WAVELET TRANSFORM AND NEURAL

NETWORKS
FOURIER TRANSFORM:
1) Random stationary signals can be represented as sums of sine wave functions thanks to
FT. If signal has time-variant frequency (as nonstationary signal), FT will be invalid.
Therefore, it is important to make sure whether signal is stationary or not, prior to
processing by FT.
2) It can be explained the Fourier Transform as follows: The signal x(t) is multipled with
exponential term at some certain frequency f, then integrated over time range. Noting that
the exponential term decomposed as real and imaginary harmonic components:. If the
original signal contains an amplitude component of certain frequency f, the signal and
harmonic component coincide, then their product will be
relatively large value. On the contrast, if signal do not contain component of frequency f,
then their product will be zero. In case, signal has amplitude component of frequency f,
however this component of frequency does not dominate on original signal, then product
will give relatively small value. Integration of product over all time range, it also means that
products will be calculated with every value of frequency.
3) The Fourier Transform reveals that whether or not component of certain frequency exits
in original signals, however, FT cannot know In-What-Times these frequency components
occur.

WAVELET TRANSFORM

Wavelet analysis is similar to Fourier analysis in the sense that it breaks a signal down
into its constituent parts for analysis. Whereas the Fourier transform breaks the signal
into a series of sine waves of different frequencies, the wavelet transform breaks the
signal into its "wavelets", scaled and shifted versions of the "mother wavelet".

When analysing signals of a non-stationary nature, it is often beneficial to be able to

acquire a correlation between the time and frequency domains of a signal.

The Fourier transform, provides information about the frequency domain, however time
localised information is essentially lost in the process. The problem with this is the
inability to associate features in the frequency domain with their location in time, as an
alteration in the frequency spectrum will result in changes throughout the time domain.

In contrast to the Fourier transform, the wavelet transform allows exceptional

localisation in both the time domain via translations of the mother wavelet, and in the
scale (frequency) domain via dilations [VIDA91].

The translation and dilation operations applied to the mother wavelet are performed to
calculate the wavelet coefficients, which represent the correlation between the wavelet
and a localised section of the signal.

The wavelet coefficients are calculated for each wavelet segment, giving a time-scale
function relating the wavelets correlation to the signal. This process of translation and
dilation of the mother wavelet is depicted below in Figure

The effect of this shifting and scaling process is to produce a time-scale representation, as
depicted in Figure. As can be seen from a comparison with the STFT, which employs a
windowed FFT of fixed time and frequency resolution, the wavelet transform offers
superior temporal resolution of the high frequency components and scale (frequency)
resolution of the low frequency components.

This is often beneficial as it allows the low frequency components, which usually give a
signal its main characteristics or identity, to be distinguished from one another in terms of
their frequency content, while providing an excellent temporal resolution for the
high frequency components which add the nuance's to the signals
behaviour.

SHORT TIME FOURIER TRANSFORM

WAVELET TRANSFORM

DISCRETE

The dilation function of the discrete wavelet transform can be represented as a tree of low
and high pass filters, with each step transforming the low pass filter as shown in Figure.

The original signal is successively decomposed into components of lower resolution,

while the high frequency components are not analysed any further. The maximum
number of dilations that can be performed is dependent on the input size of the data to be
analysed, with 2N data samples enabling the breakdown of the signal into N discrete
levels using the discrete wavelet transform.

The versatility and power of the DWT can be significantly increased by using its
generalised form, wavelet packet analysis (DWPA). Unlike the DWT which only
decomposes the low frequency components (approximations), DWPA utilises both the
low frequency components (approximations), and the high frequency components
(details). Figure shows the wavelet decomposition tree.

BEARING FAULT DETECTION USING WAVELET TRANSFORM

AND NERUAL NETWORKS - CASE STUDY
Experiments were performed on the single ball bearing. Four sets of data were obtained from the
experimental system: (i) under normal conditions; (ii) inner race faults;
(iii) Outer race faults and (iv) ball fault.

Some insights of Wavelet analysis for signal processing purposes are now briefly presented. The
wavelet analysis has the advantage of better performance for no stationary signals, representing a
time signal in terms of a set of wavelets. They are constituted for a family of functions which are

derived from a single generating function called mother wavelet, although dilation and
translation processes. Dilation is related with size, and it is also known as scale parameter while
translation is the position variation of the selected wavelet along the time axis. This process is
illustrated in Figure.

The vibration of the machine is measured and wavelet transform is used to

obtain the coefficients. These coefficients are then provided to neural
networks to classify the faults.
NEURAL NETWORK FOR FAULT CLASSIFICATION USING WAVELET TRANSFORM

The most popular network used for classification purposes is the multilayer perceptron (MLP). It
was trained with the extracted features; training data was set 75% including the validation
dataset, while the test data represents the 25% of all data. The network architecture is confirmed
by the input layer, which contains 18 neurons; there is only one hidden layer with 10, 20 and 30
neurons, in order to verify the ANNs behavior; and the output layer has 4 neurons, one per each
condition to classify. The net function is commonly linear, but the transfer function used was the
hyperbolic tangent sigmoid. This structure is the most popular one and has been proven to be
able to learn complicated functions.

The choice of training parameters is sometimes critical to the success of the neural network
training process. Unfortunately, the choice of these values is generally problem dependent. There
is no generic formula that can be used to choose these parameter values. Nevertheless, some
guidelines, which are described below, can be followed as an initial trial.

After a few trials, the neural network designer should have enough experience to set the
appropriate parameters for any given problem. In relation with the training design, the targets for
a supervised learning are the followings:
Normal conditions [-1 -1 -1 1]
Inner race faults [-1 -1 1 -1]
Outer race faults [-1 1 -1 -1]
Ball fault [1 -1 -1 -1]
A critical parameter is the speed of convergence, which is determined by the learning coefficient.
In general, it is desirable to have fast learning, but not so fast as to cause instability of learning
iterations. Some previous works have had a successful training with a learning rate of 0.05.
The training process also involves a base algorithm to performance the training iterations. The
updated error back-propagation with momentum method shown is equation was used to train the
networks. The momentum constant used in this case was 0.8.
The stopping criteria are based in the minimum error reached, in this case 5%. The network
training is also limited to 25000 epochs and the validation dataset may affect the training, with a
maximum of 1000 iteration failed.
The initial weights of the neural network play a significant role in the training convergence.
Without a priori information about the final network weights, it is common practice to initialize
all weights with random numbers of small absolute values between [-0.1, 0.1].