10. Random Vibrations and Uncertainty

Introduction and Context

Deterministic vibration analysis assumes that forces, properties, and boundary conditions are known functions of time. Many engineering systems, however, operate under loads that cannot be predicted exactly. Road roughness, turbulent wind, wave excitation, combustion irregularity, tool–workpiece interaction, and broadband acoustic pressure are examples of excitations that are better represented statistically. In such cases, the objective is not to predict one exact response history, but to estimate response statistics, probabilities of exceedance, fatigue damage, and reliability.

Random vibration analysis therefore extends the methods of mechanical vibrations from a single time history to an ensemble of possible time histories. The analysis combines dynamics with probability and stochastic-process theory. It is particularly important at the postgraduate level because real design decisions often involve incomplete information: material properties vary, damping is uncertain, manufacturing tolerances exist, and operating loads fluctuate.

A useful distinction must be made between randomness and uncertainty. Randomness describes inherent variability in a phenomenon, such as fluctuating pressure on a structure. Uncertainty describes incomplete knowledge, such as uncertainty in damping or an estimated load spectrum. Both affect the confidence that can be assigned to a design prediction.

Learning Objectives

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

1. distinguish deterministic, random, stationary, ergodic, and nonstationary vibration processes;

2. describe random signals using probability density functions, moments, correlation functions, and power spectral density;

3. calculate spectral responses of linear mechanical systems;

4. explain narrow-band and broad-band random vibration and their relevance to fatigue;

5. formulate Monte Carlo simulation and uncertainty-quantification workflows;

6. interpret probabilistic design, reliability, and failure-probability results;

7. identify modelling, statistical, numerical, and professional risks in random-vibration analysis.

From a Time History to a Stochastic Process

A random vibration signal is represented by a stochastic process, commonly written as X(t). At any particular time, X(t) is a random variable. A possible measured record is one realization of the process; another test performed under nominally identical conditions may produce a different record.

The probability density function (PDF), f_X(x), describes the likelihood of signal values near x. Its cumulative distribution function (CDF) is

F_X(x) = P[X(t) <= x].

The mean or first statistical moment is

m_X = E[X(t)],

and the variance is

sigma_X^2 = E[(X(t)-m_X)^2].

The root-mean-square (RMS) value is

x_rms = sqrt(E[X^2(t)]).

For a zero-mean process, the RMS value equals the standard deviation. The RMS value is not the maximum value; a random signal can exceed several RMS levels, although the probability decreases according to its distribution and duration of observation.

A process is stationary in the strict sense if all joint probability distributions remain unchanged under a shift in time. In engineering work, wide-sense stationarity is more commonly used. A process is wide-sense stationary when its mean is constant and its autocorrelation depends only on time separation, not on absolute time. Thus,

R_X(t1,t2) = R_X(tau), where tau = t2-t1.

A process is ergodic when its statistical properties can be estimated from one sufficiently long realization. This assumption is crucial in testing because engineers generally have one measured record rather than an infinite ensemble. Stationarity and ergodicity are not identical: stationarity concerns the process, whereas ergodicity concerns whether time averages represent ensemble averages.

Nonstationary vibration occurs when statistical characteristics change with time. A vehicle passing over changing terrain, a machine during start-up, and a launch vehicle during ascent are examples. Such signals may require time-frequency analysis, evolutionary spectra, order tracking, or segmentation into approximately stationary intervals.

Correlation and Dependence

The autocorrelation function is defined as

R_X(tau) = E[X(t)X(t+tau)]

for a stationary process. The autocovariance function is

C_X(tau) = E[(X(t)-m_X)(X(t+tau)-m_X)].

For a zero-mean process, autocorrelation and autocovariance are identical. At zero lag,

R_X(0) = E[X^2(t)] = x_rms^2.

The normalized autocorrelation coefficient is

rho_X(tau) = C_X(tau)/C_X(0).

It indicates how strongly values separated by tau are related. Rapid decay of correlation usually indicates a broad frequency content, whereas slowly decaying correlation indicates a more narrow-band process.

For two processes X(t) and Y(t), the cross-correlation function is

R_XY(tau) = E[X(t)Y(t+tau)].

Cross-correlation is useful for identifying time delay, transmission paths, and relationships between input and response. In multi-input vibration systems, correlation between inputs must be retained. Treating correlated inputs as independent can substantially underpredict or overpredict response.

Power Spectral Density and Spectral Descriptions

The power spectral density (PSD) describes how the mean-square value of a stationary signal is distributed with frequency. With angular frequency omega, the two-sided PSD is the Fourier transform of the autocorrelation function:

S_X(omega) = integral from -infinity to infinity of R_X(tau) exp(-i omega tau) d tau.

The inverse relation is

R_X(tau) = (1/2pi) integral from -infinity to infinity of S_X(omega) exp(i omega tau) d omega.

For a one-sided PSD defined over positive frequencies, the mean square is obtained by integrating the PSD over that range:

x_rms^2 = integral from 0 to infinity of G_X(f) df.

Consequently, the area under a PSD curve has units of squared response per unit frequency and equals the variance for a zero-mean process. This provides an important physical interpretation: a PSD is not an amplitude spectrum. Its square root, integrated appropriately, gives an RMS measure.

Common descriptions include white noise, for which PSD is constant over a specified bandwidth; narrow-band noise, concentrated near one frequency; and broadband noise, distributed over a relatively wide frequency range. Ideal white noise over infinite frequency has infinite variance and is not physically realizable. Engineering white-noise specifications always imply a finite bandwidth or an effective system bandwidth.

In experimental work, PSD is commonly estimated using a periodogram or, more reliably, Welch's averaged method. The record is divided into overlapping segments, each segment is windowed, and the resulting spectra are averaged. Segment length controls frequency resolution; averaging reduces variance of the estimate. Window choice, overlap, sampling rate, anti-alias filtering, detrending, and record length all influence the result.

A cross-power spectral density, S_XY(f), describes the frequency-domain relationship between two signals. The magnitude-squared coherence is

gamma^2_XY(f) = |S_XY(f)|^2/[S_XX(f)S_YY(f)].

It lies between zero and one for ideal estimates and indicates the degree to which one signal can be linearly related to another at a frequency. Low coherence may result from noise, nonlinear behaviour, unmeasured inputs, or insufficient averaging.

Spectral Response of Linear Systems

For a linear time-invariant system, the frequency-response function H(omega) relates input and output in the frequency domain. If y(t) is the output and x(t) is the input,

Y(omega) = H(omega)X(omega).

For a stationary input with PSD S_X(omega), the output PSD is

S_Y(omega) = |H(omega)|^2 S_X(omega).

This equation is the central result of linear random-vibration analysis. It shows that a structure acts as a frequency-selective filter. Frequencies near lightly damped resonances can be strongly amplified, while frequencies outside the system bandwidth may be attenuated.

For a single-degree-of-freedom system with mass m, damping c, and stiffness k, subjected to force F(t),

m x-double-dot + c x-dot + kx = F(t).

The force-to-displacement frequency-response function is

H_xF(omega) = 1/[k - m omega^2 + i c omega].

The displacement PSD is therefore |H_xF|^2 S_F. Velocity and acceleration PSDs follow from multiplication by omega^2 and omega^4, respectively, because V=i omega X and A=-omega^2X.

For base excitation, the absolute acceleration response is often more relevant than relative displacement. The appropriate transfer function must be derived from the chosen coordinates; using a force-excitation formula without accounting for base motion is a frequent error.

For a multiple-degree-of-freedom system,

M q-double-dot + C q-dot + Kq = B w(t),

where w(t) is a vector of random inputs. In the frequency domain,

Q(omega) = H(omega)B W(omega).

If S_W is the input spectral-density matrix, the response spectral-density matrix is

S_Q(omega) = H(omega)B S_W(omega) B^T H^H(omega),

where H denotes the conjugate transpose. The diagonal elements give response PSDs, while off-diagonal elements describe correlation between coordinates. Modal analysis can simplify this calculation when modal coupling and non-proportional damping are appropriately treated.

Random Fatigue and Extreme Response

Random vibration can cause fatigue even when the RMS response is modest. Fatigue depends on the number, amplitude, and sequence of stress cycles. A PSD must therefore be converted into an estimate of stress-cycle distribution or damage rate.

For a narrow-band Gaussian process, local maxima and zero crossings can be related to spectral moments. The spectral moments are

m_n = integral from 0 to infinity of f^n G_X(f) df.

The zero-crossing rate and peak rate are functions of ratios such as m_2/m_0 and m_4/m_2. Narrow-band approximations can then estimate the distribution of peaks and combine it with an S–N curve to calculate fatigue damage.

For broader-band processes, narrow-band methods may be inaccurate. Alternative approaches include Dirlik's method, Tovo–Benasciutti corrections, rainflow-cycle estimation from simulated time histories, and direct time-domain fatigue calculations. In a stress-based approach, the expected damage rate can be estimated through a cycle distribution and Miner’s rule:

D = sum_i n_i/N_i,

where n_i is the number of cycles experienced at stress range i and N_i is the corresponding life from the S–N relation. Failure is often approximated when D reaches unity, but the rule is empirical and does not fully represent sequence effects, mean-stress effects, corrosion, plasticity, or load interaction.

Extreme response is distinct from fatigue damage. A high peak may cause yielding or clearance loss even if it occurs only once, whereas many moderate cycles may cause fatigue failure. Gaussian peak assumptions should not be used automatically for impulsive, clipped, nonlinear, or strongly non-Gaussian processes.

Uncertainty Quantification

Uncertainty quantification (UQ) examines how uncertain inputs affect model outputs. It is helpful to distinguish:

• aleatory uncertainty: inherent variability in loads or operating conditions;
• epistemic uncertainty: lack of knowledge about parameters, models, or boundary conditions;
• parametric uncertainty: uncertain values such as damping, stiffness, or material strength;
• model-form uncertainty: inadequacy of the mathematical model itself.

A standard UQ workflow includes defining uncertain variables, assigning probability distributions, generating samples, evaluating the model, and analysing the output distribution. Sensitivity analysis identifies which variables most influence the response.

If a response is approximated by y=g(theta), where theta is a vector of uncertain parameters, first-order propagation gives

Var(y) approximately equals grad g^T Sigma_theta grad g,

when uncertainties are small and the model is locally linear. This method is efficient but unreliable near resonances, discontinuities, strong nonlinearities, or large parameter variations.

The first-order second-moment method uses means, variances, and derivatives. Polynomial chaos expansions approximate the response using orthogonal polynomials of random variables and can be efficient for smooth problems. Bayesian calibration combines prior information with test data to update parameter distributions, producing posterior uncertainty rather than a single fitted value.

Monte Carlo Simulation and Probabilistic Design

Monte Carlo simulation directly samples uncertain inputs and propagates them through the vibration model. A basic procedure is:

1. define the quantity of interest, such as RMS acceleration, peak stress, or fatigue damage;

2. specify distributions and correlations for uncertain parameters;

3. generate a sample of input values or random excitation records;

4. solve the dynamic model for each sample;

5. estimate the output CDF, confidence intervals, failure probability, and sensitivity measures;

6. increase the sample size until the conclusions are stable.

If failure occurs when g(theta) <= 0, the failure probability is estimated by

P_f approximately equals N_f/N,

where N_f is the number of failed simulations among N trials. For small failure probabilities, ordinary Monte Carlo may require an impractically large sample. Importance sampling, subset simulation, Latin hypercube sampling, and surrogate models can improve efficiency.

Reliability is R=1-P_f for the specified event, time interval, and uncertainty model. A reliability statement is incomplete unless these conditions are stated. For example, a component may have high reliability against one-hour RMS exceedance but lower reliability against ten-year fatigue failure.

Reliability-based design optimization (RBDO) seeks a design that minimizes cost or mass while satisfying probabilistic constraints such as

P[stress > allowable] <= target probability.

The design should include safety factors, inspection plans, maintenance assumptions, and uncertainty in both loads and resistance. A deterministic safety factor applied to a mean value does not automatically provide a known reliability level.

Applied Case: Random Vibration of an Electronics Mount

Background

Consider an illustrative case involving an electronic control unit mounted on an industrial machine. The unit is exposed to broadband floor acceleration. A preliminary deterministic analysis used a single RMS acceleration and predicted acceptable displacement, but field units showed intermittent connector failures.

Situation and Problem

The mounting bracket has a dominant mode near 180 Hz, and its damping is uncertain because of joint preload and friction. The measured base acceleration is approximately stationary over normal operation but varies between machines. The engineering questions are whether the response exceeds the connector limit and whether fatigue damage is acceptable over the planned service interval.

Analysis

A set of measured acceleration records is divided into stationary segments. Welch PSD estimates show a broadband background and a broad excitation region near the bracket mode. The input PSD is represented by a distribution rather than a single curve. Modal testing provides a distribution for natural frequency and damping, with damping treated as uncertain and correlated with joint condition.

The bracket model is reduced to a dominant single mode for initial screening. For every Monte Carlo sample, a natural frequency, damping ratio, input PSD scale, and connector stress factor are selected. The response acceleration PSD is computed using the base-excitation transfer function. RMS acceleration and peak estimates are calculated, and stress PSD is converted into an estimated fatigue-damage rate using a broad-band method. Selected high-response cases are then simulated in the time domain and analysed using rainflow counting.

Approach and Solution

The analysis identifies damping as the strongest parameter near resonance, while the input PSD scale controls response away from resonance. The deterministic nominal model lies near the median response but does not represent the upper tail. The design is modified by increasing bracket stiffness moderately, adding a controlled damping interface, and changing the connector support to reduce stress concentration. The revised design is evaluated with the same uncertainty distributions, rather than only with nominal properties.

Outcome

The revised design reduces the upper-percentile RMS response and produces a substantial reduction in predicted fatigue damage. More importantly, the analysis provides a probability of exceeding the connector stress limit under the stated operating conditions. Validation testing is performed using several independently selected records, and model parameters are updated when measured modal properties differ from predictions.

Learning Points

The case demonstrates that an RMS input alone cannot describe a resonant random-vibration problem. Frequency location, damping uncertainty, input correlation, response tails, and fatigue-cycle distribution all matter. It also illustrates that uncertainty analysis should guide design changes and test planning, not merely add a statistical appendix after the design is complete.

Measurement, Computation, and Digital Practice

Random-vibration analysis depends strongly on data quality. Sensors must have suitable bandwidth, calibration, mounting stiffness, and dynamic range. Sampling must satisfy the anti-aliasing requirement. Saturation, clipping, missing data, and electrical noise can distort both PDFs and PSDs.

Digital workflows commonly combine experimental modal analysis, finite-element models, signal-processing software, and probabilistic simulation. Reduced-order models and surrogate models permit thousands of evaluations that would be too expensive using a detailed finite-element transient model. Machine-learning regression can approximate a response surface, but it should be validated over the full design domain and should not replace physical reasoning. Active learning can select new simulations where prediction uncertainty or failure likelihood is high.

Data governance is also important. Test records should retain calibration information, sampling settings, preprocessing decisions, environmental conditions, and version-controlled analysis scripts. Reproducibility is especially important when probabilistic results influence safety decisions.

Challenges, Limitations, and Common Errors

Several errors recur in engineering practice. A PSD may be incorrectly interpreted as a peak spectrum or as an amplitude spectrum. One-sided and two-sided PSD conventions may be mixed, producing a factor-of-two error. Frequency units in hertz and radians per second may be confused. An insufficient record length may hide low-frequency content, while inadequate frequency resolution may obscure a resonance.

Assuming Gaussianity without checking the PDF can underestimate extreme loads. Treating correlated inputs as independent can produce misleading multi-input responses. Using an uncertain damping value without specifying its distribution can make a probabilistic result appear more precise than it is. A finite simulation sample may not contain rare failures, which does not prove that the failure probability is zero.

Linear transfer functions are inappropriate when clearances, friction, impacts, material yielding, nonlinear mounts, or amplitude-dependent damping dominate. In such cases, time-domain stochastic simulation, nonlinear spectral methods, equivalent-linearization methods, or experimental characterization may be required. Fatigue predictions are also limited by uncertainty in S–N data, surface condition, mean stress, manufacturing defects, and damage accumulation rules.

Professional and Ethical Considerations

A probabilistic result is conditional on the assumed model, data, distributions, and time horizon. Reporting only a favourable percentile or omitting uncertainty in damping can mislead decision-makers. Engineers should state assumptions, confidence intervals, validation evidence, and the consequences of model discrepancy. Safety-critical designs may require conservative bounding cases, independent review, traceable computation, and compliance with applicable vibration, structural, and reliability standards.

Uncertainty should not be used to conceal inadequate testing or weak modelling. Nor should a statistically small failure probability be treated as acceptable without considering the severity of failure, detectability, redundancy, and societal consequences. Professional judgement remains necessary when data are sparse or the event of interest is outside the validated range.

Emerging Directions and Best Practice

Current practice is moving toward hybrid digital models that combine physics-based dynamics with measured data and probabilistic updating. Time-varying PSD estimation, operational modal analysis, Bayesian model updating, non-Gaussian simulation, and uncertainty-aware digital twins are increasingly relevant. Reliability-based maintenance uses vibration measurements to update remaining-life estimates rather than relying only on fixed inspection intervals.

A sound engineering workflow should:

• define the response quantity and failure event before selecting a statistical method;
• verify stationarity, distributional assumptions, sampling adequacy, and input correlation;
• use PSD integration to check RMS consistency;
• validate transfer functions against modal tests or operational measurements;
• distinguish aleatory variability from lack of knowledge;
• perform sensitivity analysis before collecting expensive additional data;
• report distributions, not only means or nominal values;
• use rare-event methods when ordinary Monte Carlo cannot resolve the required probability;
• validate important tail predictions experimentally;
• document data, assumptions, software, and decisions for independent review.

Chapter Summary

Random vibration analysis treats excitation and response as stochastic processes rather than as single predictable time histories. Probability distributions describe instantaneous values, while correlation functions describe dependence across time and between signals. The power spectral density distributes mean-square content over frequency and, together with a frequency-response function, permits calculation of linear-system response spectra. Multi-degree-of-freedom systems require spectral-density matrices and input cross-correlations.

Random fatigue requires more than an RMS estimate because damage depends on stress cycles and extremes. Narrow-band approximations, broad-band spectral methods, and time-domain rainflow analysis provide complementary approaches. Uncertainty quantification propagates variability in loads, properties, models, and measurements into response and reliability predictions. Monte Carlo simulation is flexible, while surrogate, importance-sampling, polynomial-chaos, and Bayesian methods improve efficiency or information quality.

The central professional lesson is that probabilistic results are conditional statements. Their credibility depends on data quality, appropriate assumptions, validated dynamics, transparent reporting, and recognition of model limitations.

Key Terms

Stochastic process: A family of random variables indexed by time or another variable.

Realization: One possible time history generated by a stochastic process.

Stationarity: Property in which statistical characteristics do not change with time shift.

Ergodicity: Property allowing time averages of a realization to represent ensemble averages.

Autocorrelation: Statistical measure of dependence between a signal and a time-shifted version of itself.

Cross-correlation: Measure of dependence between two signals at different time lags.

Power spectral density: Distribution of signal mean-square value over frequency.

Spectral moment: Integral of a PSD weighted by a power of frequency.

Coherence: Frequency-dependent measure of linear association between two signals.

Frequency-response function: Complex ratio of output to input in the frequency domain.

Narrow-band process: Random process whose energy is concentrated in a small frequency range.

Broad-band process: Random process with significant energy over a wide frequency range.

Aleatory uncertainty: Inherent variability in a physical process.

Epistemic uncertainty: Uncertainty caused by incomplete knowledge.

Monte Carlo simulation: Repeated random sampling used to propagate uncertainty through a model.

Reliability: Probability that a system performs its required function under stated conditions.

Failure probability: Probability that a specified limit state is violated.

Rainflow counting: Cycle-counting method used to convert irregular stress histories into fatigue cycles.

Reliability-based design optimization: Design optimization subject to probabilistic performance constraints.

Surrogate model: Efficient approximation of an expensive computational model.

Review Questions

Short Answer Questions

1. What is the difference between a realization and a stochastic process?

2. Why is the autocorrelation at zero lag important in random vibration analysis?

3. State the relationship between input PSD, frequency-response function, and output PSD.

4. Distinguish aleatory uncertainty from epistemic uncertainty.

5. Why can an RMS response be insufficient for fatigue assessment?

Long Answer Questions

1. Explain wide-sense stationarity and ergodicity, and discuss their importance in vibration testing.

2. Derive and interpret the spectral response relation for a linear single-degree-of-freedom system.

3. Discuss PSD estimation using Welch’s method, including the effects of record length, windowing, and averaging.

4. Compare narrow-band fatigue methods, broad-band spectral methods, and time-domain rainflow analysis.

5. Describe a complete Monte Carlo uncertainty-quantification workflow for a multi-degree-of-freedom structure.

Analytical / Application Questions

1. A measured acceleration PSD is integrated over frequency to obtain an RMS value that disagrees with the RMS calculated directly from the time record. Identify possible causes and propose a diagnostic procedure.

2. Two base inputs to a flexible structure are strongly correlated near a resonance. Explain how ignoring their cross-spectrum could affect predicted response and design decisions.

3. A bracket has uncertain damping and stiffness, and its natural frequency overlaps a broadband excitation peak. Design a probabilistic analysis and recommend how its results should be validated before release.

22. Multiple Choice Questions with Answers

Question 1

For a zero-mean stationary process, what does the area under its one-sided PSD represent?

A. The maximum possible amplitude

B. The variance or mean-square value

C. The damping ratio

D. The mean crossing rate only

Correct Answer: B. The variance or mean-square value

Question 2

The output PSD of a linear system subjected to a scalar input with PSD Sx is given by:

A. Sy = H Sx

B. Sy = H^2 + Sx

C. Sy = |H|^2 Sx

D. Sy = Sx/|H|^2 for all systems

Correct Answer: C. Sy = |H|^2 Sx

Question 3

Which condition defines a wide-sense stationary process?

A. Its maximum value is constant

B. Its mean is constant and autocorrelation depends only on time separation

C. It contains only one frequency

D. Every realization is identical

Correct Answer: B. Its mean is constant and autocorrelation depends only on time separation

Question 4

Why are cross-spectral densities important in multi-input vibration analysis?

A. They eliminate the need for a dynamic model

B. They describe correlations between inputs

C. They convert PSD into a PDF

D. They apply only to deterministic signals

Correct Answer: B. They describe correlations between inputs

Question 5

A designer uses a nominal damping value near a lightly damped resonance. Which uncertainty is most likely to have a dominant effect on predicted response?

A. Uncertainty in damping

B. Uncertainty in unit conversion only

C. Uncertainty in display resolution

D. Uncertainty in the sign convention only

Correct Answer: A. Uncertainty in damping

Question 6

In ordinary Monte Carlo estimation of failure probability, Pf is estimated as:

A. The number of samples divided by the number of failures

B. The mean response divided by its variance

C. The number of failures divided by the total number of samples

D. One minus the number of failures

Correct Answer: C. The number of failures divided by the total number of samples

Question 7

Which statement best distinguishes fatigue damage from extreme-response assessment?

A. Fatigue depends on cycles, while extreme assessment focuses on limit-exceeding peaks

B. Fatigue uses only the maximum value and extremes use only RMS

C. They are always identical for Gaussian processes

D. Neither depends on frequency content

Correct Answer: A. Fatigue depends on cycles, while extreme assessment focuses on limit-exceeding peaks

Question 8

What is a principal risk of treating correlated vibration inputs as independent?

A. The model becomes deterministic

B. The response may be substantially misestimated

C. The PSD necessarily becomes white noise

D. The system loses all resonances

Correct Answer: B. The response may be substantially misestimated

Question 9

Which method is especially useful when ordinary Monte Carlo cannot efficiently estimate very small failure probabilities?

A. Importance sampling

B. Removing all uncertain variables

C. Using only the mean input

D. Replacing the PSD with its peak frequency

Correct Answer: A. Importance sampling

Question 10

A PSD estimate has poor frequency resolution around a structural resonance. Which change is generally most direct?

A. Use shorter records

B. Increase segment length or total record duration

C. Remove the resonance from the model

D. Set damping equal to zero

Correct Answer: B. Increase segment length or total record duration