# US 2021/0233039 A1 · Soft Measurement Method for Dioxin Emission Concentration In Municipal Solid Waste Incineration Process

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2021/0233039 A1
- Document ID: US20210233039A1
- Status: Published application
- Kind: A1
- Application number: 16/967,408
- Filed: 2019-12-02
- Published: 2021-07-29
- Earliest claimed: 2019-03-24
- Assignees: Beijing University of Technology
- Applicants: Beijing University of Technology
- Inventors: Jian TANG; Junfei QIAO; Zihao GUO; Haijun HE
- Web page: https://patentgrep.com/patent/US20210233039A1

## CPC

- G06Q 10/30: Administration of product recycling or disposal
- G05B 13/027: {using neural networks only}
- G05B 13/042: {in which a parameter or coefficient is automatically adjusted to optimise the performance}
- G05B 13/048: {using a predictor}
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- G06N 20/20: Ensemble learning
- Y02W 90/00: Enabling technologies or technologies with a potential or indirect contribution to greenhouse gas [GHG] emissions mitigation
- G06Q: INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G05B: CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS (systems for controlling or regulating non-electric variables G05D; systems for regulating electric or magnetic variables G05F; control devices or systems insofar as characterised by mechanical features only G05G)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- Y02W: CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO WASTEWATER TREATMENT OR WASTE MANAGEMENT

## Abstract

Disclosed is a soft measurement method of DXN emission concentration based on multi-source latent feature selective ensemble (SEN) modeling. First, MSWI process data is divided into subsystems of different sources according to industrial processes, and principal component analysis (PCA) is used to separately extract the subsystems' latent features and conduct multi-source latent feature primary selection according to the threshold value of the principal component contribution rate preset by experience. Using mutual information (MI) to evaluate the correlation between the latent features of the primary selection and DXN, and adaptively determine the upper and lower limits and thresholds of the latent feature reselection; finally, based on the reselected latent features, a least squares-support vector machine (LS-SVM) algorithm with a hyperparameter adaptive selection mechanism is used to establish DXN emission concentration sub-models for different subsystems, and based on branch and bound (BB) and prediction error information entropy weighting algorithm to optimize the selection of sub-models and calculation weights coefficient, a SEN soft measurement model of DXN emission concentration is constructed.

## Claims

### Claim 1

A soft measurement method of dioxin emission concentration in municipal solid waste incineration (MSWI) process, comprising the steps of:
using a latent feature extraction and primary selection module to divide municipal solid waste incineration process data into subsystems of different sources according to industrial processes;
using a principal component analysis (PCA) to separately extract the subsystems' latent features;
conducting a primary selection of the latent features according to threshold value of principal component contribution rate preset by experience;
using a latent feature evaluation and reselection module, measuring, based on mutual information (MI), the correlation between the selected latent features and dioxin (DXN);
adaptively determining upper and lower limits and thresholds of reselection of the latent features; and
using an adaptive selective ensemble modeling module, further comprising:
based on latent feature reselection, using a least squares-support vector machine (LS-SVM) algorithm with a hyperparameter adaptive selection mechanism to establish DXN emission concentration sub-models for different subsystems; and
adopting strategy optimization based on branch and bound (BB) and prediction error information entropy weighting algorithm to select the sub-models and calculate weight coefficients to construct DXN emission concentration SEN soft measurement model.

### Claim 2 (depends on 1)

The soft measurement method of dioxin emission concentration in municipal solid waste incineration process according to claim 1, wherein the step of using the latent feature extraction and primary selection module further comprising the steps of:
for ith subsystem, using PCA to extract latent features of high-dimensional input process variables, i.e., normalizing input data Xi to mean=0 and variance=1, and decomposing Xi into:
X
i
=
t
1
FeAll
i
⁡
(
p
1
FeAll
i
)
T
+
L
+
t
m
FeAll
i
i
⁡
(
p
m
FeAll
i
i
)
T
+
L
+
t
M
FeAll
i
i
⁡
(
p
M
F
⁢
eAll
i
i
)
T
⁢
⁢
wherein
,
t
m
FeAll
i
i
⁢
⁢
and
⁢
⁢
p
m
FeAll
i
i
;
(
3
)
represent the score and load vector of mFeAllith principal component (PC), T represents the transpose, and MFeAlli represents the number of latent features extracted for the ith subsystem, with the calculation formula as follows:
MFeAlli=rank(Xi) (4);
wherein based on the above expression, all the latent features extracted from the data Xi are expressed as:
T
i
=
[
t
1
FeAll
i
,
L
,
t
m
FeAll
i
i
,
L
,
t
M
FeAll
i
i
]
⁢
wherein
,
T
i
∈
R
N
×
M
FeAll
i
;
(
5
)
represents the score matrix, which is the orthogonal mapping of the data Xi in the direction of the load matrix Pi; Pi is expressed by the following formula:
P
i
=
[
p
1
FeAll
i
,
L
,
p
m
FeAll
i
i
,
L
,
p
M
FeAll
i
i
]
⁢
wherein
⁢
⁢
P
i
∈
R
M
×
M
FeAll
i
;
(
6
)
wherein the latent features extracted from the data Xi are further expressed as:
Z
FeAll
i
=
T
i
=
X
i
⁢
⁢
P
i
=
[
z
1
FeAll
i
,
L
,
z
m
FeAll
i
i
,
L
,
z
M
FeAll
i
i
]
=
[
{
(
z
1
FeAll
i
)
n
}
n
=
1
N
,
⁢
⁢
L
,
{
(
z
m
FeAll
i
i
)
n
}
n
=
1
N
,
L
,
{
(
z
M
FeAll
i
i
)
n
}
n
=
1
N
]
=
{
(
z
FeAll
i
)
n
}
n
=
1
N
⁢
⁢
⁢
wherein
⁢
⁢
Z
FeAll
i
∈
R
N
×
M
FeAll
i
,
(
7
)
wherein all latent features are further expressed as:
ZFeAll[ZFeAll1,L,ZFeAlli,L,ZFeAllI]={ZFeAlli}i=1I (8);
the feature vector corresponding to the mFeAllith load vector
p
m
FeAll
i
i
is recorded as
λ
m
FeAll
i
i
,
the contribution rate
θ
m
FeAll
i
i
corresponding to mFeAllith latent feature
z
m
FeAll
i
i
is calculated using the following formula:
θ
m
FeAll
i
i
=
λ
m
FeAll
i
i
∑
m
FeAll
i
=
1
M
FeAll
i
⁢
λ
m
FeAll
i
i
×
100
;
(
9
)
wherein threshold preset by experience is recorded as θContri, with default value being 1, where the primary selection of the latent features is performed using the following rules:
ξ
m
FeAll
i
i
=
[
1
,
if
⁢
⁢
θ
m
FeAll
i
i
≥
θ
Contri
0
,
else
⁢
⁢
θ
m
FeAll
i
i
<
θ
Contri
;
(
10
)
wherein,
ξ
m
FeAll
i
i
represents the mark value denoting whether the mFeAllith latent feature is selected, and a value of 1 indicates that the latent feature is selected for the primary selection;
wherein, the latent features of the primary selection for the ith subsystem are expressed as:
Z
FeSe
⁢
⁢
1
⁢
st
i
=
[
z
1
FeSe
⁢
⁢
1
⁢
st
i
,
L
,
z
m
FeSe
⁢
⁢
1
⁢
stl
i
i
,
L
,
z
M
FeSe
⁢
⁢
1
⁢
st
i
i
]
=
[
{
(
z
1
FeSe
⁢
⁢
1
⁢
st
i
)
n
}
n
=
1
N
,
⁢
⁢
L
,
{
(
z
m
FeSe
⁢
⁢
1
⁢
st
i
i
)
n
}
n
=
1
N
,
L
,
{
(
z
M
FeSe
⁢
⁢
1
⁢
st
i
i
)
n
}
n
=
1
N
]
=
{
(
z
FeSe
⁢
⁢
1
⁢
stl
i
)
n
}
n
=
1
N
;
(
11
)
wherein, all latent features from the primary selection zFeSe1st are expressed as:
ZFeSe1st=[ZFeSe1st1,L,ZFeSe1sti,L,ZFeSe1stI]={ZFeSe1sti}i=1I (12).

### Claim 3 (depends on 1)

The soft measurement method of dioxin emission concentration in municipal solid waste incineration process according to claim 1, wherein using the latent feature evaluation and reselection module further comprising the steps of:
for the ith subsystem, a mutual information (MI) value between each primary feature
z
m
FeSe
⁢
⁢
1
⁢
st
i
i
and the DXN emission concentration is marked as
ξ
MI
m
FeSe
⁢
⁢
1
⁢
st
i
and calculated using the following formula:
ξ
MI
m
FeSe
⁢
⁢
1
⁢
st
i
=
∫
∫
p
prob
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
,
y
)
⁢
log
⁡
(
p
prob
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
,
y
)
p
prob
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
)
⁢
p
prob
⁡
(
y
)
)
⁢
d
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
)
⁢
dy
=
H
⁡
(
y
)
-
H
(
y
⁢

z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
)
⁢
⁢
⁢
wherein
,
p
prob
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
)
;
(
13
)
and pprob(y) represent the marginal probability density of
z
m
FeSe
⁢
⁢
1
⁢
⁢
st
i
and y;
p
prob
⁡
(
z
m
FeSe
⁢
⁢
1
⁢
st
i
,
y
)
represents the joint probability density;
H
(
y
⁢

z
m
FeSe
⁢
⁢
1
⁢
st
i
)
represents the conditional entropy, and H(y) represents the information entropy;
the threshold of reselection of the latent features is adaptively determined according to the prediction performance of DXN soft measurement model; further comprising the steps of:
the upper limit of the threshold θContriUplimit, the lower limit of θContriDownlimit and the fixed step size θContriStep are calculated using the following formula:
θ
Contri
Uplimit
=
min
⁡
(
max
⁡
(
ξ
M
⁢
I
m
FeSel
⁢
⁢
1
⁢
t
1
)
,
L
⁢
,
⁢
max
⁡
(
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
i
)
,
L
⁢
⁢
max
⁡
(
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
I
)
)
;
(
14
)
θ
Contri
Downlimit
=
max
⁡
(
min
⁡
(
ξ
M
⁢
I
m
FeSel
⁢
⁢
1
⁢
t
1
)
,
L
⁢
,
min
⁡
(
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
i
)
,
L
⁢
⁢
min
⁡
(
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
I
)
)
;
(
15
)
⁢
θ
Contri
Step
=
θ
Contri
Uplimit
-
θ
Contri
Downlimit
N
Contri
Step
;
(
16
)
wherein the functions max(·) and min(·) represent taking the maximum and minimum values respectively; NContriStep, represents the number of candidate thresholds preset by experience, and the default value is 10; and
the determined threshold is recorded as θContri, and is adaptively selected between θContriUplimit and θContriDownlimit based on the prediction performance of DXN soft measurement model;
using the following rules to reselect the latent features of the primary selection:
ξ
m
FeSe
⁢
⁢
1
⁢
st
i
i
=
{
1
,
if
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
i
≥
θ
Contri
0
,
else
ξ
M
⁢
I
m
FeSe
⁢
⁢
1
⁢
st
i
<
θ
Contri
;
(
17
)
Wherein,
ξ
m
FeSe
⁢
⁢
1
⁢
st
i
i
represents the mark value of whether the mFeSe1stith latent feature is selected, and the value of 1 indicates that the latent feature is selected again;
expressing the reselected latent features for the ith subsystem as:
Z
FeSe
⁢
⁢
2
⁢
nd
i
=
[
z
1
FeSe
⁢
⁢
2
⁢
nd
i
,
L
,
z
m
FeSe
⁢
⁢
2
⁢
⁢
nd
i
i
,
L
,
z
M
FeSe
⁢
⁢
2
⁢
⁢
nd
i
i
]
=
[
{
(
z
1
FeSe
⁢
⁢
2
⁢
nd
i
)
n
}
n
=
1
N
,
L
,
{
(
z
m
FeSe
⁢
⁢
2
⁢
⁢
nd
i
i
)
n
}
n
=
1
N
,
L
,
{
(
z
M
FeSe
⁢
⁢
2
⁢
⁢
nd
i
i
)
n
}
n
=
1
N
]
;
and
=
{
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
n
}
n
=
1
N
(
18
)
expressing all the re-selected latent features zFeSe2nd as:
ZFeSe2nd=[ZFeSe2nd1,L,ZFeSe2ndi,L,ZFeSe2ndI]={ZFeSe2ndi}i=1I (19)

### Claim 4 (depends on 1)

The soft measurement method of dioxin emission concentration in municipal solid waste incineration process according to claim 1, the step of using an adaptive selective ensemble modeling module further comprises the steps of:
for the ith subsystem, describe the process of constructing a DXN emission concentration sub-model based on the reselection of latent features zFeSe2ndi, and modeling hyperparameter pairs {Keri,Regi}; further comprising:
transform the reselected latent feature {(zFeSe2ndi)n}n=1N into a high-dimensional feature space by mapping φ(·), and subsequently solve the following optimization problem:
{
min
W
i
,
b
i
O
LS
-
SVM
=
1
2
⁢
(
w
i
)
T
⁢
w
i
+
1
2
⁢
R
eg
i
⁢
∑
n
=
1
N
⁢
(
ζ
n
i
)
2
s
.
t
:
y
^
n
i
=
(
w
i
)
T
⁢
φ
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
n
)
+
b
i
+
ζ
n
i
;
(
20
)
wherein, wi represents the weight coefficient, bi represents the offset, and ζni is the prediction error of the nth sample;
using Lagrangian method to obtain the following formula:
L
i
⁡
(
w
i
,
b
i
,
ζ
i
,
β
i
)
=
1
2
⁢
(
w
i
)
T
⁢
w
i
+
1
2
⁢
∑
n
=
1
N
⁢
(
ζ
n
i
)
2
-
∑
n
=
1
N
⁢
β
n
i
⁡
[
(
w
i
)
T
⁢
φ
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
n
)
+
b
i
+
ζ
n
i
-
y
^
n
i
]
;
(
21
)
wherein, βi=[β1i,L,βni,L,βNi] represents the Lagrange operator vector, and ζi=[ζ1i,L,ζni,L,ζNi] represents the prediction error vector;
solve the above formula:
∂
L
i
∂
w
i
=
0
,
∂
L
i
∂
b
i
=
0
,
∂
L
i
∂
ξ
i
=
0
,
∂
L
i
∂
β
i
=
0
,
;
(
22
)
express the adopted kernel function as follows:
Ωkeri(zFeSe2ndi,(zFeSe2ndi)n)=<φ(zFeSe2ndi)·φ((zFeSe2ndi)n)> (23);
transform LS-SVM problem into solving the following linear equation system:
[
0
1
L
1
1
Ω
ker
i
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
1
,
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
1
)
+
1
R
eg
i
L
Ω
ker
i
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
1
,
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
N
)
M
M
M
M
1
Ω
ker
i
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
N
,
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
1
)
L
Ω
ker
i
⁡
(
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
N
)
,
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
N
)
)
+
1
R
eg
i
]
·
[
⁢
b
i
β
1
i
M
β
N
i
⁢
]
=
[
1
y
1
i
M
y
N
i
]
;
(
24
)
solve the above formula to obtain βi and bi;
express the DXN emission concentration sub-model based on LS-SVM as:
y
^
i
=
∑
n
=
1
N
⁢
β
n
i
·
Ω
ker
i
⁡
(
z
FeSe
⁢
⁢
2
⁢
nd
i
,
(
z
FeSe
⁢
⁢
2
⁢
nd
i
)
n
)
+
b
i
;
(
25
)
wherein the above hyperparameter adaptive selection mechanism of emission concentration sub-model is implemented using the following two steps:
step 1: adopt a grid search strategy with the prediction performance of the sub-model as objective function, and adaptively select the initial hyperparameter pair {(Kerinitial)i,(Reginitial)i} in the candidate hyperparameter matrix Mpara, wherein the hyperparameter matrix Mpara can be expressed as follows:
M
para
=
[
[
K
e
⁢
r
1
,
R
e
⁢
g
1
]
L
[
K
e
⁢
r
1
,
R
e
⁢
g
r
]
L
[
K
e
⁢
r
1
,
R
e
⁢
g
R
]
M
M
M
M
M
[
K
e
⁢
r
k
,
R
e
⁢
g
1
]
L
[
K
e
⁢
r
k
,
R
e
⁢
g
r
]
L
[
K
e
⁢
r
k
,
R
e
⁢
g
R
]
M
M
M
M
M
[
K
e
⁢
r
K
,
R
e
⁢
g
1
]
L
[
K
e
⁢
r
K
,
R
e
⁢
g
r
]
L
[
K
e
⁢
r
K
,
R
e
⁢
g
R
]
]
J
;
(
26
)
where k=1,L,K K represent the number of candidate kernel parameters; r=1,L,R represent the number of candidate penalty parameters; [Kerk,Regr] represents the kth kernel parameter and rth penalty parameter, the formed hyperparameter pair also is the jth parameter pair in the hyperparameter matrix Mpara, that is, there are Mparai=[Kerk,Regr]; j=1,L,J, J=K×R represent the number of all hyperparameter pairs in the hyperparameter matrix Mpara; wherein, the hyperparameter pair {(Kerinitial)i,(Reginitial)i} selected for the first time using a grid search strategy is an element in matrix Mpara, i.e., there is {Kerinitial,Reginitial}∈Mpara;
step 2: based on the {(Kerinitial)i,(Reginitial)i} selected by the above step, a new set of candidate hyperparameters is obtained using the following formula:
(
K
er
vector
)
i
=
(
K
er
initial
)
i
/
k
supara
down
:
k
supara
up
×
(
K
er
initial
)
i
-
(
K
er
initial
)
i
/
k
supara
down
N
ker
:
k
supara
up
×
(
K
er
initial
)
i
;
(
27
)
(
R
er
vector
)
i
=
(
R
eg
initial
)
i
/
k
supara
down
:
k
supara
up
×
(
R
eg
initial
)
i
-
(
R
eg
initial
)
i
/
k
supara
down
N
ker
:
k
supara
up
×
(
R
eg
initial
)
i
;
(
28
)
wherein, (Kervector)i and (Regvector)i represent new candidate hyperparameter sets, corresponding to kernel parameter vector and penalty parameter vector, respectively; Nker and Nreg represent the number of new hyperparameters set based on experience; ksuparadown and ksuparaup represents the hyperparameter shrinkage and expansion factor based on experience, the default value is 10;
by adopting the grid search strategy, adaptively obtain the hyperparameter pair of the ith sub-model {Keri,Regi};
perform the above process on all subsystems, the set of sub-model prediction outputs is expressed as:
Y
^
=
[
y
^
1
,
L
,
y
^
i
,
L
,
y
^
I
]
⁢
{
y
^
i
}
i
=
1
I
=
[
f
1
⁡
(
K
er
1
,
R
eg
1
,
Z
FeSe
⁢
⁢
2
⁢
nd
1
)
,
L
,
f
i
⁡
(
K
er
i
,
R
eg
i
,
Z
FeSe
⁢
⁢
2
⁢
nd
i
)
,
L
,
f
I
⁡
(
K
er
I
,
R
eg
I
,
Z
FeSe
⁢
⁢
2
⁢
nd
I
)
]
=
{
f
i
⁡
(
K
er
i
,
R
eg
i
,
Z
FeSe
⁢
⁢
2
⁢
nd
i
)
}
i
=
1
I
;
(
29
)
wherein, fi(·) represents the ith sub-model;
by combining the optimization selection algorithm based on branch and bound (BB) and the prediction error information entropy weighting algorithm, adaptively optimize the above sub-models and calculate the weighted coefficients;
given the candidate sub-model and weighting algorithm, the optimal sub-model selection is similar to the weighted optimal feature selection;
for a limited number of candidate sub-models, by running optimization and weighting algorithms multiple times, SEN models with ensemble sizes of 2 to (I−1) is obtained; and
rank the optimized SEN models are ranked, and select the best prediction performance model as final DXN soft measurement model;
wherein assuming the ensemble size of final DXN soft measurement model is Isel, the predicted output value ŷ is calculated by the following formula:
y
^
=
∑
i
se
⁢
⁢
1
=
1
I
se
⁢
⁢
1
⁢
w
i
sel
⁢
y
^
i
se
⁢
⁢
1
=
∑
i
se
⁢
⁢
1
=
1
I
se
⁢
⁢
1
⁢
w
i
sel
·
f
i
se
⁢
⁢
1
⁡
(
K
er
i
se
⁢
⁢
1
,
R
eg
i
se
⁢
⁢
1
,
Z
FeSe
⁢
⁢
2
⁢
nd
i
se
⁢
⁢
1
)
.
(
30
)
