Method for fault detection and diagnosis of a rotary machine

0Associated
Cases 
0Associated
Defendants 
0Accused
Products 
4Forward
Citations 
0
Petitions 
2
Assignments
First Claim
1. Method for fault detection and diagnosis of a rotary machine, in particular a balancing machine, wherein a rotor having an imbalance is rotated and excites a vibration in said rotary machine due to the imbalancecaused force, and wherein the rotational speed of said rotor and said vibrations are measured in order to obtain input data quantitative for said rotational speed and said vibrations, the method further comprising:
 assuming that a process of dynamic behavior of said rotary machine can be modeled by a linear system in the unfaulty state;
forming an overdetermined set of linear equations which contains input and output data of the process and unknown states of the assumed linear system;
extracting the number of states needed to accurately model the process by using mathematic operations such as orthogonal or oblique projections to form a matrix of which the rank equals the assumed linear system;
computing singular values by using Singular Value Decomposition for obtaining an approximate indication for the order of the assumed linear system.
2 Assignments
0 Petitions
Accused Products
Abstract
Modem rotary machine production requires builtin fault detection and diagnoses. The occurrence of faults, e.g. increased friction or loose bonds has to be detected as early as possible. Theses faults generate a nonlinear behavior. Therefore, a method for fault detection and diagnosis of a rotary machine is presented. Based on a rotor system model for the faulty and unfaulty case, subspacebased identification methods are used to compute singular values that are used as features for fault detection. The method is tested on an industrial rotor balancing machine.
9 Citations
View as Search Results
Method for the automatic detection and identification of errors in a balancing machine  
Patent #
US 20110067493A1
Filed 08/11/2010

Current Assignee
Schenck Rotec GmbH

Sponsoring Entity
Schenck Rotec GmbH

METHOD FOR CORRECTING PERMANENT CALIBRATION AND FORCEMEASURING BALANCING MACHINE  
Patent #
US 20130186170A1
Filed 12/21/2012

Current Assignee
Schenck Rotec GmbH

Sponsoring Entity
Schenck Rotec GmbH

Method for the automatic detection and identification of errors in a balancing machine  
Patent #
US 8,561,463 B2
Filed 08/11/2010

Current Assignee
Schenck Rotec GmbH

Sponsoring Entity
Schenck Rotec GmbH

Method for correcting permanent calibration and forcemeasuring balancing machine  
Patent #
US 9,057,650 B2
Filed 12/21/2012

Current Assignee
Schenck Rotec GmbH

Sponsoring Entity
Schenck Rotec GmbH

Impending failure and failure annunciation systems for actuators and drive train components  
Patent #
US 20070250246A1
Filed 03/23/2007

Current Assignee
Pendicom LLC

Sponsoring Entity
Pendicom LLC

Method for vibration analysis  
Patent #
US 6,779,404 B1
Filed 05/03/2002

Current Assignee
STRUCTURAL VIBRATIONS SOLUTIONS AS

Sponsoring Entity
Andersen Palle, Brincker Rune

Process and apparatus device for analysis of roller bearings in machines  
Patent #
US 6,553,837 B1
Filed 08/14/2000

Current Assignee
PRFTECHNIK Dieter Busch AG

Sponsoring Entity
PRFTECHNIK Dieter Busch AG

Modal analysis method and apparatus therefor  
Patent #
US 20020183942A1
Filed 05/13/2002

Current Assignee
Centre de recherche industrielle du Qubec

Sponsoring Entity
Centre de recherche industrielle du Qubec

Fault inspection system for rotary machines  
Patent #
US 4,988,979 A
Filed 09/08/1989

Current Assignee
Nippondenso Co. Ltd.

Sponsoring Entity
Nippondenso Co. Ltd.

5 Claims
 1. Method for fault detection and diagnosis of a rotary machine, in particular a balancing machine, wherein a rotor having an imbalance is rotated and excites a vibration in said rotary machine due to the imbalancecaused force, and wherein the rotational speed of said rotor and said vibrations are measured in order to obtain input data quantitative for said rotational speed and said vibrations, the method further comprising:
assuming that a process of dynamic behavior of said rotary machine can be modeled by a linear system in the unfaulty state; forming an overdetermined set of linear equations which contains input and output data of the process and unknown states of the assumed linear system; extracting the number of states needed to accurately model the process by using mathematic operations such as orthogonal or oblique projections to form a matrix of which the rank equals the assumed linear system; computing singular values by using Singular Value Decomposition for obtaining an approximate indication for the order of the assumed linear system.  View Dependent Claims (2, 3, 4, 5)
1 Specification
Applicants claim priority under 35 U.S.C. §119 of European Application No. 07017478.4 filed on Sep. 6, 2007.
The invention relates to a method for fault detection and diagnosis of a rotary machine, in particular a balancing machine, wherein a rotor having an imbalance is rotated and excites a vibration in the rotary machine due to the imbalancecaused force, and wherein the rotational speed of the rotor and the vibrations are measured in order to obtain input data quantitative for the rotational speed and the vibrations.
A method for fault detection and diagnosis of a rotary machine, in particular a balancing machine is provided. A rotor having an imbalance is rotated and excites a vibration in the rotary machine due to the imbalancecaused force. The rotational speed of the rotor and the vibrations are measured in order to obtain input data quantitative for the rotational speed and the vibrations.
In accordance with the method, it is assumed that a process of dynamic behavior of the rotary machine can be modeled by a linear system in the unfaulty state. An overdetermined set of linear equations is formed, which contains input and output data of the process and unknown states of the assumed linear system. The number of states needed to accurately model the process is extracted by using mathematical operations such as orthogonal or oblique projections to form a matrix of which the rank equals the assumed linear system. Singular values are computed by using Singular Value Decomposition for obtaining an approximate indication for the order of the assumed linear state.
Fault detection and diagnosis is increasingly important for modern rotary machines. Currently, mostly limit checking and periodic maintenance cycles are used to detect faults. Sometimes signalbased fault detection is applied. However, these methods mainly rely on the knowledge of experts. This situation can be improved by modelbased fault detection. As more information (proper excitation, process model, several measurements) is used, more accurate fault detection can be performed. Standard rotary systems behave basically linear. In case of specific faults, e.g. sliding friction, loose bonds and motion blocks, linear relations no longer hold. An indication for the nonlinearity can be used to detect these faults. Taking into account the noisy environment the method presented will use subspace approaches to estimate singular values. These singular values can be used to detect these faults.
In order to design a modelbased fault detection and diagnosis system, the dynamic behavior of the rotor system needs to be modeled. In a first step, a general model with two degrees of freedom for stiff rotors is given. For lower rotary speeds a simplified model can be applied.
It is assumed that the rotor is not fully balanced, so that an imbalance force F_{u }and torque M_{u }exist. The rotor is situated on two independent bearing supports. Their movement speeds in horizontal plane are denoted by {dot over (x)}_{1},{dot over (x)}_{2}. Plunger coil sensors are used to measure these speeds, resulting in measurement values {dot over (s)}_{1},{dot over (s)}_{2 }(see also
Rotor and bearings are assumed to be stiff, the ground connection of the two bearing supports is modeled by two springdamper systems. It is assumed furthermore that the sensor force feedback on the rotor movement can be neglected.
m_{r}>>m_{s1 }m_{r}>>m_{s2 }
c_{1}>>c_{s1 }c_{2}>>c_{s2 }
As long as the system stays in normal condition, it can be described by a linear state space system.
{dot over (x)}_{m}(t)=A_{m}x_{m}(t)+B_{m}u_{m}(t) (1)
y_{m}(t)=C_{m}x_{m}(t)
Applying Newton'"'"'s law of motion for a rotary mass it follows
However, this detailed model is not needed for the regarded rotary system. As it is operated in subcritical region, certain simplifications can be made.
Under the assumption that the machine is driven with subcritical rotary speed, i.e. (ω_{r}<<ω_{crit}) where ω_{r }is the actual rotary speed and ω_{crit }is the lower of the two critical speeds according to system (1) it can be assumed that
m_{1}{umlaut over (x)}_{1}<<d_{1}{dot over (x)}_{1}<<c_{1}x_{1 }
m_{2}{umlaut over (x)}_{2}<<d_{2}{dot over (x)}_{2}<<c_{2}x_{2 }
With these simplifications the model reduces to a model of order two:
The system is observable and controllable, the poles are on the stability limit. Inputs and outputs are exchanged in order to match the state space structure. Discretization with small sampling time T_{0 }leads to
{dot over (x)}_{d}(k+1)=A_{d}x_{d}(k)+B_{d}u_{d}(k) (16)
y_{d}(k)=C_{d}x_{d}(k) (17)
This model will be used as basis for fault detection.
The presented method is used to detect two fault states where sliding friction occurs.
If the sensor is not connected properly to the left moving rotor support, the force is propagated via sliding friction. The propagated force F_{Rs}=f({dot over (x)}_{1}−{dot over (s)}_{1}) is modeled as Coulomb dry friction. The sensor dynamics are described by the frequency response G_{s}(w).
The sensor force feedback on the rotor movement is neglected. Thus, the relation is nonlinear in the outputs equation only.
If the left rotor support is not properly connected to the ground, the bearing support socket (mass m_{g}) may move on the ground. FIG. (3) shows the dynamic behavior of this fault state. It is assumed that dry Coulomb friction persists between rotor support socket and the ground.
The dynamics may be described by a linear system with nonlinear feedback according to FIG. (3).
{dot over (x)}_{1}(t)=A_{1}x_{1}(t)+B_{1}u_{1}(t)
y_{1}(t)=C_{1}x_{1}(t) (32)
with
The two described fault states introduce nonlinear behavior into the state space relation, either in the output equation or in the system equation. An approximation of this behavior with the reduced model according to section 2.2 turns out to be inaccurate.
It is assumed that the rotor is not fully balanced. The remaining imbalances cause an imbalance force F_{u }and torque M_{u}. It is assumed that the rotor speed ω_{r }and rotor roll angle φ_{r }are known and the imbalance amplitudes A_{u1}, A_{u2 }and angles φ_{u1}, φ_{u2 }are measured or known. The imbalance force and torque can be modeled as
As an approximate indication for the degree of linearity, common subspacebased methods are wellsuited. The mathematical approach that is used can be described as follows:
 It is assumed that the process can be modeled by a linear state space system in the unfaulty state.
 Given the input and output data of the process, an overdetermined set of linear equations is constructed. The set of equations contains input data, output data and unknown states of the state space system.
 The number of states needed to accurately model the process is extracted by mathematic operations such as orthogonal or oblique projections. The number of observable and controllable states equals the system order.
 The problem of determining what number of states is needed to model the system is transformed to a matrix rank determination. This determination is performed approximately by computing the singular values via Singular Value Decomposition (SVD).
The computed singular values give an approximate indication for the order of the presumed state space system. If the given process strongly obeys the reduced model equations according to section (2.2), the method indicates a process of order two. If nonlinear behavior resides and the linear model does not fit, the indication becomes indistinct.
This subsection briefly describes the computation of the features for linearity indication. The used algorithm is partially known as MOESP (Multivariable OutputError State Space).
The input/output relation is assumed to match a linear state space relation according to equation (44). N samples of inputs and outputs are available.
x(k+1)=Ax(k)+Bu(k)+Bn(k) (44)
y(k)=Cx(k)+Cm(k) (45)
The system is observable and controllable of order n. m(k) and n(k) represent white noise sequences. Matrices A, B, C, D and the states x(k) can be transformed by a regular transformation to
Ā=T^{−1}AT
The measured data is aligned in block Hankel Matrices
with
u(k)=({dot over (s)}_{i}(k){dot over (s)}_{2}(k))^{T }
y(k)=(F_{u}(k)M_{u}(k))^{T }
N number of measurements 2i maximum order that can be indicated. Userchosen.
j=N−2i+1 if all measurements are used.
The matrices contain all available data and therefore all available information. A set of linear equations is formed which contains these Hankel Matrices and the state Vectors x(k). To explain the procedure, the noise influence is set to zero at this stage.
Y_{p}=Γ_{i}X_{p}+H_{i}U_{p }
Y_{f}=Γ_{i}X_{f}+H_{i}U_{f} (51)
X_{f}=A^{i}X_{p}+Δ_{i}U_{p }
To develop this set of equations, following matrices are used:
The state matrices X_{p }and X_{f }are defined analogously to the input/output Hankel Matrices:
X_{p}=(x(0)x(1) . . . x(j−1)) (55)
X_{f}=(x(i)x(i+1) . . . x(i+j−1)) (56)
The set of equations (51) can easily be verified by direct insertion. By removing the unknown states from this set of equations the solution for Y_{f }yields:
The notation † stands for the MoorePenrosePseudoinverse. Equation (57) yields direct information on the linearity indication features. If the model is purely linear, the row space of Y_{f }can be fully described by the row spaces
The row space of a matrix is the space spanned by its row vectors. If a matrix J is of full rank, its row space equals the row space of J_{z }if J=J_{s}J_{z}.
For order extraction many different methods are known. The most common are N4SID, MOESP and CVA. As the underlying system is on the stability limit, the algorithm with the most direct order computation, MOESP, is used. Tests with real data as described in the following have approved this choice. MOESP uses a direct RQ decomposition of aligned Block Hankel Matrices:
From the first part, a matrix β_{r}, with rank=system order is extracted.
The rank of β_{r }equals the number of linear independent vectors in its column space, which means that
rank(β_{r})=rank(Γ_{i})=n (61)
The column space of a matrix is the space spanned by its column vectors. If a matrix J is of full rank, its column space equals the column space of J_{z }if J=J_{s}J_{z}.
A proper method can extract the ‘true’ order of the underlying system. In the case of a faulty, nonlinear behavior, the linearity indication differs from the described model of order two. To extract the matrix rank in an approximate way, Singular Value Decomposition (SVD) is used. The SVD of β_{r }yields 3 matrices U_{1}, S_{1}, V_{1}
U, and V_{1 }are orthogonal matrices. S_{1 }is a diagonal matrix which contains the singular values σ_{i}. In case of a faultfree, not noisy system, n singular values are nonzero while all other singular values are zero.
σ_{1}> . . . >σ_{n}>0 (64)
σ_{n+1}= . . . =σ_{2i}=0 (65)
Under the influence of noise (process noise as well as measurement noise), the SVD no longer yields clear order decisions. To represent noise influences, it will be assumed that the measurements y(k) are contaminated by white noise n(k)=(n_{1}(k)n_{2}(k))^{T}. The excitations u(k) contain noise m(k)=(m_{1}(k)m_{2}(k))^{T}. With
and subsequent formation of M_{p }and M_{f}, equation (51) is enhanced to
Y_{p}=Γ_{i}X_{p}+H_{i}U_{p}+H_{i}N_{p}+M_{p }
Y_{f}=Γ_{i}X_{f}+H_{i}U_{f}+H_{i}N_{f}+M_{f} (66)
X_{f}=A^{i}X_{p}+Δ_{i}U_{p}+Δ_{i}N_{p }
The solution for Y_{f }then yields
and the SVD is performed on
where
β_{i}=2i×2i matrix with rank that has to be estimated
β_{r}=2i×2i matrix with rank n (n=system order)
E_{N}=2i×2i Matrix with full rank (noise representation)
The notation
stands for the orthogonal projection onto the row space of
where only the part lying in the row space of
is considered. In literature, this projection is referred to as ‘Oblique Projection’.
It is shown in [VOdM96] that E_{n}→0 for N→∞.
An infinite number of measurements is not achievable. However, if the SignaltoNoiseRatio (SNR) and the measurement number N is sufficiently high, the system impact on the Singular Values is larger than the noise influence (see [Lju99]) and E_{n}<<β_{r }holds and therefore rank (β_{i})=rank (β_{r}). Normally, the influence of E_{n }is not fully negligible and the Singular Values computation yields
σ_{1}> . . . >σ_{n}>>σ_{n+1}, . . . , σ_{2i} (69)
The singular values representing the system structure dominate, all successive singular values represent the noise influence and are considerably lower.
As test rig, a industrial production machine for rotors with mass m_{r}≈25 kg is used. It is equipped with 2 standard plunger coil sensors. The fundamentation can be adjusted by common clamps. The machine is driven in subcritical rotary speed. The environment consists of normal industrial surrounding, e.g. other machines, noise etc.
Three different runs are analyzed:
 Unfaulty standard run
 Run with sliding friction in sensor connection
 Run with sliding friction in ground connection
The features described in the preceding section are computed for the 3 described states. All tests are done on the same machine with equal bearings and rotor. For each run, a timespan of 5 seconds is examined. The singular values are computed according to the preceding chapter. FIG. (4) gives an overview over the different features. The values are given in logarithmic scale, the standard deviation over all regarded runs is indicated.
The first two Singular Values, which represent the linear behavior, remain nearly equal. The third and fourth singular value refer to model inaccuracies and noise. In case of nonlinear behavior, their values are considerably higher than in the purely linear, unfaulty case. Principally, these values can be used as a feature for the appearance of nonlinear faults in linear systems.
The invention described above presents a subspacebased method to detect the occurrence of nonlinear fault states in linear systems. A rotor system was used as example. The faultfree state as well as two different fault cases have been modeled and tested on an industrial rotor balancing machine. It has been shown that the computed singular values are highly sensible to nonlinear fault states and are wellsuited as features for the occurrence of the considered friction faults.