• No results found

Fuzzification and membership functions

Literature Review

SIMPLE FL CONTROL SYSTEM

4.2.2 Fuzzification and membership functions

To apply the rules the system uses the membership function that is a graphical

representation o f the magnitude o f participation of each input. It associates a

weighting with each of the inputs that are processed, defines functional overlap

There is a unique membership function associated with each input parameter. The

membership functions associate a weighting factor with values of each input and the

effective rules. These weighting factors determine the degree of influence or degree of

membership (DOM) each active rule has. By computing the logical product of the

membership weights for each active rule, a set o f fuzzy output response magnitudes

are produced.

The shape of membership functions can be different. In general, the triangular is

common (see Figure 4.3), but bell, trapezoidal, haversine and, exponential have been

used. More complex functions are possible but require greater computing overhead to

implement4. MEMBERSHIP FUNCTIONS

I t

Q ) 2 (B v - o O Q . <D ,>s Shoddered Centers

/

Negative Zero Positive Z & P

[4

Width » | EngheeringUnits

(Typically lbs, deg F, or deg/m, ftfsec, etc.)

Fig. 4.3 - The features of a m em bership function

(SOURCE: Mitsuishi T., Endou T., Shidama Y., (2003) - The concept o f fuzzy set

and membership function and basic properties o f fuzzy set operation - Journal

of formalized mathematics, Vol. 12)

The degree of membership (DOM) is determined by plugging the selected input

parameter into the horizontal axis and projecting vertically to the upper boundary of

the membership function(s).

4.2.3 Defuzzification

Once the functions are inferred, scaled, and combined, they are defuzzified into a

crisp output which drives the system.

Before undertaking the defuzzification process for crisp output generation, the logical

products for each rule must be combined or inferred. There are several inference

methods:

The MAX-MIN method tests the magnitudes o f each rule and selects the

highest one. The horizontal coordinate o f the "fuzzy centroid" of the area

under that function is taken as the output. This method does not combine the

effects of all applicable rules but does produce a continuous output function

and is easy to implement.

The MAX-DOT or MAX-PRODUCT method scales each member function to

fit under its respective peak value and takes the horizontal coordinate of the

"fuzzy" centroid of the composite area under the fimction(s) as the output.

Essentially, the member fimction(s) are shrunk so that their peak equals the

magnitude o f their respective function ("negative", "zero", and "positive").

This method combines the influence of all active rules and produces a smooth,

continuous output.

example, if three "negative" rules fire, but only one "zero" rule does,

averaging will not reflect this difference since both averages will equal 0.5.

Each function is clipped at the average and the "fuzzy" centroid of the

composite area is computed.

The ROOT-SUM-SQUARE (RSS) method combines the effects of all

applicable rules, scales the functions at their respective magnitudes, and

computes the "fuzzy" centroid o f the composite area. This method is more

complicated mathematically than other methods, but was selected for this

example since it seemed to give the best weighted influence to all firing rules.

The defuzzification of the data into a crisp output is accomplished by

combining the results of the inference process and then computing the "fuzzy

centroid" of the area. The weighted strengths o f each output member function

are multiplied by their respective output membership function center points

and summed.

Finally, this area is divided by the sum o f the weighted member function

strengths and the result is taken as the crisp output as shown in Figure 4.4.

OUTPUT M E M B E R SH IP FUNCTION

P o sitiv e

-1 0 0 t - 5 0 0 +50 +100

(-63.5%)

P e rc e n t O utput - (-100 to 0 = C o o lin g , 0 to + 1 0 0 = H e ^ in g )

Fig. 4.4 - The horizontal coordinate of the centroid is taken as the crisp output

(SOURCE: Mitsuishi T., Endou T., Shidama Y., (2003) - The concept o f fuzzy set

and membership function and basic properties o f fuzzy set operation - Journal

of formalized mathematics, Vol. 12)

The logical product of each rule is inferred to arrive at a combined magnitude for each

output membership function. Once inferred, the magnitudes are mapped into then-

respective output membership functions, delineating all or part of them.

4.3 Artificial Neural Networks

Artificial Neural Networks (ANNs) are relatively crude electronic models based on

the neural structure of the brain5. The brain basically learns from experience.

Neural networks attempt to produce the human way o f processing data and learning.

They exhibit certain analogies to the way in which arrays o f neurons function in

biological learning and memory. The fundamental blocks are processing units, called

nodes, which can be likened to neurons, and weighted connections, which can be

The basic unit o f neural networks, the artificial neurons, simulate the four basic

functions of natural neurons as shown in Figure 4.5.

Sun T ra n s fe r

Wei a r ts

Fig. 4.5 - A basic Artificial Neuron

(SOURCE: Mehrotra K., Mohan C.K., and Ranka S., (1996) - Elements o f Artificial

Neural Networks — Cloth)

The various inputs to the network are represented by the mathematical symbol, x(n).

Each of these inputs are multiplied by a connection weight. These weights are

represented by w(n). In the simplest case, these products are simply summed, fed

through a transfer function to generate a result, and then output. This process lends

itself to physical implementation on a large scale in a small package.

In currently available software packages these artificial neurons are called

"processing elements" and have many more capabilities than the simple artificial

neuron described above.

Simmadcn F in c tio n Transfer Function Inputs Learning and Recall Schedule Learning C ycle Hyperbolic Tangent Linear Sgmoid etc. And Min etc. Max Sum Outputs

Fig. 4.6 - A model of a “Processing Element”

(SOURCE: Mehrotra K., Mohan C.K., and Ranka S., (1996) - Elements o f Artificial

Neural Networks — Cloth)

In Figure 4.6, inputs enter into the processing element from the upper left. The first

step is for each of these inputs to be multiplied by their respective weighting factor

(w(n)). Then these modified inputs are fed into the summing function, which usually

just sums these products. Yet, many different types o f operations can be selected.

These operations could produce a number o f different values which are then

propagated forward; values such as the average, the largest, the smallest, etc.

Sometimes the summing function is further complicated by the addition of an

activation function which enables the summing function to operate in a time sensitive

way. The output of the summing function is then sent into a transfer function. This

algorithm that takes the input and turns it into a zero or a one, a minus one or a one, or

some other number. The transfer functions that are commonly supported are sigmoid,

sine, hyperbolic tangent, etc. This transfer function also can scale the output or

control its value via thresholds. The result o f the transfer function is usually the direct

output of the processing element.

Finally, the processing element is ready to output the result of its transfer function.

This output is then input into other processing elements, or to an outside connection,

as dictated by the structure of the network.

A typical neural network consists of many inter -connected nodes that are organized

into a sequence of layers. The first layer is called the input layer, and it is responsible

for converting the input data in order to be processed it in the next layer. The next

layer, may be more than one layer, is called hidden layer in which nodes perform the

operations on the data received from the first layer. The output layer produces the

estimation by performing a differently defined mathematical function. Evidently, the

calculated output will not be similar or close to the existing output data in one cycle.

Therefore, by repeating and memorizing with unique weight and performing error

calculation of each cycle, the best estimation result is gained.

INPUT LAYER

HIDDEN LAYER

(there may be several hidden layers)

OUTPUT LAYER

Fig. 4.7 - A Simple Neural Network Diagram

(SOURCE: Hassoun M.H., (1995) - Fundamentals o f Artificial Neural Networks

(Hardcover) - The MIT Press)

Once a network has been structured for a particular application, that network is ready

to be trained.

n

The Operation of a Neural Network is controlled by three properties :

1. The pattern of its interconnections, architecture.

2. Method of determining and updating the weights on the interconnections, training.

3. The function that determines the output of each individual neuron, activation or