Solution of the Breakage Matrix Reverse Problem

Aleksandar Fistes, Dušan Rakić, Aleksandar Takači, Mirjana Brdar

PII: S0032-5910(14)00772-4
DOI: doi: 10.1016/j.powtec.2014.08.061
Reference: PTEC 10506

To appear in: Powder Technology

Received date: 28 May 2014

Revised date: 7 July 2014
Accepted date: 22 August 2014

Please cite this article as: Aleksandar Fistes, Dušan Rakić, Aleksandar Takači, Mirjana
Brdar, Solution of the Breakage Matrix Reverse Problem, Powder Technology (2014), doi:
10.1016/j.powtec.2014.08.061

This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Solution of the Breakage Matrix Reverse Problem

Aleksandar Fistes, Dušan Rakić, Aleksandar Takači, Mirjana Brdar*

PT
Faculty of Technology, University of Novi Sad, Bulevar cara Lazara 1, 21000 Novi Sad, Serbia

RI
SC
Abstract

This paper is a complete overview of the reverse problem in the breakage matrix context. The

NU
procedure for calculating the input particle size distribution to a milling operation in order to
MA
obtain the desired output particle size distribution is given. The reverse problem was solved by

the “precision” and “approximation” approaches. For the precision case, a detailed algorithm is
ED

presented, where in every step a precise range for the current output is provided. This approach is

limited with a number of output size fractions that can be freely chosen. On the other hand, the
PT

approximation approach gives the best-fit solution which corresponds to the desired output
CE

particle size distribution. In this case the problem is solved by the orthogonal decomposition and

non-negative least squares method and by semilogarithmic loss function. Applicability of the
AC

proposed methods was confirmed by examples related to wheat milling.

Keywords: breakage matrix, reverse problem, linear programming, least squares problem,

positive linear inverse problem

*
Corresponding author: e-mail: [email protected]; tel: +381214853631; fax: +38121450413
 ACCEPTED MANUSCRIPT

1. Introduction

PT
Over the years, various modeling approaches (such as discrete element method, finite

RI
element method, population balance models, breakage matrix approach, etc.) were used for the

analysis of particle breakage in the comminution processes. Population balance type of modeling

SC
(PBM) is probably most often used to mathematically describe the comminution process (for

NU
examples refer to [1, 2, 3, 4]. PBMs are based on monitoring the evolution of particle size

distribution (PSD) over time. Therefore, continual sampling within the process is required which
MA
is difficult or sometimes essentially impossible to achieve. The breakage matrix approach treats

an entire process as a single breakage event. Although it is less informative than the population
ED

balance approach, it is more applicable to practical experiments [5]. The relationship between the
PT

input size vector (f) and the resulting output size vector (o) is described by the matrix equation:

B  f  o, (1)
CE

or in expanded form:

 b11 b12 b1n   f1   o1 

AC

b b2 n   f 2   o2 
 21 b22  
     
, (2)
     
bm1 bm 2 bmn   f n  om 

where f and o are column vectors of the PSD in the input and output material from a milling

operation, and B is the breakage matrix. The breakage matrix approach is convenient for studies

where PSDs are measured discretely using sieve analysis [6]. Elements of the breakage matrix

B can be determined by milling narrow sized range fractions of inlet material and the obtained

PSDs from the sieve analysis form the corresponding columns of the breakage matrix. Therefore,
 ACCEPTED MANUSCRIPT

coefficients bij represent the weight fraction of the size range j of the output material obtained

by milling the size range i of the input material. Since the elements of the breakage matrix, as

well as the elements of input and output vectors, are weight fractions their values must be within

the unit interval  0,1 . Moreover, the sum of the elements in every column equals 1.

PT
RI
The breakage matrix is associated with the PBM concept and it is effectively a

SC
combination of the selection function and the breakage function. The selection function is a mass

fraction of particles that are selected and broken in time (during the process). The breakage

NU
function is a mass fraction of breakage products from size x that fall below size y , where x  y
MA
[7]. The selection function is encompassed in diagonal elements of the breakage matrix, and the

breakage function is in the off-diagonal elements [5].

ED

Breakage matrix approach was first introduced by Broadbent and Callcot [8, 9, 10] to

describe cone milling of coal, and later used to study milling of stone [11, 12], coal [13] and
PT

wheat [6, 14, 15].

CE

Broadbent and Callcott [8, 9, 10] used the same sieve sizes for both feed and product size

distributions, giving a square breakage matrix (dimension N x N), where the above diagonal
AC

values are equal to zero.

The breakage matrix approach described in Eqs. (1)–(2) is appropriate for modeling

comminution processes in once-through type of mills where retention time of the particles in the

mill is short, such as roller mills or cone mills. One of the prime examples of roller mills

application is wheat flour milling. Applicability of the breakage matrix approach for predicting

the PSD on the first-break roller milling of wheat was confirmed by [14] and [6]. Retention time

of wheat kernels in the roller mills grinding zone is very short. It is measured in m/sec, as it was

shown by [16, 17] using high-speed video system to observe wheat kernel breakage. Wheat flour
 ACCEPTED MANUSCRIPT

milling is a gradual reduction process consisting of sequential and consecutive size reduction and

separation with up to 16 roller milling operations (so called passages). After every grinding step

the ground material is sieved and the undersize material removed before regrinding. This means

PT
that the grinding passages throughout the flour milling process differ in characteristics of the

RI
input material to the rolls, as well as in the set of roll parameters which are adjusted according to

the characteristics of the input material. Although there is number of grinding steps in the

SC
process, theoretically every step can be observed as once-through type of mill with very short

NU
retention time and therefore with its own breakage matrix.

Usually, the PSDs of the feed and milling output occur in different size ranges. Also,
MA
under the compressive crushing forces, it is entirely possible for a wheat kernel to break so as to

produce flattened bran flake with dimensions larger than the kernel from which it came [16, 17].
ED

These facts are not accounted in square and triangular matrices of Broadbent and Callcott. Using
PT

different sized sieves and different number of sieves for inlet and outlet size distributions, the

breakage matrices became non-square and non-triangular.

CE

In a retention-type of mills multiple breakage events are likely to occur and PBM models
AC

are preferred [1, 4]. The discrete linear PBM (DL-PBM) observe comminution process as series

of elementary breakage events during the time. Then Eq. (1) is written as follows

Bk  o(0)  o( k ) , (3)

where o(0) and o( k ) are mass fraction frequency distribution at time n  0 and n  k.

In spite of the advantages of the breakage matrix method, there are some limitations as the

breakage matrix determined for one set of conditions cannot be used for another set of conditions

[2]. Also, Eq. (1) presumes that particle breakage is independent of any inter-particle interactions

(irrespective of the size distribution of surrounding particles). This assumption seems improbable
 ACCEPTED MANUSCRIPT

in a retention-type of mill, especially in dense-phase systems, because it ignores multi-particle

interactions during breakage. Some recent papers of [4, 5, 18, 19] deal with the modified

breakage matrix methodology in order to characterize multi-particle interactions during

PT
breakage. They include the influence of the multi-particle interactions among particles of

RI
different sizes during breakage. The problems of discrete non-linear PBM (DNL-PBM) and

time-continuous non-linear PBM (TCNL-PBM) are introduced and studied in [3, 4, 18, 20]. In

SC
particular, inverse problem and parameters estimation are studied in [18] which define breakage

NU
matrix methodology for characterization of multi-particle interactions in dense-phase systems.

Considering Eq. (1) the output PSD (vector o) can be altered by changing the breakage
MA
matrix (B) or by changing the input PSD (vector f). Usually, the most efficient way to influence

the outlet PSD is by changing the process. In such way the desired outlet PSD is obtained by
ED

changing the breakage matrix (B). The significance of the forward problem and the inverse
PT

problem (parameter identification) is well-established in literature. The other possible approach,

which has been addressed in this paper, is to identify the input PSD that would result in a desired
CE

output PSD. This approach implies that the breakage matrix (B) is constant and therefore it
AC

applies to the same set of milling conditions.

Examples given in the paper are related to roller milling of wheat. Therefore it is assumed

that the particle breakage is independent of any inter-particle interactions. Also, it is presumed

that the PSDs of the input and output occur in different size ranges. The study is carried out as a

general case where a number of inputs and outputs could be different and the elements of the

breakage matrix are arbitrary values between 0 and 1.

Fistes et al. [21] started the study of the reverse problem in the breakage matrix context

in. This could be useful in milling operations where the possibility of controlling PSD of the
 ACCEPTED MANUSCRIPT

input material exists. In that paper, the procedure for calculating the input PSD in a milling

operation, in order to obtain the desired output PSD, is given. This is done by the transformation

of the matrix Eq. (1) using standard Gaussian elimination techniques. Considering the number of

PT
input (n) and output (m) size fractions three different cases (m=n, m>n and m<n) were studied.

RI
The case with more output than input size fractions (m>n) seems to be the most interesting case

not only mathematically but also for its potential for practical application. In this case with the

SC
overdetermined set of linear equations, the values of n-1 outputs are precisely defined, which

NU
then dictate the remaining m-n output fractions. However, using this technique it is still possible

to specify an output PSD that is unachievable with positive input values. To prevent that, the
MA
problem is interpreted as a system of linear inequalities and solved by a linear programming

method. Acceptable outputs are the points in the interior of the convex polyhedron. Moreover,
ED

vertices of the polyhedron are the columns of the breakage matrix.

PT

The other possible approach is to find a best-fit solution which corresponds to desired m

output values. In the case of overdetermined set of linear equations associated by Eq. (1), the
CE

problem could be observed in the linear least square sense [22, 23, 24, 25, 26, 27]. These
AC

problems are usually solved by orthogonal decomposition methods. For example, authors in [28]

proposed that in their study on reconstruction of bubble size distributions from slices could be

solved by singular value decomposition (SVD) method.

Thus, there are two possible ways for solving the reverse problem, namely “precision”

and “approximation”. Both of them have been extended and thoroughly discussed in this paper.

In the “precision” approach, instead of choosing n-1output simultaneously it is possible to

set outputs one at the time, where in every step a precise range for the current output is provided

[21]. The order of the outputs could be defined by the user’s preferences. This approach leads to
 ACCEPTED MANUSCRIPT

an algorithm (given in Section 2.) hence the problem could easily be implemented. For

dimension n=4 a geometrical interpretation of the procedure is illustrated by an example.

In the “approximation” approach different numerical methods could be used such as: LU

PT
and Cholesky decomposition, modified Gram-Schmidt orthogonalization, SVD and QR

RI
decomposition with many modifications [22, 23, 25]. It should be noted that although the

elements of the breakage matrix and chosen output PSD (weight fractions of output size

SC
fractions) are positive, the solution (weight fractions of input size fractions) obtained by the least

NU
square methods could be negative. Positive linear inverse problems are intensively investigated,

usually in a strongly mathematical context, with many interesting applications [29, 30, 31, 32].
MA
Negative solutions could be avoided by imposing additional conditions or by using some other

method instead of the least square. First option is known as non-negative least squares problem
ED

(NNLS) with a well known algorithm [23]. Also, NNLS could be observed as a linear
PT

programming problem, where the least square estimate of error term is a function which is

minimized. As another option, instead of least squares it is possible to use some other estimates
CE

of error in order to avoid negative solutions. In that case, positivity condition is integrated in the
AC

error approximation norm. For this purpose, semilogarithmic loss function [33] is used. In both

cases (NNLS and semilogarithmic loss function) it is possible to add weighted coefficients to the

minimization function and give different importance to the outputs. QR decomposition, NNLS

and semilogarithmic loss function are used in the examples given in the paper which are related

to wheat milling.

2. Theoretical part
 ACCEPTED MANUSCRIPT

Let us recall Eq. (1), by B is denoted the breakage matrix and by f and o are denoted input

and output vectors, respectively. Then, the following conditions hold:

 f j  1,  oi  1 b  1, j  1, 2, , n.
n m m
and

PT
j 1 i 1 i 1
ij

(4)

RI
By direct computation Eq. (4) implies that

oi   mi , M i  , i {1, 2, , m},

SC
NU
(5)

where mi  min bij and M i  max bij . Introducing matrix C  [cij ]( m1)( n1) and vector
1 j  n 1 j  n
MA
d  [di ]( m1)1 , where

cij  bij  bin , i {1, 2, , m  1}, j {1, 2, , n  1}, di  oi  bin , i {1, 2, , m 1},
ED

(6)
PT

matrix equation Eq. (2) is transformed into

C  f  d. (7)
CE

Case with more outputs then inputs (m  n) will be considered in the sequel. Then, n  1

outputs are freely chosen and by Eq. (7) inputs and the remaining m  n outputs are uniquely
AC

determined [21]. Considering (4) acceptable outputs are solutions of a system of linear

inequalities

 gij  o j   gij  b jn  0, i  1, 2, , n  1 ,

n 1 n 1

j 1 j 1

   ij  j    gij   b jn  1,
 n 1 
n 1

n 1 n 1

(8)
 
j 1  i 1  j 1  i 1 
g o
 ACCEPTED MANUSCRIPT

where coefficients g ij are the elements of inverse matrix for upper square submatrix of C.

Geometrically, the solutions of the system Eq. (8) are the points in the interior of the convex

polyhedron, where the vertices are determined by the columns of the breakage matrix B :

V j  b1 j , b2 j , ,bn1, j  , j  1, 2,

PT
, n. (9)

RI
First, the case m 3 will be considered. Without loss of generality suppose that

b11  b12  b13 . Then the triangle with vertices V1  b11 , b21  , V2  b12 , b22  and V3  b13 , b23  is given.

SC
NU
Instead of choosing values o1 and o2 simultaneously values are set one at the time by the

following procedure.

Step 1. Let o1 be a value from interval  b11 , b13  .

MA
Step 2. Having chosen the value of o1 , the value of o2 is determined in the following way:
ED

if o1   b11 , b12  then o2   l12  o1  , l13  o1   ,

if o1   b12 , b13  then o2   l23  o1  , l13  o1   ,

PT

where lines l23 , l13 and l12 are defined by pairs of points V2 ,V3 , V1 ,V3  and V1 ,V2  ,
CE
AC

respectively.

The procedure given in the case m 3 will be extended to the general case m  n, m  n .

The main idea is to reduce problem dimension (step by step) by using hyperplanes. Input

parameters are points V j , j  1, 2, , n, given in Eq. (9). These points are vertices of the convex

n1
polyhedron in whose interior contains the acceptable solution of the problem. Given points

 n 
determine    n hyperplanes in
 n  1
n1
expressed by equations:
 ACCEPTED MANUSCRIPT

hk ,1 : xn1   aik  xi  an 1,k , k  1, 2, , n ,

n2

i 1

(10)

which represent sides of the polyhedron. Then, for some k0 1, 2, , n  1 arbitrary

PT
ok0  (mk0 , M k0 ) is chosen, where mk0 and M k0 are defined in Eq. (5). Without loss of generality

RI
let k0  1 and set x1  o1 in Eq. (10). This leads to n hyperplanes in

SC
n 2
are obtained:

hk ,2 : xn 1   aik  xi  an 1,k  a1k  o1 , k  1, 2, , n.

n2

NU
i 2

(11)
MA
 n  n  (n  1)
The intersections of hyperplanes Eq. (11) are  
 n  2
points:
2

I  1, 2, , n , i, j  I , i  j.
ED

Vij 
kI \i , j
hk ,2 ,
PT

(12)

The r th coordinate of the point Vij is denoted by Vij , r  1, 2, , n  2. The points Vij such that
CE

mr  Vij  M r , r  2, , n  1 ,
AC

r 1

(13)

are the starting parameters for the next iteration. Now, the whole procedure is repeated in

dimension n 2
. After n  2 steps the algorithm ends by choosing the last output similarly as in

the Step 2 of the case m 3 . In Figure 1 the described procedure is represented as a diagram.

In the approximation approach the case m  n could be observed as a linear least

square problem and solved using QR decomposition. Then, the breakage matrix Bmn is

decomposed as:
 ACCEPTED MANUSCRIPT

B  Q  R, (14)

where Qmn is the orthogonal matrix and Rnn is the upper triangular matrix. Best least square

approximation is

PT
f  R 1  QT  o, (15)

i.e. R 1  QT is the pseudoinverse matrix of the non-quadratic matrix B . Condition number,

RI
SC
which describes the sensitivity of the problem, is given by

 max  B 
 ( B)   B B ,
 min  B 

NU
(16)

where  min  B  and  max  B  are minimal and maximal singular value of B, and B is norm
MA
of the pseudoinverse matrix of B . Namely, it is recommended to use QR decomposition if the
ED

condition number is not too big (problem is well-conditioned) [24, 26].

However, the results provided by this method could have negative values. With non-
PT

negative requirements for the input PSD, the equation (1) is a NNLS problem and could be
CE

observed as a linear programming problem

min o  B  f f  0,
AC

2
, (17)

where

o B f 
  oi   B  f i . 
m 2

i 1
2

On the other hand, the non-negativity condition could be incorporated in a different norm

whose nature does not allow negative solutions. For this purpose norm defined by the least

semilogs [33]
 ACCEPTED MANUSCRIPT


oi  Bi  f
o B f
m
o
Bi  f Bi  f
log i
i 1
semi log

(18)

PT
will be used in the paper. In addition, if desired outputs are chosen by some priority, then

weights i are added to the minimization function:

RI
min  i   i ,
m

SC
(19)
i 1
f

where  i  d (oi ,  B  f i ) are error estimates defined by the observed norm. The coefficient of

NU
determination r 2 , defined by
MA

  

 i i
2
m

i 1  
o o
r 
  oi  oi     oi  oi 
  
 
2
ED

 2
2
m m

i 1   i 1  
PT


is used to test the best-fit solution oi relative to the experimental data o .
CE

So, by r 2 is estimate error between solution arised from model and experimental one.
AC

Also, it is known that these type of errors (modeling errors) are negligible in once-through mills

[4]. Moreover, in the case of single breakage event the numerical time-discretization errors are

not critical [4]. Let us mention that the PSD of the milling output obtained by experimental

milling was compared with the PSD determined using the procedure for solving the reverse

problem presented in [21], where the average deviation was 2.62% .

3. Materials and Methods

 ACCEPTED MANUSCRIPT

Wheat was separated into fractions using Sortimat (Perten AB, Sweden). Slotted sieves

were used to separate the grains by thickness into four different size fractions: >2.8 x 20 mm, 2.8

x 20/2.5 x 20 mm, 2.5 x 20/2.2 x 20 mm and <2.2 x 20 mm. In order to determine the initial

PT
moisture content and hence the amount of water to be added during conditioning, for each size

RI
fraction of wheat moisture content was determined, according to ICC standard method No.110/1

[34].

SC
In the first step batches of 0.5 kg of the wheat kernels were conditioned to 13.5%

NU
moisture for 20 hours. In the second step, samples were conditioned to 16% moisture for 12

hours. Finally, 30 minutes before milling, moisture content in the samples was raised for another
MA
0.5%.

The conditioned samples were milled on the laboratory roll stand Variostuhl, model C Ex 2
ED

(Miag, Braunschweig, Germany). Fluted rolls 100 mm in length (250 mm diameter) with 3.2
PT

flutes per cm, were used and operated in the dull to dull disposition with flute angles 40o/70o and

8%. Differential of 2.5 was used relative to the fast roll speed of 6 m/s. Roll gap was set at 0.5
CE

mm and samples were fed at the rate of 0.21 kg/cm/min. Sieve analysis was performed on the
AC

Bühler laboratory sifter model MLU-300 (Uzwil, Switzerland) using the entire milled stock.

Samples were sieved for 3 min and separated into eight fractions using square aperture sieves

(size 2000, 1180, 850, 600, 450, 300 and 150 m) along with a bottom collecting pan. The

amount of material remaining on each sieve was measured to ±0.1 g using a Sartorius Precision

balance (Sartorius AG, Germany). Breakage matrix B of dimension 8  4 was constructed from

the collected sieve analysis data.

4. Results and discussion

 ACCEPTED MANUSCRIPT

The proposed breakage matrix approaches were tested by examples related to wheat

milling. Table 1 shows sieve analysis data given as weight fraction of the milling output size

fraction produced by milling different sized wheat kernel fractions.

PT
The output size distributions represent the columns of the breakage matrix, B84 . Examples 1 and

RI
2 cover precision and approximation approaches, respectively.

SC
Example 1. Breakage matrix B84 , input values f and output values o are given by:

 0.3441 0.4112 
 0.2503 0.2807 
0.3654 0.3955

NU
 0.2492 0.2666
 0.0888 0.0448   f1   o1 
  f  o 
0.0750 0.0578

B  0.0847 
f   2
, o   2 .
MA
0.0427 0.0349 
0.1286 0.1300 0.0969
 f3   
,
     
0.0432 0.0360
 0.0490 0.0423  f4   o8 
 0.0464 0.0456 
0.0461 0.0464

 
ED

0.0445 0.0470
 0.0500 0.0466 0.0539 0.0558 
PT

By (8) the set of acceptable outputs is the interior of a tetrahedron (Figure 2.) defined by

the vertices V1  0.3441,0.2503,0.0888 , V2  0.3654,0.2492,0.0750  , V3  0.3955,0.2666,0.0578

CE

and V4  0.4112,0.2807,0.0448 . This tetrahedron is bounded by the planes

z  pi  x, y  , i 1, 2,3, 4, where:

AC

p1  x, y   0.108 x – 0.802 y  0.314

p2  x, y   0.480 x – 0.387 y  0.351

p3  x, y   0.649 x – 0.015 y  0.316

p4  x, y   0.642 x  0.121 y  0.279 . (20)

The maximum and minimum values in corresponding rows of the breakage matrix are

m1  0.3441, M1  0.4112, m2  0.2492, M 2  0.2807,

 ACCEPTED MANUSCRIPT

m3  0.0448, M 3  0.0888, m4  0.0847, M 4  0.1300.

Considering Eq. (5) let o1  0.3500. From Eq. (11), the intersections of the planes given by Eq.

(20) and the plane x  o1 are obtained:

PT
z  pi  o1 , y   li  y  , i 1, 2,3, 4. (21)

Intersection points of the lines z  li  y  and z  l j  y  , i, j 1, 2,3, 4 , i  j, are denoted by Vij :

RI
SC
V12  0.2500, 0.0850 , V13  0.2522,0.0852 , V14  0.2403, 0.0838 ,

V23  0.2530, 0.0849 , V24  0.2386, 0.0851 , V34  0.2257, 0.0955 .

NU
By Eq. (13) points Vij such that
MA
m2  Vij  M 2 and m3  Vij  M 3
1 2
ED

are vertices of a convex polyhedron in lower dimension.

PT

Set of acceptable values for outputs o2 and o3 is the interior of triangle with vertices

V12  0.2500, 0.0850 , V13  0.2522, 0.0852  and V23  0.2530, 0.0849 . The triangle (Figure 3.) is
CE

bounded by lines

l1  y   0.122 · y  0.054
AC

l2  y   – 0.015 · y  0.089

l3  y   – 0.388 · y  0.183 . (22)

To satisfy the condition o2   0.2500, 0.2530 , we take, for example, o2  0.2520. Then

l1  o2   0.0852, l2  o2   0.0849, l3  o2   0.0853,

and o3   0.0849, 0.0852 . By choosing o3  0.0851 the input PSD

f1  0.8737, f 2  0.0140 , f3  0.1119

 ACCEPTED MANUSCRIPT

leads to the desired output PSD. Remaining outputs are

o4  0.1251 , o5  0.0419 , o6  0.0487 , o7  0.0464 and o8  0.0504.

Complete algorithm for this example is given in Figure 4. The shaded path refers to the

PT
output values chosen through the example.

RI
Example 2. Problem B  f  o, for the breakage matrix given in Example 1, will be solved by

SC
QR decomposition. Condition number  ( B)  191.0623 can not be considered too big thus QR

NU
decomposition is used. Then the breakage matrix B84 is decomposed as B  Q  R, where Q84
MA
and R44 are matrices given by:

0.7432 0.5361 0.0671 0.3299 

0.5406 0.3355 0.3488 0.6371
 
ED

 0.1918 0.7052 0.1033 0.3840 

 
 0.2778 0.0864 0.8689 0.1461
Q
0.0922 0.0272 0.1992 0.1963
PT

and
 
 0.1058 0.1830 0.0602 0.4445 
0.1002 0.1363 0.0890 0.2660 
CE

 
0.1080 0.2060 0.2386 0.0976 
AC

0.4630 0.4751 0.4948 0.5078 

 0 0.0229 0.0465 0.0609 
R
 0 0.0420 0.0603 
.
 
0
 0 0 0 0.0088 

The best-fit solution o (in the least square sense) for the desired output PSD o is obtained by

the aid of the pseudoinverse matrix R 1  QT . For example,

 ACCEPTED MANUSCRIPT

0.3640  0.3640 
0.2520  0.2516 
   
0.0770  0.2342   0.0768
   0.6765  
o  0.1270 
f  R 1  QT  o   , o  B f   0.1269 
0.0410  0.0874   0.0425
, .
   

PT
 
0.0470   0.0400  0.0469 
0.0450   0.0453
   

RI
0.0470  0.0482 

SC
On the other hand, for some other o,

 0.3632   0.3634 

NU
 0.2546   0.2539 
   
 0.0768   0.3610   0.0768 
   0.5835   
MA
o  0.1262 
, f    , o  B f   0.1257 
 0.0398   0.1224   0.0425 
,
     
 0.0468   0.1808   0.0466 
0.0457   0.0452 
ED

   
 0.0469  0.0487 
PT

the approximation o is provided with the input PSD such that one coordinate is negative. Hence,
CE

conditions with non-negative constraints for input PSD must be included. In view of Eq. (17) the
AC

problem is solved by a linear programming method and the following solution is obtained:

 0.3632  0.3636 
 0.2546   0.2535
   
 0.0768     
  0.5577   
0.3483 0.0771

o  0.1262 
, f    , o  B f   0.1255
 0.0398   0  0.0424 
. (23)
     
 0.0468   0.0972  0.0469 
0.0457  0.0454 
   
 0.0469   0.0488

As can be seen negativity of the input size fraction f 3 is overcome by setting its value to 0. This

means that one of the size fractions is excluded from the input material to a milling operation,
 ACCEPTED MANUSCRIPT

which is potentially unacceptable in practice. On the other hand, if the method of least semilogs

Eq. (18) is used, different solution is obtained:

 0.3632   0.3570 
 0.2546   0.2484 
   

PT
 0.0768     
     
0.3352 0.0757

o  0.1262 
, f   0.5062 
, o  B f   0.1225 

RI
 0.0398   0.1215   0.0413 
. (24)
     
 0.0468  0.0210   0.0463

SC
0.0457  0.0447 
   
 0.0469   0.0481

NU
MA
Coefficients of determination for approximations obtained in Eq. (23) and Eq. (24), respectively,

are:
ED

r 2  0.9988 and r 2  0.9990,

PT

which can be considered as highly significant goodness of fit.

CE

Considering wheat flour milling it is well-established in literature that wheat millability is

AC

highly influenced by the dimensions of individual wheat kernels and the kernel size distribution

[35, 36, 37, 38]. In a given mill with fixed roll parameters and sieve apertures, change in kernel

size affects grinding performance [38]. Ratio of the minimum dimension of wheat kernel to roll

gap critically determines the breakage patterns [6, 14, 17, 39]. Therefore, the miller should

consider the parameter of kernel size distribution and strive to be within a certain tolerance for

optimum results [38].

5. Conclusions
 ACCEPTED MANUSCRIPT

The paper represents an improvement of the recent studies of the reverse problem in the

breakage matrix context (defining the input PSD to a milling operation which would give the

desired output PSD). We propose two ways for solving the problem, the “approximation” and the

PT
“precision” methods.

RI
The solution of the reverse problem could be useful in milling operations where the

possibility of controlling the input PSD exists. This paper provides a solving procedure for the

SC
reverse breakage matrix problem via algorithm making it practically applicable because it could

NU
be implemented (as illustrated by Example 1).

The “precision” approach is limited to n-1 outputs that can be freely chosen. Sometimes,
MA
especially in higher dimensions where geometrical interpretation of the solution is not possible, it

is not easy to provide the set of acceptable outputs such that it is possible to obtain acceptable
ED

solution (non negative values for the weight fraction of the input size fractions). In contrast to the
PT

solving procedure where the values of desired outputs are chosen simultaneously, the algorithm

given in this paper provides a possibility to set outputs one at the time. This generates a precise
CE

range for the current output. After choosing one output the range of acceptable values for the
AC

next one is precisely defined.

The “approximation” approach is not limited to n-1 outputs that can be freely chosen,

however it is limited in a sense that it is only possible to find a best-fit solution which would

correspond to desired output PSD (Example 2). In case of the overdetermined set of linear

equations, the solution could be obtained by some of the orthogonal decomposition methods such

as the QR decomposition used in this paper. However, the results provided by these methods

could have negative values. This indicated the necessity of imposing the additional non-negative

constrains on input PSD, and the solution is then obtained by the non-negative least square
 ACCEPTED MANUSCRIPT

method. A step forward is to integrate the condition of positivity in the error approximation form

and this is done by the semilogarithmic loss function.

Finally, let us conclude that the precision and approximation approach gives the complete

PT
overview of the reverse problem in the breakage matrix context.

RI
Acknowledgement

SC
The work presented is supported by the Serbian Ministry of Education and Science

NU
projects No. III44006, No. 031014 and No. 174009 , PSNTR project No. 114-451-2167.
MA
References

[1] L. G. Austin, A review: introduction to the mathematical description of grinding as a rate

ED

process, Powder Technol. 5 (1971) 1-17.

PT

[2] T. P. Meloy, M. C. Williams, Problems in population balance modeling of wet grinding,

Powder Technol. 71 (1992) 273-279.

CE

[3] E. Bilgili, B. Scarlett, Population balance modeling of non-linear effects in milling processes,
AC

Powder Technol. 153 (2005) 59-71.

[4] E. Bilgili, M. Capece, Quantitative analysis of multi-particle interactions during particle

breakage: A discrete non-linear population balance framework, Powder Technol. 213 (2011)

162-173.

[5] J. Baxter, A. Abu-Nahar, U. Tüzün, The breakage matrix approach to inadvertent particulate

degradation: dealing with intra-mixture interactions, Powder Technol. 143-144 (2004) 174-178.

[6] G. M. Campbell, P. J. Bunn, C. Webb, and S. C. W. Hook, On predicting roller milling

performance: Part II. The breakage function, Powder Technol. 115 (2001) 243-255.
 ACCEPTED MANUSCRIPT

[7] R. G. Holdich, Fundamentals of Particle Technology, Midland Information Technology and

Publishing, Leicestershire, 2002.

PT
[8] S. R. Broadbent, T. G. Callcott, Coal breakage processes: I. A new analysis of coal breakage

RI
processes, Journal of the Institute of Fuel 29 (1956a) 524-528.

[9] S. R. Broadbent, T. G. Callcott, Coal breakage processes: II. A matrix representation of

SC
breakage, Journal of the Institute of Fuel 29 (1956b) 528-539.

NU
[10] S. R. Broadbent, T. G. Callcott, Coal breakage processes: IV. An exploratory analysis of the

cone mill in open circuit grinding, Journal of the Institute of Fuel 30 (1957) 18-21.
MA
[11] L. G. Austin, D. R. Van Orden, J. W. Perez, A preliminary analysis of smooth roll crushers,

Int. J. Miner. Process. 6 (1980) 321-336.

ED

[12] S. R. Krogh, Crushing characteristics, Powder Technol. 27 (1980) 171-181.

PT

[13] R. S. C. Rogers, K. Shoji, A double-roll crusher model applied to a full scale crusher,

Powder Technol. 35 (1983) 123-129.

CE

[14] G. M. Campbell, C. Webb, On predicting roller milling performance: Part I. The breakage
AC

equation, Powder Technol. 115 (2001) 234-242.

[15] A. Fistes, G. Tanovic, Predicting the size and compositional distributions of wheat flour

stocks following first break roller milling using the breakage matrix approach, J. Food Eng. 75

(2006) 527-534.

[16] C. Fang, G. M. Campbell, Stress-strain analysis and visual observation of wheat kernel

breakage during roller milling using fluted rolls, Cereal Chemistry, 79 (2002a) 511-517.

[17] C. Fang, G. M. Campbell, Effect of roll fluting disposition and roll gap on breakage of

wheat kernels during first-break roller milling, Cereal Chemistry, 79 (2002b) 518-522.
 ACCEPTED MANUSCRIPT

[18] E. Bilgili, M. Capece, A rigorous breakage matrix methodology for characterization of

multi-particle interactions in dense-phase particle breakage, Chem. Eng. Res. Des. 90 (2012)

1177-1188.

PT
[19] M. Capece, E. Bilgili, R. N. Dave, Emergence of falsified kinetics as a consequence of

RI
multi-particle interactions in dense-phase comminution processes, Chem. Eng. Sci. 66 (2011)

5672-5683.

SC
[20] E. Bilgili, J. Yepes, B. Scarlett, Formulation of a non-linear framework for population

NU
balance modeling of batch grinding: Beyond first-order kinetics, Chem. Eng. Sci. 61 (2006) 33-

44.
MA
[21] A. Fistes, D. Rakić, A. Takači, M. Brdar, Using the breakage matrix approach to define the

optimal particle size distribution of the input material in a milling operation, Chem. Eng. Sci.
ED

102 (2013) 346-353.

PT

[22] W. H. Press, B. P. Flannery, S. A. Teukolsky, W. T. Vetterling, Numerical Recipes in C:

The Art of Scientific Computing, 2nd Edition, Cambridge University Press, 1992.
CE

[23] C. L. Lawson, R. J. Hanson, Solving least squares problems, Classics in applied

AC

mathematics vol. 15, SIAM, Philadelphia, 1996.

[24] G. H. Golub, C. F. Van Loan, Matrix Computations, 3rd Edition, The John Hopkins

University Press, Baltimore, 1996.

[25] A. Björck, Numerical methods for least squares problems, SIAM, Philadelphia, 1996.

[26] D. S. Watkins, Fundamentals of Matrix Computations, 3rd Edition, John Wiley & Sons, Inc,

New Jersey, 2010.

[27] L. N. Trefethen, D. Bau III, Numerical Linear Algebra, SIAM, Philadelphia, 1997.
 ACCEPTED MANUSCRIPT

[28] G. M. Campbell, C. D. Rielly,P. J. Fryer, P. A. Sadd, Reconstruction of bubble size

distributions from slices, In: Campbell GM, Webb C, Pandiella SS, Niranjan K (Eds.), Bubbles

in Food. Eagan Press, St. Paul, 1999, pp. 207-220. (Chapter 23).

PT
[29] M. Bertero, P. Boccacci, Introduction to inverse problems in imaging, IOP Publishing,

RI
Bristol, 1998.

[30] G. D. De Villiers, B. McNally, E. R. Pike, Positive solutions to linear inverse problems,

SC
Inverse Probl. 15 (1999) 615-635.

NU
[31] I. Daubechies, M. Defrise, C. De Mol, An iterative thresholding algorithm for linear inverse

problems with a sparsity constraint, Commun. Pur. Appl. Math. 57 (2004) 1413-1457.
MA
[32] Y. Vardi, D. Lee, From image deblurring to optimal investments: Maximum likelihood

solutions for positive linear inverse problems, J. R. Stat. Soc. B, 55 (1993) 569-612.
ED

[33] K. Yoneda, A Parallel to the Least Squares for Positive Inverse Problems, J. Oper. Res. Soc.
PT

Jpn. 49 (2006) 279-289.

[34] ICC Standard No. 110/1: Determination of Moisture Content of Cereals and Cereal Products
CE

(Practical Method), 1976.

AC

[35] M. K. Dattaraj, A. B. Ward, F. F. Niernberge, The relation of certain physical characteristics

of wheat to milling properties, Association of Operative Millers-Bulletin, January (1975) 3537-

3542.

[36] D. R. Marshall, D. J. Mares, H. J. Moss, F. W. Ellison, Effects of grain shape and size on

milling yields in wheat, I Theoretical Analysis based on simple geometric models, Aust. J. Agric.

Res. 35 (1984) 619-630.

[37] D. R. Marshall, D. J. Mares, H. J. Moss, F. W. Ellison, Effects of grain shape and size on

milling yields in wheat, II Experimental studies. Aust. J. Agric. Res. 37 (1986) 331-342.
 ACCEPTED MANUSCRIPT

[38] E. S. Posner, A. N. Hibbs, Wheat Flour Milling, AACC, St.Paul, Minnesota, 2005.

[39] Y. Z. Li, E. S. Posner, The Influence of kernel size on wheat millability, Association of

Operative Millers-Bulletin. November (1987) 5089-5098.

PT
RI
SC
NU
MA
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Figure 1. Precision approach algorithm

Figure 2. Solution space for acceptable outputs

Figure 3. Solution area for acceptable outputs

PT
Figure 4. Precision approach algorithm (Example 1.)

RI
SC
NU
MA
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

PT
Notation

RI
bij elements of breakage matrix
B breakage matrix

SC
DL discrete linear
DNL discrete non-linear
f input size vector

NU
hk hyperplanes in n1
lij line determined by points Vi and V j
mi minimal value in ith row of breakage matrix
MA
Mi maximal value in ith row of breakage matrix
m number of output fractions
n number of input fractions
NNLS non-negative least squares problem
ED

o output size vector


o best-fit aproximation of o
PBM population balance model
PT

PSD particle size distribution

Q orthogonal matrix from QR decomposition
R upper triangular matrix from QR decomposition
CE

r2 coefficient of determination
SVD singular value decomposition
TCNL time-continuous non-linear
AC

Vj vertices of the convex polyhedron

Vr r th coordinate of the point V
i
 ( B)
error estimates

  B
condition number of matrix B
singular value of matrix B
i weights
B norm of the matrix of B

Subscripts

i, j size class indices

m,n size class with finest particles
min minimal
max maximal
 ACCEPTED MANUSCRIPT

semilog semilog norm of the matrix

1 size class with coarsest particles
2 squared norm of the matrix

PT
Superscripts

T transpose matrix
th

RI
(n) n breakage event
(0) initial mass fraction
-1 inverse matrix

SC
* pseudoinverse matrix

NU
MA
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
MA
ED
PT
CE
AC

Figure 1
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
MA
ED

Figure 2
PT
CE
AC
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
MA
ED

Figure 3
PT
CE
AC
 ACCEPTED MANUSCRIPT

PT
RI
SC
NU
MA
ED
PT
CE
AC

Figure 4
 ACCEPTED MANUSCRIPT

Table 1 Breakage matrix for the first break milling of wheat

Output size class Input (kernel) size class (mm)
(m) >2.8 2.5-2.8 2.2-2.5 <2.2
>2000 0.3441 0.3654 0.3955 0.4112

PT
1180-2000 0.2503 0.2492 0.2666 0.2807
850-1180 0.0888 0.0750 0.0578 0.0448

RI
600-850 0.1286 0.1300 0.0969 0.0847

SC
450-600 0.0427 0.0432 0.0360 0.0349
300-450 0.0490 0.0461 0.0464 0.0423

NU
150-300 0.0464 0.0445 0.0470 0.0456
<150 0.0500
MA 0.0466 0.0539 0.0558
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Graphical abstract

PT
RI
SC
NU
MA
ED
PT
CE
AC
 ACCEPTED MANUSCRIPT

Highlights:

 The reverse problem in the breakage matrix context is considered.

 Solving procedures for the overdetermined case are given.
 Best-fit solution of the problem in the least square sense is obtained.

PT
 Applicability of the proposed model was confirmed by some examples.

RI
SC
NU
MA
ED
PT
CE
AC