International Congress on Computational Mechanics and Simulation (ICCMS),
IIT Hyderabad, 10-12 December 2012
Page 1 of 10 Auto-Regressive with eXogenous Input (ARX) based Computational Model for Fault Diagnosis of Spacecraft Systems Dipta Ranjan Roy, Nibir Chakraborty, D. Roy. Mahapatra Department of Aerospace Engineering, Indian Institute of Science, Bangalore, India, 560012 Presenting author: diptaroy2001@yahoo.co.in
ABSTRACT An effective fault diagnostic method should be supported by a robust damage estimation technique, as well as, suitable computational techniques for fault identification, isolation etc. The present work is structured into two components. The first one aims at developing a PID based control system, for a spacecraft nozzle system mounted on a single axis gimbal, which uses a servo actuator. The second part is aimed at developing a dynamic diagnostic system, which incorporates a process model based damage detection and system identification scheme. For computational viability, an AutoRegressive with eXogenous inputARX model is proposed for feature generation as well as anomaly detection, which uses statistical inference rules. 1. Introduction Any physical system can be represented by a single degree of freedom, or a combination of several single degree of freedom, spring-mass-damper system. Whenever a fault is injected into such a system, it brings a change in the system parameters, and we observe a deviation from its normal behavior. By system parameters, we are commonly referring to the physical quantities like mass, stiffness and damping. Thus, the first step towards an effective fault diagnosis of a system is associated with the identification of system parameters, which actually governs its dynamic behavior. In this work, we have tried to establish and validate a process model based system identification methodology, based on input-output data. We have used a very fundamental approach, associated with the construction of the Frequency Response Function (FRF) of a system under investigation and then extracting the system parameters, i.e., mass, stiffness and damping from the FRF. For physical realization and experimental validation of the methodology, we have used a servomotor actuated spacecraft thruster nozzle. For simplicity, we have modeled the spacecraft thruster nozzle as an open-loop PID control system. Then, a recursive filtering or auto- regressive eXogenous input (ARX) based system identification procedure is used to construct the Bode plot (which effectively is also equivalent of finding the FRF of the system). After the FRF has been identified, we used simple mathematical techniques to estimate the system parameters. We used a single degree of freedom spring-mass- damper system to analytically represent our spacecraft thruster nozzle and performed experimental validations to implement the ARX based modeling, for estimation of system parameters. The major advantage of the process is that we can estimate the frequency response of the system, from the experimentally obtained time-domain data. The experimental results are quite satisfactory and show good coherence with the analytical results. Once the system parameters are identified using the given technique, it can be used as a reference state, to identify fault induced state in the system. This accounts for the fault diagnosis part, in context of this work. 2. Transfer Function approach for System Identification The most simplistic approach for identifying a single degree of freedom spring-mass-damper system is by identifying the Transfer Function. This work is based on the verification and validation of system identification using the Transfer Function approach. For a second order spring-mass-damper system, we will assume that that the forcing function is also harmonic in nature, which is expressed in the form,
International Congress on Computational Mechanics and Simulation (ICCMS) IIT Hyderabad, 10-12 December 2012 Page 2 of 10 Substituting the harmonic solutions in the second order differential equation we obtain the equation of motion in the frequency domain, given by
which is entirely an frequency dependent function. Thus, the Transfer Function of the system is given as,
where,
and
3. Estimation of parameter from analytical models
LMS-AMESim is a commercially available software which is used for vibration simulation [1]. In this work, we used it to theoretically analyze and study the vibration characteristics of a single degree of freedom spring-mass-damper system. Secondly, we also used it to generate the Frequency Response Function (FRF) of the system and from the response characteristics, we estimated the parameters (mass, stiffness and damping) of the system. The process is very simple and easy to handle and it gives very accurate results. The complete procedure is described in this section. Later, we used the same procedure to estimate the system parameters for our experimental model. The parameters used for simulation are, Physical Mass 2 Kg Physical Stiffness 6000 kN/m Physical Damping 5 N/(m/s)
The spring-mass-damper representation in the LMS-AMESim software is shown below,
Fig.1 LMS-AMESim block diagram of the spring mass damper system International Congress on Computational Mechanics and Simulation (ICCMS), IIT Hyderabad, 10-12 December 2012 Page 3 of 10 The extreme right block represents the sinusoidal frequency that is imparted to the system. The next block represents the 2-port mass which in turn is connected to a spring-damper system. On running the simulation the time history for the mass, as well as, the Frequency Response Function for the system can be obtained. The frequency response function graph which is used to estimate the physical parameters of the system is shown below.
Fig.2 Frequency Response of the single degree of freedom spring-mass-damper system a. Stiffness Estimation The Transfer Function for the one-degree of freedom spring-mass-damper system is given by,
Putting =0 (corner frequency) and substituting the corresponding value of G(0) from the Frequency Response graph, we can obtain the stiffness of the system as follows,
Here, G(0)= -75.4481 Thus, k= 5291.1355 N/m b. Natural Frequency Estimation Natural frequency (fn) can be determined from the phase response of the spring-mass-damper system. It is the frequency where the phase reversal takes place. Thus, from the phase response graph, it can be determined by extracting the corresponding frequency value where the phase is close to 90 degree.
International Congress on Computational Mechanics and Simulation (ICCMS) IIT Hyderabad, 10-12 December 2012 Page 4 of 10 Thus, n = 55.257 Hz c. Mass Estimation Mass of the system can be directly estimated using the relation
Thus, m = 1.9744 Kg d. Non-dimensional damping co-efficient Estimation The non-dimensional damping co-efficient can be determined by analyzing the difference in response between the corner-frequency and the natural-frequency of the system. The relation, to obtain it graphically is given by,
Thus, =0.0243 e. Damping Estimation Damping of the system can be directly estimated using the relation
Thus, c= 5.3229 N/m/s
4. ARX Modeling Approach
The system identification problem is to estimate a model of a system based on observed input-output data. There are several ways to describe a system to be estimated. The basic single input-single output configuration of a general system, in the time domain is given as,
where, k=1,2,3..........N where u(k) is the input signal, y(k) is the output signal, e(k) is the white noise with variance , v(k) is the additional, immeasurable disturbance (noise) and can be filtered as white noise. Here, z is the shift operator, so that
International Congress on Computational Mechanics and Simulation (ICCMS), IIT Hyderabad, 10-12 December 2012 Page 5 of 10
Thus, it can be said
where g(i) is the weighting function (or impulse response), G(z-1) is called the transfer function operator or the transfer function of the linear system.
To be able to estimate the functions G and H they typically, have to be parameterized, most often as rational functions in the delay operator z-inverse. Let the parameters be the numerator and denominator co-efficient. Thus we give a general parametric model structure:
where A,B,C,D,F are polynomials in the delay operator z-inverse. The polynomials in the delay operators are given by the form,
The structure variables na, nb, nc, nd and nf are called the orders of the respective polynomials. The number nk is the number of delays from input to output. More on the ARX based modeling is dispersed elsewhere in the literature [2], [3], [4], [5] and can be used for further usage. The model based FDI method uses a mathematical model of the plant under investigation and computes the redundancy between the measured values and the analytical value of the plant variable. The resultant difference, also known as the residuals, gives a proper indication of the fault that has incurred in the system [6]. An online identification scheme is also possible using the ARX based model. The online identification of the ARX model uses recursive least-squares (RLS) algorithm based on a Kalman filter approach. The ARX model, in this case, can be expressed as
where, y[k] is the discrete output signal and u[k] is the discrete input signal; measured at a discrete time k. The structure of a recursive algorithm based system identification scheme is given below, International Congress on Computational Mechanics and Simulation (ICCMS) IIT Hyderabad, 10-12 December 2012 Page 6 of 10
Fig 3. Recursive System Identification Scheme. The error of the system output is defined as
where the adaptive model generates a one-step ahead predicted response ) n ( y 1 + ) based on the u(n+1) after adjusting the parametric vector ) n ( w ) based on the error e(n). The goal of the adaptive algorithm is to minimize the cost function J(n) described as
5. Experimental setup for system characterization
An instrumented testbed is designed and assembled to achieve the desired objectives, as discussed in the previous sections. The designed model consists of an accelerometer mounted on the thruster nozzle. The thruster nozzle is in turn fitted to a shaft, where the rotation of the shaft is controlled by a servomotor. The accelerometer gives the magnitude of vibration that occurs due to the torque produced by the servomotor. In this experimental setup, the servomotor acts as the actuator, this produces the necessary angular displacements. S e r v o m o t o r
p o c k e t S h a f t
B e a r i n g Base Plate Adjustable Screw Holes Countersunk Screw Holes Vertical Side Stand S e r v o m o t o r
p o c k e t S h a f t
B e a r i n g Base Plate Adjustable Screw Holes Countersunk Screw Holes Vertical Side Stand
Fig 4. Schematic of the model Fig 5. Schematic of the base plate To keep our experimental model simple, we decided to go with an open-loop loop control. Even though the thruster nozzle is excited by the servomotor torque, we are not using the accelerometer reading to counterbalance the torque produced by the servomotor. That is why we are referring to the system as an open loop control problem. However, there is full scope of using the same system as a closed loop PID control system, where the accelerometer reading International Congress on Computational Mechanics and Simulation (ICCMS), IIT Hyderabad, 10-12 December 2012 Page 7 of 10 can be used as a feedback to stabilize the trajectory of the thruster nozzle. Since we tried to estimate the system parameters using frequency response analysis (Bode plot estimation), another objective was to use a nominal thruster nozzle configuration. The advantage is that the whole system can be represented as a single degree of freedom spring-mass-damper system, which makes it easy to compare the analytical calculations and the experimental results. Thus, we decided to go with the open-loop system configuration, where we used the servomotor as the actuation source, to analyze the vibration characteristics of the thruster nozzle. Hence, the overall objective reduces to a simple System Identification problem, where the plant matrix can be obtained from the input- output values. Once the co-efficient for the system matrix are determined we can easily introduce faults in it to predict and determine its behavior. 6. Model independent System Identification
This section deals with the experimental approach to characterize the system, using auto-regressive and adapting filtering based technique (ARX Model), from input-output data. This accounts for the Model-independent approach in our work. For the input configuration, the measurements were taken without the cantilever beam. In order to achieve this, the accelerometer was mounted on the top of the thruster nozzle. In the next step, i.e. for the output configuration, the cantilever beam was attached to the thruster nozzle body and the accelerometer was mounted on the top of the cantilever beam. The purpose of doing this it that, when the accelerometer is mounted on the tip of the cantilever beam, we can ,mathematically model the whole system as a single degree of freedom, spring-mass- damper system with a tip mass. Once the input and output measurements were properly captured, we used the System Identification toolbox available in MATLAB, to determine the Frequency Response function of the system. Once the FRF of the system was determined, we used it to estimate the system parameters (mass, stiffness and damping) using the methodology described in the previous section. The corresponding input-output response and the sampled data which is used for identification of the system is shown below.
Fig.6 Schematic to capture the input response of the system
Fig 7 Schematic to capture the input response of the system International Congress on Computational Mechanics and Simulation (ICCMS) IIT Hyderabad, 10-12 December 2012 Page 8 of 10 The corresponding real-time response of the system which is acquired by the usage of LABView data acquisition tool and NI-PCX DAQ card. The data is then transferred to MATLAB and the corresponding plots are shown below.
Fig.8 Trapezoidal pulse as Servomotor input (generated using AUDRINO)
Fig.9 Input Response of the Thruster Nozzle (without the cantilever beam)
Fig.10 Output Response of the Thruster Nozzle (without the cantilever beam) International Congress on Computational Mechanics and Simulation (ICCMS), IIT Hyderabad, 10-12 December 2012 Page 9 of 10
Fig. 11 System identification Toolbox in MATLAB
Fig.12 Frequency Response Function of the Experimental Model
The estimated values of the parameters from the Experimental Model are described in the table below. PARAMETRS ESTIMATED VALUES Corner Frequency 1.0003 [Hz] Natural Frequency 4.805456 [Hz] 1 st Peak value (Corner Frequency) 0.01191 [dB] International Congress on Computational Mechanics and Simulation (ICCMS) IIT Hyderabad, 10-12 December 2012 Page 10 of 10 2 nd Peak value (Natural Frequency) 0.0009389 [dB] Non-dimension damping constant () 0.5100
7. References [1] http://nupet.daelt.ct.utfpr.edu.br/_ontomos/paginas/AMESim4.2.0/doc/pdf/manuals/amesim.pdf [2] http://alexandria.tue.nl/extra1/erap/publichtml/8835148.pdf [3] System Identification Toolbox, MATLAB [4] Karel J. Keesman, System Identification-An Introduction, Springer-Verlag London Limited 2011, ISBN 978-0- 85729-521-7, DOI 10.1007/978-0-85729-522-4 [5] Rolf Isermann, Marco Munchhof, Identification of Dynamic Systems-An Introduction with Applications, Springer- Verlag London Limited 2011, ISBN 978-3-540-78878-2, DOI 10.1007/978-3-540-78879-9 [6] Rolf Isermann, Fault-Diagnosis Applications, Model-Based Condition Monitoring: Actuators, Drives, Machinery, Plants, Sensors, and Fault-tolerant Systems, Springer-Verlag London Limited 2011, ISBN 978-3-642-12766-3, DOI 10.1007/978-3-642-12767-0