Professional Documents
Culture Documents
Neural Networks
Jarmo Lehtonen Heikki N. Koivo
Department of Automation and Systems Technology Department of Automation and Systems Technology
Helsinki University of Technology Helsinki University of Technology
Espoo, Finland Espoo, Finland
jarmo.lehtonen@tkk.fi heikki.koivo@tkk.fi
4 Data
The next step is to look at the data and see if any
regularity can be noted under different load and fault
conditions. The data used in the study is generated using a
FEM (Finite Element Method) model of a 35kW induction
motor. The data is generated for three different load
conditions: no load, half load and full load. The load is
determined with respect to the nominal load present in the
motor nameplate. The voltages used to feed the motor are
obtained from a voltage converter. In some cases the motor
can be fed with sinusoidal input voltage, but this study
centers on motors with converter feed.
4.4 Results
The procedure for using the method is listed below as a
reminder for the reader for how the results are obtained: Table 1. The Bayesian probabilities for the different motor
conditions with no load.
The condition for Data from Data from rotor Data from
The condition for Data from Data from rotor Data from which the classifier healthy faulty motor with 1 stator
which the classifier healthy faulty motor with 1 stator probability is given motor broken bar faulty motor
probability is given motor broken bar faulty motor Healthy condition 0.65 0.64 0.06
Healthy condition 0.41 0.38 0.28 Rotor faulty
Rotor faulty condition
condition with 3 broken bars
& end ring 0.01 0.04 0.0
with 3 broken bars
Stator faulty condition 0.34 0.32 0.94
& end ring 0.37 0.44 0.21
Stator faulty condition 0.22 0.18 0.51
Table 3. The Bayesian probabilities for the different motor
Table 2. The Bayesian probabilities for the different motor conditions with full load.
conditions with half load.
The condition for Data from Data from rotor Data from
which the classifier healthy faulty motor with 1 stator
The condition for Data from Data from rotor Data from
probability is given motor broken bar faulty motor
which the classifier healthy faulty motor with 1 stator
Healthy condition 0.59 0.12 0.05
probability is given motor broken bar faulty motor
Rotor faulty
Healthy condition 0.59 0.26 0.26
condition
Rotor faulty
with 3 broken bars
condition
& end ring 0.09 0.70 0.02
with 3 broken bars
Stator faulty condition 0.31 0.18 0.98
& end ring 0.18 0.71 0.0
Stator faulty condition 0.22 0.03 0.74
The results show that a mild rotor fault cannot be
Table 3. The Bayesian probabilities for the different motor diagnosed properly under no-load and half-load operation.
conditions with full load. The fault diagnosis system cannot make a distinction
between the mild rotor fault and the healthy condition. The
stator fault is diagnosed properly for the no load and half
The condition for Data from Data from rotor Data from load conditions with good probability. For full load the
which the classifier healthy faulty motor with 1 stator
probability is given motor broken bar faulty motor
classification results are proper. This is probably due to the
Healthy condition 0.64 0.02 0.28 fact that load amplifies the fault harmonics in the stator
Rotor faulty current, and therefore the fault indicator is stronger.
condition When tested with validation data the rotor fault was
with 3 broken bars
& end ring 0.08 0.98 0.03
diagnosed with a 0.98 probability in the full load condition
Stator faulty condition 0.28 0.00 0.70 and for the independent testing data the rotor fault is diag-
nosed with a 0.70 probability. The result is good, as it
Next, independent data was used in testing the models. The shows that the rotor fault can be detected gradually to some
neural network models were not taught again. degree even when the model for the fault is created for a
fault of different severity. This only applies for the full load
Table 4. The Bayesian probabilities for the different motor conditions, however.
conditions with no load.
The condition for Data from Data from rotor Data from
5 Conclusions
which the classifier healthy faulty motor with 1 stator
probability is given motor broken bar faulty motor
This paper studied the use of neural networks for model-
Healthy condition 0.36 0.35 0.10 based fault diagnostics of induction motors. The idea was
Rotor faulty to use data provided by accurate FEM simulations of motor
condition operation under different fault and load conditions to create
with 3 broken bars
& end ring 0.37 0.36 0.10
identification based neural network model models which
Stator faulty condition 0.27 0.28 0.80 can be used for online fault diagnosis. The neural network
models are created for each phase of the three phase
Table 5. The Bayesian probabilities for the different motor induction motor, and each load condition is handled
conditions with half load. separately. Most of the faults cases have to be left out due
to
the unavailability of data for different fault conditions. The
data is often a restriction for data-based models, because
each different situation would require its own set of data,
along with data for validation and testing. For this work a
FEM model of a 35 kW induction motor was simulated for
cases of a healthy, rotor faulted and stator faulted motor.
The parameters for the model were obtained from a real simple visual inspection. The classifier results are not
operational motor. No real measurement data was used in completely unanimous, because the probability was divided
the thesis, but studies with FEM simulations of induction between the different models in each condition and load.
motors show that the data is very realistic with respect to The results indicate that as the load increases the faults
the data obtained from real motor measurements, see e.g. characteristics are more evident in the data. The correct
[Arkkio 1990]. Measurement data from a real motor was, conditional classification probabilities increase with the
however, used as an input to the FEM models in the data load.
group that was used to build the time series models. Use of It should be noted that the performance of a fault
real measurement data includes some measurement error detection algorithm is proportional to the effectiveness of
into the data and hence makes it more suitable for testing the fault indicator. Stator current, although being a
the method. FEM simulations are not appropriate for on- noninvasive indicator, is not a very clear indicator.
line fault detection because of excessive computational Effectiveness of different fault indicators is studied for
time, but they can be easily used to simulate vast amounts example in [Negrea 2004]. It is clear that if measured data
of data for different motor operating conditions and faults. is available from some other fault indicator then the
Models were built separately for the data using methods of presented FDI method can be applied in a similar way and
identification in order to get necessary computational speed the better the indicator the better the results. Better
for a possible on-line implementation. indicators for common induction motor faults than stator
A few notes about what has to be taken into consideration current are e.g. circulating currents and fluxes. Preliminary
if the method were to be implemented on a real machine: studies show forces to be good fault indicators also, but
In a practical application an estimate of the motor load they are problematic to measure directly.
would be needed in order to use the correct models for that The faults in the stator current are manifested by certain
particular load condition. This can be obtained e.g. by harmonic frequencies and hence it is important that the
estimating the slip. Also the models here are made for and models are accurate enough. Especially when using
tested only for three discrete load conditions. It was not nonlinear neural network models there is a risk that the
tested how the method works for in-between values of neural network parameter estimation algorithm converges
load. to a local minimum and the model is not the best possible
The validity of the method taught using FEM data would one. Hence it is important to retrain the neural network
have to be proven to work on real machine data as well. models until the model achieves the necessary accuracy.
The models could also be taught using real machine data
from a machine on which different faults are imposed in References
different load conditions. The feasibility of FEM data is
that it can be used to relatively easily generate great [1] M.E.H. Benbouzid, Bibliography on induction motors
amounts of motor data for different operational states faults detection and diagnosis, IEEE Trans. on Energy
(different faults and load conditions) Conversion, Vol. 14 Issue 4, pp. 1065 -1074, Dec. 1999.
Lots of descriptive data is needed for the implementation
and tuning of a neural network model based FDI scheme, if
[2] M.E.H. Benbouzid,, M. Vieira, and C. Theys, In-
it is to be robust under harsh conditions. This is a common
duction motors faults detection and localisation using
factor for all data-based methods.
stator current advanced signal processing techniques,
Steady-state operation is assumed. Hence an algorithm
IEEE Trans. on Power Electronics, Vol. 14, No. 1. pp. 14
should be implemented for steady-state operation. The
22, January 1999.
algorithm should also decide when the motor is in steady-
state. One possibility is to deduct this from the slip and
[3] S. Altug, M-Y. Chow, and J.H. Trussell, Fuzzy in-
motor speed. The model construction, testing, simulation
ference systems implemented on neural architectures for
and fault classification were implemented in a Matlab-
motor fault detection and diagnosis, IEEE Tr. on Industrial
environment. For the creation of neural network based
Electronics, Vol. 46, No. 6, pp. 1069-1079.
identification models a special Matlab-toolbox called
NNSYSID was used. The fault classification was
[4] S.L. Ho and K.M. Lau, Detection of faults in induction
implemented using a classical Bayes classifier, which
motors using artificial neural networks. Electrical machines
calculates the conditional probability for how well a given
and drives, IEEE Conf. publication No. 412. 11-13
model represents the data which is inputted to the FDI
September 1995.
system.
The results indicate that for most of the load cases the
[5] N. Arthur, and J. Penman, Induction machine con-
classification was successful. The highest probability given
dition monitoring with higher order spectra, IEEE Trans.
by the classifier for a model to present the data inputted
on Industrial Electronics, Vol. 47, No. 5, pp. 1031-1041,
to the fault diagnosis system was the criteria for the
Oct. 2000.
classification. The fault characteristics inflicted on the
stator current were reasonably small and hard to see by
[6] S. Moreau, J.C. Trigeassou, G. Champenois and J.P.
Gaubert, Diagnosis of Induction machines: A procedure
for electrical fault detection and localisation, IEEE Int.
Symp. on Diagnostics for Electrical Machines, Power
Electronics and Drives, Sept. 1-3 1999, Gijon Spain, pp.
225-230.