant when the sensor mass is not small compared with the effective modal mass.

Force measurement is essential in experimental modal analysis. An instrumented impact hammer contains a force transducer and often interchangeable tips. A hard tip produces a short-duration force pulse with broad frequency content; a soft tip lengthens the pulse and concentrates energy at lower frequencies. Shaker tests use electrodynamic, hydraulic, or piezoelectric exciters. The shaker must deliver adequate force over the required frequency range without imposing unwanted constraints through its stinger, fixture, or reaction structure.

The Measurement Chain and Data Acquisition

A typical acquisition chain consists of sensor, cable, signal conditioning, anti-alias filter, analogue-to-digital converter, timing system, and analysis software. Voltage-mode piezoelectric sensors commonly use constant-current power, often called an ICP-type interface. Charge-output sensors are advantageous at high temperature but require charge amplifiers and careful cable management. IEPE-type systems simplify field measurement but require compatible input channels.

The dynamic range of an acquisition system must accommodate both the smallest signal of interest and the largest expected transient. If the input range is too small, clipping occurs; if it is too large, quantization and electronic noise may obscure low-level vibration. A 24-bit converter does not automatically provide 144 dB of usable dynamic range because analogue noise, conditioning electronics, grounding, and sensor noise impose practical limits.

Simultaneous sampling is important when phase relationships are required. Multiplexed systems sample channels sequentially, introducing time delays that can corrupt phase measurements unless they are accurately compensated. A common sampling clock should be used for force and response channels in frequency-response measurements. Triggering must also be considered: impact tests often use the hammer signal to define the record start, whereas rotating machinery may require a tachometer or angular reference.

Ground loops, electromagnetic interference, cable triboelectric noise, loose connectors, and poor shielding can contaminate measurements. Cables should be secured so that motion does not create spurious charge, especially with piezoelectric sensors. Before testing, channels should be inspected for bias, overload, open circuits, and implausible phase or coherence behaviour.

Sampling, Aliasing, and Filtering

If a continuous signal is sampled at frequency \(f_s\), frequency components above the Nyquist frequency \(f_N=f_s/2\) can appear in the sampled record at incorrect lower frequencies. This phenomenon is aliasing. Once aliased, the original component cannot be removed by digital processing because its identity has been lost. Anti-alias analogue filters must therefore attenuate out-of-band content before conversion.

The sampling frequency should be selected with reference to the highest frequency of interest, the transition band of the anti-alias filter, and the possibility of unexpected high-frequency excitation. Sampling only slightly above twice the upper analysis frequency is risky because a real filter does not provide infinite attenuation at the Nyquist limit. Oversampling followed by digital decimation is often preferable.

Filtering is useful but potentially dangerous. High-pass filtering can remove drift and quasistatic offsets; low-pass filtering suppresses unwanted high-frequency noise; band-pass filtering isolates a frequency region. However, filters can change amplitude and phase. Zero-phase offline filtering avoids phase shift by processing in forward and reverse directions, but it is not causal and cannot represent real-time behaviour. Filter settings must be documented with the measurement results.

FFT Analysis: Resolution, Leakage, and Windows

The discrete Fourier transform represents a finite time record in the frequency domain. The fast Fourier transform (FFT) is an efficient algorithm for computing it. For a record of duration \(T\), the frequency-bin spacing is approximately

\[\Delta f=1/T.\]

Longer records improve frequency resolution, whereas increasing the sampling frequency alone does not improve resolution unless record length in samples is also increased. Modal tests require sufficient resolution to separate closely spaced modes and to estimate damping reliably.

Spectral leakage occurs when the recorded segment does not contain an integer number of cycles of a periodic component, or when the record truncates a transient. Energy then spreads into adjacent frequency bins. Window functions reduce leakage by smoothly reducing the record at its ends. A rectangular window preserves amplitude for coherent sampled sinusoids but has high sidelobes. Hanning windows are widely used for general vibration spectra and random signals. Flat-top windows offer improved amplitude accuracy for steady sinusoidal measurements but have poorer frequency resolution. Exponential windows are commonly used for impact-response records to reduce the effect of late-time noise.

Averaging reduces random variation in spectral estimates. In impact testing, several impacts may be averaged, provided the structure and boundary conditions remain unchanged. In stationary random testing, averaging independent or partially overlapping segments produces more stable power spectra and frequency-response estimates. Overlap increases the number of averages for a given record length but does not create fully independent data.

The power spectral density (PSD) describes mean-square content per unit frequency, commonly expressed in acceleration squared per hertz. RMS values are obtained by integrating the PSD over the frequency range. Amplitude spectra, PSDs, cross-spectra, transfer functions, and coherence answer different questions and should not be interchanged.

Frequency-Response Functions and Modal Parameters

A frequency-response function (FRF) describes the output response relative to an input as a function of frequency. For force input and acceleration output, the measured FRF is accelerance; displacement/force is receptance, and velocity/force is mobility. With input \(x\) and output \(y\), spectral estimators include

\[H_1(f)=G_{yx}(f)/G_{xx}(f),\qquad H_2(f)=G_{yy}(f)/G_{xy}(f),\]

where \(G\) denotes auto- or cross-spectral density. The \(H_1\) estimator is generally preferred when output noise dominates, while \(H_2\) is more suitable when input noise dominates. Coherence is commonly calculated as

\[\gamma^2(f)=|G_{xy}(f)|^2/[G_{xx}(f)G_{yy}(f)].\]

Coherence near unity indicates a strong linear relationship under the tested conditions, although high coherence does not prove that the test is free from all errors. Low coherence may result from noise, nonlinear behaviour, multiple unmeasured inputs, leakage, inadequate averaging, or poor excitation.

A mode is characterized experimentally by its natural frequency, damping, and mode shape. Resonance peaks indicate candidate natural frequencies, but peak picking alone is unreliable for closely spaced modes, heavily damped modes, or systems with strong residual effects. Curve-fitting methods use complex FRFs and fit mathematical models to estimate modal parameters. Single-degree-of-freedom methods are useful for isolated modes; multiple-degree-of-freedom methods are required when modal overlap is significant.

Mode shapes must be interpreted with attention to measurement coordinates, reference points, phase convention, scaling, and sign. A mode shape displayed with arbitrary normalization is adequate for visual comparison, but model correlation may require mass normalization or consistent scaling. Modal assurance criterion (MAC) compares two mode-shape vectors:

\[MAC(\phi_a,\phi_b)=\frac{|\phi_a^H\phi_b|^2}{(\phi_a^H\phi_a)(\phi_b^H\phi_b)}.\]

A value close to one indicates strong correlation, but MAC can conceal local errors when only a few coordinates are measured. Frequency error, mode-shape correlation, damping discrepancy, and physical plausibility should all be considered.

Planning and Conducting a Modal Test

A defensible test begins with a test specification. It should state the purpose, frequency range, response quantities, boundary conditions, excitation type, number and locations of measurement points, environmental conditions, sampling parameters, acceptance criteria, and safety controls. The expected mode shapes from a preliminary model can guide sensor locations. Nodes should generally be avoided because responses there may be too small to distinguish from noise; locations should also be selected to distinguish competing mode shapes.

In an impact test, the structure is usually instrumented with fixed response sensors and the hammer is applied sequentially at multiple force locations. This roving-hammer arrangement is efficient when the number of response channels is limited. In a roving-response arrangement, the impact location remains fixed while sensors move. The choice affects force consistency, test duration, and the ability to maintain reference measurements.

The force pulse must contain energy across the frequency range of interest. Double hits, an unsuitable hammer tip, or impacts that are too weak can invalidate the FRF. The force spectrum should be examined directly. A hard impact may excite high frequencies but overload a delicate structure; a soft impact may fail to excite high modes.

Shaker testing provides controlled, repeatable excitation. Sine sweeps are intuitive and useful for identifying resonances and nonlinearities. Stepped sine tests allow accurate amplitude and phase measurements at each frequency, but can be time-consuming. Random excitation estimates broad-band response efficiently and is valuable for linear systems. Burst random excitation reduces leakage by allowing decay intervals. Shaker tests require attention to force-control loops, fixture resonances, stinger alignment, structural constraints, and emergency shutdown limits.

Operational modal analysis (OMA) identifies modal properties from responses measured during normal operation, without a measured input. It is valuable for bridges, wind turbines, vehicles, buildings, and large machinery where artificial excitation is impractical. OMA assumes that the operational forces provide sufficiently broad and informative excitation and that the system is approximately linear and stationary over the analysis interval. Because the input is unknown, absolute modal scaling is generally unavailable, and closely spaced modes or harmonic components can be difficult to interpret. Operational spectral analysis, stochastic subspace identification, and frequency-domain decomposition are common approaches.

Calibration, Uncertainty, and Data Quality

Calibration establishes the relationship between a known physical input and the measured output. A sensor certificate typically reports sensitivity, frequency response, and calibration conditions, but the complete measurement uncertainty also includes mounting, cabling, conditioning, digitization, environmental variation, repeatability, and analysis choices. Field calibration with a reference vibration calibrator can reveal sensor or channel problems before and after a test.

Uncertainty should be reported in a manner appropriate to the decision being made. For a resonance frequency, repeatability and frequency resolution may dominate. For damping, record duration, noise, nonlinearities, and curve-fitting method may dominate. For mode shapes, spatial sampling and sensor orientation are important. Significant digits should not imply greater certainty than the experiment supports.

A useful quality-control sequence is: inspect time histories; verify sensor polarity and units; check overload and clipping; examine force and response spectra; inspect coherence; compare repeated measurements; and document all changed settings. A measurement that looks smooth is not necessarily valid. Physical consistency, repeatability, and agreement between independent indicators are stronger evidence.

Applied Case: Illustrative Modal Test of a Machine-Tool Bed

Background

A precision machine-tool bed exhibits excessive vibration during high-speed cutting. A finite-element model predicts the first bending mode at 182 Hz and a torsional mode at 236 Hz, but operators report a strong response near 205 Hz. The objective is to determine the actual modal properties and identify a design modification.

Situation and Analysis

The bed is tested in its installed support condition, with the spindle locked and ancillary masses attached. Eight lightweight accelerometers are placed at points selected from the predicted bending and torsional mode shapes. One reference accelerometer remains fixed while the instrumented hammer is roved across 32 points. The sampling frequency is set well above the 500 Hz analysis limit, and an anti-alias filter is enabled. Several impacts are rejected because of double hits or insufficient force spectrum.

The resulting FRFs show resonances at 176, 207, and 241 Hz. Coherence is high near the first two modes but falls near 241 Hz at several points because the response is small. Curve fitting identifies a bending mode at 177 Hz and a coupled bending-torsional mode at 208 Hz. The predicted 236 Hz mode is shifted to 241 Hz. The unexplained 208 Hz mode is traced to a flexible bolted support bracket omitted from the finite-element model.

Approach and Solution

The model is updated in stages rather than by indiscriminately changing material properties. First, the bracket is added with measured geometry and joint stiffness is represented by calibrated connector elements. Second, uncertain support stiffness is varied within physically credible bounds. The updated model predicts modes at 178, 210, and 239 Hz, with improved MAC values for the measured mode shapes. A redesigned bracket increases local stiffness and separates the coupled mode from the cutting-force harmonic.

Outcome

The revised test and model indicate that the original design problem was not primarily an error in the bed material properties. It was a boundary-condition and joint-modelling error. The redesign reduces the measured response at the operating speed without requiring a heavier bed.

Learning Points

The case illustrates the importance of testing the real support condition, measuring joints and attachments, using coherence as a diagnostic rather than an ornament, and updating uncertain physical parameters in a controlled sequence. It also shows why agreement of one natural frequency alone is insufficient for model validation.

Model Updating and Digital Practice

Model updating seeks parameter values that reduce the discrepancy between analytical and experimental results. Candidate parameters may include joint stiffness, damping, material modulus, density, geometry, boundary stiffness, and added mass. Updating should be constrained by physical knowledge; otherwise, an optimization algorithm may produce mathematically close but physically meaningless parameters.

A typical objective function combines normalized frequency errors, mode-shape discrepancies, and possibly FRF residuals. Sensitivity analysis identifies which parameters have meaningful influence on the measured outputs. Regularization, parameter bounds, and independent validation data help prevent overfitting. The data used for updating should not be the only data used to claim validation.

Digital test systems increasingly automate channel configuration, calibration tracking, synchronized acquisition, spectral estimation, and cloud-based asset histories. Machine-learning methods can classify spectra, detect anomalies, or assist operational modal identification. They should supplement—not replace—physical reasoning. Training-data bias, changing operating conditions, false alarms, and poor explainability are important governance concerns in automated diagnostic systems.

Challenges, Professional Practice, and Common Errors

Common errors include placing sensors at modal nodes, ignoring sensor mass, using insufficient sampling frequency, allowing aliasing, selecting an inappropriate window, accepting low-coherence FRFs, confusing resonance peaks with modes, and comparing unscaled mode shapes without considering coordinate systems. Other failures arise from testing a structure in a boundary condition different from service, omitting attached equipment, or modifying the structure during sensor installation.

Safety is central to vibration testing. Shakers can generate large forces, rotating machinery can eject components, and resonant amplification can cause sudden structural failure. Test plans should specify exclusion zones, emergency stops, force and displacement limits, electrical isolation, and procedures for abnormal response. Data integrity is also a professional responsibility: calibration records, raw data, processing settings, rejected records, and uncertainty statements should be retained so that results are reproducible.

Best practice is to begin with a physical hypothesis, design measurements that can discriminate between competing explanations, acquire raw time histories as well as processed results, repeat critical measurements, and communicate limitations explicitly. Experimental modal testing is strongest when measurement, modelling, and engineering judgement are treated as complementary activities.

Current Developments

Recent developments include dense wireless sensor networks, fibre-optic sensing, scanning laser Doppler vibrometry, distributed MEMS arrays, operational testing under variable conditions, and hybrid experimental–numerical digital twins. Advanced identification methods increasingly combine time-domain and frequency-domain information, while Bayesian model updating represents uncertainty in both measurements and model parameters. These developments improve spatial coverage and automation, but they do not remove the need for sensor validation, appropriate excitation assumptions, and engineering interpretation.

Chapter Summary

Experimental vibration testing links real structural behaviour with analytical prediction. Reliable results require appropriate sensors, adequate dynamic range, synchronized acquisition, anti-alias filtering, suitable sampling, and careful control of leakage and resolution. FFT spectra describe frequency content, while FRFs relate measured response to excitation and support modal identification. Coherence, repeated measurements, calibration, and uncertainty analysis provide essential quality checks.

Impact testing is efficient for many laboratory structures; shaker testing offers controlled excitation; and operational modal analysis is useful when inputs cannot be measured. Modal parameters must be identified using both spectral evidence and physical interpretation. Model updating should modify physically uncertain parameters, such as joint or support stiffness, and should be validated with independent evidence. Safety, data traceability, and honest reporting of limitations are integral parts of professional experimental practice.

Key Terms

Aliasing: Misrepresentation of a frequency component caused by sampling below the required rate.

Anti-alias filter: Analogue low-pass filter that attenuates frequencies liable to alias during digitization.

Accelerance: Frequency-response function defined as acceleration divided by force.

Coherence: Frequency-dependent measure of the linear consistency between input and output signals.

Damping: Mechanism by which vibration energy is dissipated.

FFT: Fast algorithm for computing the discrete Fourier transform.

Frequency resolution: Separation between adjacent spectral frequency lines, approximately the reciprocal of record duration.

Frequency-response function: Complex ratio of output response to input excitation versus frequency.

Leakage: Spreading of spectral energy among frequency bins due to finite or non-coherent records.

Modal analysis: Identification or prediction of natural frequencies, damping, and mode shapes.

Modal assurance criterion: Correlation measure between two mode-shape vectors.

Mode shape: Spatial pattern of relative motion associated with a natural frequency.

Operational modal analysis: Modal identification from response measurements under unknown operational excitation.

Power spectral density: Mean-square signal content expressed per unit frequency.

Sampling frequency: Number of analogue samples acquired per unit time.

Sensor mass loading: Modification of structural dynamics caused by attaching a sensor.

Shaker test: Vibration test using a controlled mechanical excitation device.

Window function: Weighting applied to a finite record to control spectral leakage.

Model updating: Adjustment of analytical-model parameters to improve agreement with measurements.

Review Questions

Short Answer Questions

1. Why is anti-alias filtering required before analogue-to-digital conversion?

2. Define frequency resolution and state how record duration affects it.

3. What information does coherence provide in an FRF measurement?

4. Distinguish impact testing from operational modal analysis.

5. Why can an accelerometer alter the natural frequencies of a lightweight structure?

Long Answer Questions

1. Explain the complete measurement chain from sensor output to FFT spectrum, identifying major sources of error.

2. Discuss leakage, window functions, averaging, and frequency resolution in experimental vibration analysis.

3. Describe a procedure for planning and conducting a roving-hammer modal test.

4. Explain the assumptions, advantages, and limitations of operational modal analysis.

5. Discuss how experimental modal data can be used to update and validate a finite-element model.

Analytical / Application Questions

1. A test has an analysis limit of 800 Hz, but significant machinery vibration may occur at 1.6 kHz. Select a sampling and filtering strategy and justify it.

2. An FRF displays a sharp resonance but coherence falls to 0.35 at the same frequency. Identify possible causes and propose diagnostic checks.

3. A finite-element model predicts modes at 150 and 220 Hz, while testing gives 153 and 198 Hz with poor correlation for the second mode. Formulate a systematic model-updating investigation rather than simply changing Young’s modulus.

22. Multiple Choice Questions with Answers

Question 1

A vibration signal contains a strong component at 1.3 kHz. If it is sampled at 2 kHz without adequate anti-alias filtering, what is the principal risk?

A. The component will be represented at an incorrect lower frequency

B. The sensor sensitivity will automatically increase

C. The frequency resolution will become infinite

D. The signal will become independent of the sampling clock

Correct Answer: A. The component will be represented at an incorrect lower frequency

Question 2

For an FFT record, which change directly improves frequency resolution?

A. Reducing the sensor mass

B. Increasing the record duration

C. Using a harder hammer tip only

D. Increasing amplifier gain without changing the record

Correct Answer: B. Increasing the record duration

Question 3

Which FRF estimator is generally preferred when output noise dominates?

A. H1

B. H2

C. The rectangular window

D. The coherence estimator alone

Correct Answer: A. H1

Question 4

A low coherence value near a resonance may indicate:

A. Perfect linearity under all conditions

B. Noise, leakage, nonlinear behaviour, or unmeasured inputs

C. That the mode shape is necessarily mass normalized

D. That aliasing has been completely eliminated

Correct Answer: B. Noise, leakage, nonlinear behaviour, or unmeasured inputs

Question 5

What is the main purpose of an exponential window in an impact test?

A. To increase sensor sensitivity

B. To reduce the effect of late-time noise in a decaying response

C. To remove the need for force measurement

D. To convert acceleration directly into strain

Correct Answer: B. To reduce the effect of late-time noise in a decaying response

Question 6

Which excitation method is most appropriate when the input force cannot realistically be measured during normal operation?

A. Operational modal analysis

B. Roving-hammer testing only

C. A calibration shaker with no structure

D. A static displacement test

Correct Answer: A. Operational modal analysis

Question 7

A hard impact hammer tip generally produces:

A. A longer pulse concentrated only at very low frequencies

B. A shorter pulse with broader high-frequency content

C. No measurable force spectrum

D. Only a static load

Correct Answer: B. A shorter pulse with broader high-frequency content

Question 8

Why should model updating avoid unrestricted changes to material properties?

A. Material properties never affect natural frequencies

B. It may produce numerical agreement with physically meaningless parameters

C. Finite-element models cannot contain material properties

D. Experimental frequencies are unrelated to model parameters

Correct Answer: B. It may produce numerical agreement with physically meaningless parameters

Question 9

What does a high Modal Assurance Criterion value primarily indicate?

A. Strong correlation between two mode-shape vectors

B. That damping is exactly zero

C. That the sampling frequency is adequate

D. That the input force was measured without noise

Correct Answer: A. Strong correlation between two mode-shape vectors

Question 10

Which practice best supports reproducibility of a modal test?

A. Saving only screenshots of final spectra

B. Changing processing settings without recording them

C. Retaining raw data, calibration records, settings, and rejected-record decisions

D. Reporting only the largest resonance peak

Correct Answer: C. Retaining raw data, calibration records, settings, and rejected-record decisions