Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Mincoef Maxcoef Range: DSP Scaling, A Pid Example: Difference Equations Must Be Scaled To

Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

DSP Scaling, a PID Example: Difference equations must be scaled to

fit into the numerical space of a microprocessor. There are 4 parts to the
scaling problem. First, the dynamic range of the multiply-accumulate,
MAC, instructions must be less than the word length of the
microprocessor. Second, the scaling of the multiply must be defined as
Integer, Fractional or Mixed. Third, the value of the binary word that
represents a unit of input must be decided. Finally the MAC coefficients
must be scaled to get an output with sufficient authority from the integral
term to swing the output over the range of Duty ratio required by the
controller.

MAC dynamic range can be defined as the largest coefficient divided
by the smallest coefficient times the A/D converter range or:
MinCoef
MaxCoef
Range
ADCbits
2 =
For a PID controller[1], the coefficients in continuous time may be
defined as P,I and D for the proportional, Integral and Differential
coefficients. Using standard z transform definitions [2] for integration
and differentiation gives the following coefficients for the z transform
domain.
P K P
coef z
=
IT K I
coef z
=

T
D
K D
coef z
=
Where T is the sample interval and Kcoef is a scaling parameter to be
determined. The dynamic range is then

2
IT
D
I
D
Range
z
z
= =
As an example, using 16 bit signed arithmetic, if
14
2 > Range , and
14
2 <
PREVIOUS
V , and
14
2 >
PRESENT
V , then
PREVIOUS PRESENT
V V overflows and the controller can get stuck with
the output at the maximum value. If dynamic range is too large then
there are 3 choices; limit the output, reduce A/D converter bits or limit
the error signal range. It turns out that limiting the error signal range has
beneficial side effects. Thats because the current charging the output
filter is also limited to:
T
V
C
dt
dv
C I
LIM
LIM
= =
Charge current limiting protects parts from overstress and automatically
gives a soft-start characteristic. Limiting usually does the trick; the next
choice is to carefully examine the A/D converter accuracy. Its quite
possible the vendor has overstated the accuracy so that leaving out
some bits has no impact. Some microprocessors have saturation
limiting built into the MAC instruction. But derivative saturation can lead
to turn on overshoot.

Arithmetic Scaling revolves around the definition of a binary word and
examination of the arithmetic operations on that word. Three word
definitions are possible; Integer, Mixed and Fractional. Figure
scaling_word shows how the bit weighting works.
Binary point
Fractional part, y bits Integer part, x bits
... 4, 2, 1 1/2,1/4,1/8,... sign bit Bit weights

Figure scaling_word, Bit arrangement for 7.8 scaling.

If the number of bits to the left of the binary point are measured using x,
and the bits to the right measured by y; then, x.y =7.8 represents the
scaling shown in the figure. This representation can also define integer
scaling, 15.0 and fractional scaling, 0.15. Notice that the sign bit isnt
included which makes the x.y notation work for both integers and
fractions.

Addition and Subtraction work the same for all scaling methods.
However multiplication works differently. Integer notation is the
commonly accepted method used in programming languages from
assembly language to high level languages such as the C programming
language. The product fits into the originally scaled 16 bit register as
long as it doesnt overflow. The problem for the DSP designer is that
integer multiplication can overflow; while fractional multiplication cannot
overflow. Moreover, coefficient rounding is more pronounced with
integer scaling. Some microprocessors have a fractional mode that
keeps the result in the product accumulator scale as 0.30. The product
register in integer based multiplication scales as 1.30, shifted one bit
right. If one is defined to be 2^y, then we want one * one to be
one. That requires the product register to be shifted right by y bits
following the MAC operations. The result can also be shifted into the
upper accumulator by shifting the result left by (16-y) bits. The shift and
store operation are usually available as single operations in assembly
language so that one would expect saving ACCL(p >> 15) to be
replaced by saving ACCH(p<<1) by a C compiler if the former shift is
out of range.


Coefficient Accuracy: The three methods of arithmetic scaling result in
different coefficient accuracy. The best accuracy is obtained using
fractional scaling. The affect of coefficient errors is to misplace the pole-
zero roots from their specified positions. But these specifications come
from estimating component values that may have tolerances over
temperature, age and operating points as high as 40%.

As you might expect, the definition of the arithmetic scaling has little
effect on the net DSP computational result, at least for SMPS controllers
because the specification accuracy. For more details see how the
problem is solved using each of the 3 word definitions:

http://www.intusoft.com/ DSP/Drawings/MCHP2zLim.DWG

One: Next, it is necessary to define the binary value that represents a
unit input. Its interesting to note, that with mixed scaling, if the quantity
is known as one, then casting one as a double and dividing variables
by one will scale them back to their input units. If one =256, then
placing Verr/256.0 in a debugger window will scale it back into plant
voltage. Input scaling is shown in Figure scaling_input.

Kadc inscale
Kadc*inscale=one

Figure scaling_input, Input scaling broken into 2 parts.

Notice that there are 2 parts to the input error scaling, the A/D scaling
and an inscale multiplier. inscale is a shifting pre-scaler that
increases the input value to one and its inverse is applied after the
computations are complete. Kadc describes how the A/D converter
measurement is processed, including scaling resistors, reference
voltage and whether the ADC is left or right justified. one*FullScale
must be greater than 2^ADCbits in order to retain all of the ADC
precision, or

FullScale
one
ADCbits
2
>

FullScale is measured using the plant output voltage that saturates the
A/D. Next, the coefficient multiplier must be selected to give maximum
coefficient precision and provide the integrator sufficient authority to
make the output controllable.

As the input limit is made smaller, there is more latitude in selection of
inscale and Kcoef. These parameters tend to center the computation
between underflow and overflow.

C Language programming: It is convenient, at least for prototyping, to
do this arithmetic in a high level language. The C programming
language is known to have a close relationship with hardware
architectures. Unfortunately the C language integer types are 15.0 for
int16. The fractional and mixed multiply, accumulate function can be
computed using the C language as follows:

#def i ne PRODUCT( a, x) ( i nt 32) a*x
I nt 16 mac2( const i nt 16 coeof 1, i nt 16 var 1,
const i nt 16 coef 2, i nt 16 var 2,
const i nt 16 ybi t s)

{ / / r et ur ns a scal ed 16 bi t r esul t
/ / f or coef 1*var 1+coef 2*var 2
r et ur n( ( i nt 16) ( ( PRODUCT( coef 1, var 1) +
PRODUCT( coef 2, var 2)
) >> ybi t s
)
)
}


Notice the caste of the coefficient to 32 bits forces the C compiler to
make the product a 32 bit word so that it can be shifted right, back into
the int16 format. This can be made an inline function to maximize
speed. It would be nicer to use C++ and override the arithmetic
functions, but C++ adds extra code layer that slows down DSP
operation and obscures the eventual hand translation into assembly
language.


Selecting Kcoef: Whatever word scaling is chosen, the integrator
should be scaled to operate between 0 and less that 0x7fff. Since the
integrator ramps up slowly, its possible to limit its values in a slower
timing loop. Then selecting a limit value or *0x7fff or 0x628F defines a
slower timing loop requirement. The gain from the integrator output to
the the duty ratio control value is then 1/(Kcoeff*one). With the limit set
at .75*max, the duty ratio max value is 0x628F/one / Kcoeff, or Kcoef >
0x628f/one.

one
f x
K
coef
628 0
>

Then the conditions can be satisfied for any value of one from 1 to
32767; or from fractional through mixed to integer and the same
controller operation is achieved.

The integrator output needs to be set to zero when input voltage falls
below the specified under voltage; setting an under voltage lockout flag,
UVLO, low. That keeps the integrator from locking up to its high limit
when the control loop is unable to provide enough output voltage.
That is,
OUT
MAX
IN
V
D
V
UVLO < = , where is the input voltage
and is the maximum duty ratio and is the desired output.
When UVLO is true, the PWM shuts down.
IN
V
MAX
D
OUT
V

Slow timing loop: A slower interrupt can be selected for the integrator
control by UVLO and saturation. The time is set by

Z LIM
SLOW
I I one
C
T
*
2
13
<
where is the integer value of integrator headroom.
13
2

The function of this interrupt is to limit the integrator and test for UVLO,
resetting the integrator if UVLO is true. Typically, this loop needs to be
evaluated every 10 or 20 T, sampling intervals.


Matrix Solution: The DSP control equations can be expressed using
matrix algebra as shown in Figure scaling_2. Assume there are j states
that need to be evaluated, with k of them having a delay history. The
equations can be arranged as shown with all trivial solutions at the
bottom of the matrix. Let Hn be the history value Hn. Then the j+k by j
sub matrix at the top will have its right hand side equal to zero. After
solving the matrix the Vn values are substituted into the Hn RHS for the
next iteration. There may be more states than history because some of
the states may include input and outputs. The main diagonal is scaled to
be 1 so that there is no divide required in the solution. For large j, the
matrix coefficients should be sparse and non zero values should be
near the main diagonal. Thats equivalent to having a number of blocks
with a single input and output cascaded. If the original matrix had non-
zero coefficients below the main diagonal, then the matrix solves an
algebraic set of simultaneous equations. DSPs can be made to have to
all zero values below the main diagonal by judicious use of backward
euler integration to break up the signal flow. That has the side effect of
adding delays and reducing controller bandwidth.

1
1
1
a12
1
1
1
1
a13 V1
0
0
0
V1p
Vkp
V3
0
0
0
...
...
0
0
0
...
...
0
0 0
0 0 0 0 0 0
0 0 0
0 0 0 0 0 0
0 0
a1j
a2j
a3j
a4j
ajj
a23
a34
... ...
0
0 0
0
V2
...
...
Vj
H1
H3
Hk
...
...
... a1(j+k)
a(j+k)(j+k) ...
...
...
...
...
...
...
...
...

Figure Scaling_2, A matrix solution has RHS(0 thru j)=0

LU decomposition, following the forward substitution gives us exactly
whats needed [2]. Then backward substitution is a multiply accumulate
series for all non zero coefficients followed by division by the main
diagonal value. If mixed precision is used, the main diagonal can be
normalized to unity; eliminating the division. If integer or fractional
scaling is used, the result can be multiplied by a predetermined
constant, formed by dividing the scaling value by the main diagonal
value, then applying the inverse of the scaling value to the outputs. The
solution proceeds from the jth row and j+1 column, summing the
products of the non-zero coefficient with their associated states. An
array of coefficients is made in the order they will be used and a
corresponding array of state-pointers can be made to make maximum
use of the DSP multiply accumulate capability.

const i nt 16 coef [ numRowCoef ] ;
i I nt 16 * var pt r [ numRowCoef ] ;
i nt 16 Vn;
whi l e ( numRowCoef - - )
Vn += *Coef ++ * *( *var pt r ++) ) ;

C compilers will figure this out; but theres always hand coded assembly
language to fall back on.

For Reduced Instruction Set Computers, RISC, it may be necessary to
limit the range of variable index change from one computation to the
next. This can be accomplished by moving the rows with H coefficient
up until they are just below the first coefficient used in that column, and
the moving the column left to place the unit value on the main diagonal.
Such a movement doesnt change the Lower triangle zero condition; but
it tends to cluster coefficients along the main diagonal. Then the varptr
usage shown above is replaced as shown below:

const i nt 16 coef [ numRowCoef ] ;
i I nt 16 of f set [ numRowCoef ] ;
i nt 16 *var pt r ;
i nt 16 Vn;
whi l e ( numRowCoef - - )
Vn += *Coef ++ * *( var pt r + *of f set ++) ;

This form may need some adjustment depending on how the C compiler
does its optimization. If the user identifies states that need a solution,
then unwanted states can be eliminated by matrix manipulation. That
reduces the number of MAC initializations and result storage; making a
faster solution.

Noise: Quantizing of the input and output leads no noise. The RMS
noise is equivalent to the LSB/sqrt(12). The PID controller A/D noise
can be referred to the controller input and combined with the PWM
noise. A/Ds usually have more noise than this theoretical value, its
easy for a vendor to inflate accuracy by adding bits, so you need to
make a few measurements around 0 and at your set point to test the
A/D accuracy. Using a modified PID controller results in the block
diagram as shown in Figure scaling_3. Referring the A/D noise to the
PWM duty ratio results in amplifying A/D noise by the reciprocal of the
power filter gain at the controller bandwidth. Taking the gain at the
controller bandwidth gives a peak value; however, it is observed that
quantizing noise tends to run close to the frequency domain peak.
Moreover, its best to add the terms together rather than use an RSS
approach because these values are no longer in a gausian world. They
may very well tend to sync with one another rather than behave as
independent random variables.
A
H
SUM2
K1
K2
X3
SUM2
K1 =1
K2 =1
SUM2
K1
K2
X4
SUM2
K1 =-1
K2 =1
Onoise
ADCnoise
A =
L*C*s^2+(R1+R2)*C*s+1
R2*C*s+1
* Vi
choose H so that A*H =B/s
then onoise =PWMnoise +ADCnoise/A*(B/(s+B))
and if s =jB, and R2 is small,
. onoise =ADCnoise*.707*L*C*B^2/Vi+PWMnoise
and if R2 is large
. onoise =ADCnoise*.707*L*B/(Vi*R2) +PWMnoise
B =Bandwidth (1/sec)
R2=capacitor ESR
R1=Inductor ESR
L=PWM inductor
C=PWM capacitor
PWMnoise

Figure scaling_3, Calculation of quantizing noise for an ideal controller

The 2 computations form an upper and lower bound for noise without
concern for B with respect to 1/(R2*C). For both cases, increasing B
increases the quantizing noise at the PWM, for the case studies done
here, it runs between 2 and 4 times larger, so that the PWM accuracy
can be 1 or 2 bits less than the effective A/D accuracy. If round off error
occurs anywhere in the computation, additional noise sources must be
added at those points and summed into the resulting PWM noise.







[1] PID control equations

Given a plant described by Figure PID_1,

R1 L
C
R2
Vout
Vin
D
D*Vin

Figure PID_1, the plant model

Its transfer function, Aplant, is:
1 ) (
1
2 1
2
2
+ + +
+
=
Cs R R LCs
Cs R
K A
p plant



Where
C is the capacitor value
L is the inductor value
R1 is the inductor ESR
R2 the capacitor ESR
Kp is the PWM gain =Vin

And a controller that compensates for the plant by matching the plant
poles and zeros, where P,I,D are controller coefficients shown below.

Ds
s
I
P A
comp
+ + =

+
+ +
=
1
1
2
Cs R
s
I
P
s
I
D
s
I
A
s
comp



If the controller cancels the open loop poles and zeros, then the loop
gain G is given by:

s
B
I K A A G
p comp plant
= = =

Then for bandwidth B,

p
K
B
I =
LCI D =
CI R R P ) (
2 1
+ =
If R2, the capacitor ESR is large, it may be necessary to insert a
canceling lag,
1
1
2
+ Cs R
or other compensation in the controller.

[2] z transform PID controller

Using
1
1
1

=
z
T
s
for forward euler integration where T is the sample
time


-Z
-1
Iz


and
T
z
s
1
1

= for differentiation

Z
-1
Dz


then substituting into the PID equation we find the z coefficient become

P P
z
=
IT I
z
=

T
D
D
z
=


[3] Quantizing Noise

Assume the measurement of an input variable is equally likely to be
made anywhere in between A/D bits. Then drawing a straight line
between 1/2LSB to +1/2LSB, the error is


12
1
2 2 /
2 /
2 2

=


ds E
RMS

Where LSB =
Then

12 12
LSB
E
RMS
=

=

This calculation assumes the measured input is uniformly spread across
the measurement space. Unfortunately, in control systems, the noise
tends to synchronize and move to the frequency where gain peaks. For
most controllers thats near the controller bandwidth.

You might also like