Spilios D. Fassois

Time-series methods for fault detection and identification in vibrating structures

An overview of the principles and techniques of time-series methods for fault detection, identification and estimation in vibrating structures is presented, and certain new methods are introduced. The methods are classified, and their features and operation are discussed. Their practicality and effectiveness are demonstrated through brief presentations of three case studies pertaining to fault detection, identification and estimation in an aircraft panel, a scale aircraft skeleton structure and a simple nonlinear simulated structure.

Identification of dynamic myoelectric signal-to-force models during isometric lumbar muscle contractions

A 14-muscle myoelectric signal (MES)-driven muscle force prediction model of the L3-L4 cross section is developed which includes a dynamic MES-force relationship and allows for cocontraction. Model parameters are estimated from MES and moments data recorded during rapid exertions in trunk flexion, extension, lateral bending and axial twist. Nine young healthy males participated in the experimental testing. The model used in the parameter estimation is of the output error type. Consistent and physically feasible parameter estimates were obtained by normalizing the RMS MES to maximum exertion levels and using nonlinear constrained optimization to minimize a cost function consisting of the trace of the output error covariance matrix. Model performance was evaluated by comparing measured and MES-predicted moments over a series of slow and rapid exertions. Moment prediction errors were on the order of 25, 30 and 40% during attempted trunk flexion-extensions, lateral bends and axial twists, respectively. The model and parameter estimation methods developed provide a means to estimate lumbar muscle and spine loads, as well as to empirically investigate the use and effects of cocontraction during physical task performances.

Friction Identification Based Upon the LuGre and Maxwell Slip Models

The problem of friction identification within the presliding and sliding regimes is addressed and three identification methods, designated as the LuGre (LG) method, the nonlinear regression (NLR) method, and dynamic nonlinear regression with direct application of the excitation (DNLRX) method, are postulated. The first employs the LG model structure, the second the basic Maxwell slip model structure, and the third an extended version of it. The Maxwell Slip model structure accounts for the presliding hysteresis with nonlocal memory, but is confined to providing constant sliding friction. This limitation is circumvented by the postulated extended version. In all methods identification is based upon signals obtained from a single experiment. The methods are successfully assessed via Monte Carlo experiments, as well as via a laboratory setup. The DNLRX is shown to achieve the best overall performance, followed by the NLR and LG methods. A simple DNLRX-based feedforward friction compensation scheme is also postulated and assessed. The results indicate that it is capable of yielding effective friction compensation.

Maximum likelihood identification of stochastic Weiner-Hammerstein-type non-linear systems

The identification problem for non-linear Wiener-Hammerstein-type systems is considered. Unlike alternative techniques that are based on deterministic system representations, a stochastic model structure that explicitly accounts for both the input-output and noise dynamics is postulated. The uniqueness properties of this structure are analysed, and appropriate necessary and sufficient conditions derived. A new time-domain identification method based on the Maximum Likelihood principle is then introduced. Unlike alternative approaches that are mainly in the frequency and correlation domains, the proposed method offers statistically optimal estimates from a single record of normal operating data, and is capable of operating directly on the time-domain data and overcoming errors associated with the evaluation of correlation functions/Fourier transforms or multi-stage procedures. The effectiveness and accuracy of the proposed method are verified via numerical simulations with a number of different systems and noise to signal ratios.

Duhem modeling of friction-induced hysteresis

In this article we recast the Dahl, LuGre, and Maxwell-slip models as extended, generalized, or semilinear Duhem models. We classified each model as either rate independent or rate dependent. Smoothness properties of the three friction models were also considered. We then studied the hysteresis induced by friction in a single-degree-of-freedom system. The resulting system was modeled as a linear system with Duhem feedback. For each friction model, we computed the corresponding hysteresis map. Next, we developed a DC servo motor testbed and performed motion experiments. We then modeled the testbed dynamics and simulated the system using all three friction models. By comparing the simulated and experimental results, it was found that the LuGre model provides the best model of the gearbox friction characteristics. A manual tuning approach was used to determine parameters that model the friction in the DC motor.

Identification of pre-sliding friction dynamics

The hysteretic nonlinear dependence of pre-sliding friction force on displacement is modeled using different physics-based and black-box approaches including various Maxwell-slip models, NARX models, neural networks, nonparametric (local) models and dynamical networks. The efficiency and accuracy of these identification methods is compared for an experimental time series where the observed friction force is predicted from the measured displacement. All models, although varying in their degree of accuracy, show good prediction capability of pre-sliding friction. Finally, we show that even better results can be achieved by using an ensemble of the best models for prediction.

A Statistical Method for the Detection of Sensor Abrupt Faults in Aircraft Control Systems

Aircraft sensors are important for proper operation and safety, and their condition is conventionally monitored based upon the hardware redundancy principle. In this work a statistical method capable of independently monitoring a single sensor, and thus enhancing reliability and overall system safety, is introduced. The methods main advantages are simplicity, applicability to a wide variety of aircraft operating conditions, the handling of uncertainties, no need for additionally monitored signals, and no need for physics based aircraft dynamics models. The method is based on a statistical time series framework accounting for random effects and uncertainties, and exploits the fact that abrupt faults are characterized by time constants smaller than those of the aircraft. It employs monitored signal nonstationarity removal, signal whitening via novel pooled autoregressive modeling, statistical decision making, as well as electronic spike/glitch removal logic. The method effectiveness is demonstrated within the simulation environment of a small commercial aircraft via test cases and Monte Carlo experiments with abrupt faults occurring in an angle-of-attack sensor.

Presliding friction identification based upon the Maxwell Slip model structure.

The problem of presliding friction identification based upon the Maxwell Slip model structure, which is capable of accounting for the presliding hysteresis with nonlocal memory, is considered. The model structures basic properties are examined, based upon which a priori identifiability is established, the role of initial conditions on identification is investigated, and the necessary and sufficient conditions for a posteriori identifiability are derived. Using them, guidelines for excitation signal design are also formulated. Building upon these results, two new methods, referred to as Dynamic Linear Regression (DLR) and NonLinear Regression (NLR), are postulated for presliding friction identification. Both may be thought of as different extensions of the conventional Linear Regression (LR) method that uses threshold preassignment: The DLR by introducing extra dynamics in the form of a vector finite impulse response filter, and the NLR by relaxing threshold preassignment through a special nonlinear regression procedure. The effectiveness of both methods is assessed via Monte Carlo experiments and identification based upon laboratory signals. The results indicate that both methods achieve significant improvements over the LR. The DLR offers the highest accuracy, with the NLR striking a very good balance between accuracy and parametric complexity.

A polynomial-algebraic method for non-stationary TARMA signal analysis—part I: the method

Abstract Time-dependent autoregressive moving-average (TARMA) models, with parameters belonging to a subspace spanned by pre-defined time-domain functions, offer appropriate representation for a fairly wide class of non-stationary signals. The methods available for their estimation have, however, been limited to the subclass of pure autoregressive (TAR) models. In this paper, a polynomial-algebraic (P-A) method for the estimation of mixed TARMA models is introduced. The method is based upon a skew polynomial operator algebra, through which the AR and MA polynomial operators are related to the model’s inverse function, and through which appropriate filtering operations, essential for the construction of a restricted quadratic approximation of the prediction error (PE) criterion, are introduced. The P-A method is characterized by low computational complexity, no need for initial guess parameter values, and avoidance of local extrema problems associated with PE optimization. The performance characteristics of the P-A method are assessed via numerically synthesized non-stationary signals, while PE-based model refinement is also examined. A first application of the non-stationary TARMA approach and the P-A method to the problem of modeling and prediction of an actual non-stationary engineering signal is also presented, and critical comparisons with non-stationary ARIMA (integrated ARMA) and conventional ARMA methods made. The paper is divided into two parts. The polynomial-algebraic method is introduced in the first, whereas its application to the non-stationary modeling and prediction of an automotive active suspension power consumption signal paradigm is presented in the second.

On the problem of stochastic experimental modal analysis based on multiple-excitation multiple-response data, part II: The modal analysis approach

In this part of the paper the stochastic multiple-excitation multiple-response modal analysis problem is considered. The relationship between the actual structural and noise dynamics and their discrete special-form ARMAX-type representation is studied for each one of the vibration displacement, velocity and acceleration data cases, and a novel and effective modal analysis approach is introduced that, unlike previous schemes, is capable of operating on any one of these types of data records. By accounting for issues such as the required excitation signal type and stochastic model form, algorithmic instability occurrence and other well-known estimation difficulties, model structure estimation and model validation, as well as model reduction and analysis based on the dispersion analysis methodology introduced in the first part of the paper [1], the proposed approach not only overcomes the limitations and drawbacks of current schemes but also constitutes the first comprehensive procedure for stochastic multiple-excitation multiple-response experimental modal analysis. The effectiveness of the approach is demonstrated through numerical experiments with structural systems characterized by well-separated and closely spaced modes, and data records of various lengths and signal-to-noise ratios. Comparisons with the classical frequency domain method and the deterministic eigensystem realization algorithm are also made, and the approach is finally used for the experimental modal analysis of a three-span beam from laboratory data.