US 8,429,100 B2Granted

Method for building adaptive soft sensor

AssigneeNational Tsing Hua University

InventorsShi-Shang Jang, Tain-Hong Pan, Shan-Hill Wong

Life of this patent

Estimating the term…

Abstract

The invention discloses a method for building adaptive soft sensor. The method comprises the following steps. The input and schedule vectors are constructed, and a novel learning algorithm that uses online subtractive clustering is used to recursively update the structure and parameters of a local model network. Three rules are proposed for updating centers and local model coefficients of existing clusters, for generating new clusters and new models as well as for merging existing clusters and their corresponding models. Once verified, the online inferential model can be created to generate the predicted value of process. Thus, it does not need much memory space to process the method and can be easily applied to any other machine.

Claims6

Dependent claims

A method for building an adaptive soft sensor, comprising the following steps:

using a distributed control system (DCS) for sampling and inspecting a manufacturing process;
using sensors configured in the distributed control system to obtain online data for recording process variables to form input vectors of an interferential model;
using the distributed control system to preprocess the online data to remove approximately identical data records and/or remove the data records that out of a reasonable range in order to be meaningful for model generation and optimization;
operating an online subtractive cluster algorithm to divide the online data into a plurality of local regions according to a schedule vector, which is a sub set of the input vectors, and the plurality of local regions compose of a plurality of clusters;
building a linear model in each of the local regions, and constructing the inferential model based on each of the linear models;
converting the online subtractive cluster algorithm into a recursive regression algorithm;
calculating density of the schedule vector with
P
Φ
=
κ
-
1
(
k
-
1
)
⁢
(
γ
⁡
(
k
)
+
1
)
-
2
⁢
η
⁡
(
k
)
⁢
ϕ
⁡
(
k
)
+
σ
⁡
(
k
)
,
where Pφ is the density of the scheduled vector k is a time interval, γ is the covariance of scheduling vector at the current sampling time k, which denotes the current change of soft sensors, η is the accumulation of changes of the scheduling vector for the past sampling time 1, 2, . . . k−1, which denotes the past change of soft sensor, σ is the accumulation covariance of scheduling vector for the past sampling time 1, 2, . . . k−1, which denotes the known knowledge of soft sensors φ is the schedule vector, γ(k)=φ(k)Tφ(k),
η
⁡
(
k
)
=
∑
j
=
1
k
-
1
⁢
ϕ
⁡
(
j
)
T
=
η
⁡
(
k
-
1
)
+
ϕ
⁡
(
k
-
1
)
T
,
⁢
σ
⁡
(
k
)
=
∑
j
=
1
k
-
1
⁢
ϕ
⁡
(
j
)
T
⁢
ϕ
⁡
(
j
)
=
σ
⁡
(
k
-
1
)
+
ϕ
⁡
(
k
-
1
)
T
⁢
ϕ
⁡
(
k
-
1
)
;
merging two nearest clusters, updating existing clusters or adding a new cluster to eliminate the recording of the redundant online data for saving memory space of the distributed control system, wherein merging two nearest clusters, updating existing clusters or adding the new cluster further comprise:
merging two nearest clusters into a new cluster when the two cluster centers are closer than a threshold;
updating existing clusters;
while obtaining new online data;
and
adding the new cluster when the new online data is far away from the plurality of clusters already exist;
operating the recursive regression algorithm to update a plurality of local models in real-time according to the linear model of the local regions;
and
obtaining a predicted output value by using each of the linear models.

147 more elements

Dependencies are inferred from the claim text.

patentgrep

Loading patent…

Abstract

A method for building an adaptive soft sensor, comprising the following steps:

using a distributed control system (DCS) for sampling and inspecting a manufacturing process;
using sensors configured in the distributed control system to obtain online data for recording process variables to form input vectors of an interferential model;
using the distributed control system to preprocess the online data to remove approximately identical data records and/or remove the data records that out of a reasonable range in order to be meaningful for model generation and optimization;
operating an online subtractive cluster algorithm to divide the online data into a plurality of local regions according to a schedule vector, which is a sub set of the input vectors, and the plurality of local regions compose of a plurality of clusters;
building a linear model in each of the local regions, and constructing the inferential model based on each of the linear models;
converting the online subtractive cluster algorithm into a recursive regression algorithm;
calculating density of the schedule vector with
P
Φ
=
κ
-
1
(
k
-
1
)
⁢
(
γ
⁡
(
k
)
+
1
)
-
2
⁢
η
⁡
(
k
)
⁢
ϕ
⁡
(
k
)
+
σ
⁡
(
k
)
,
where Pφ is the density of the scheduled vector k is a time interval, γ is the covariance of scheduling vector at the current sampling time k, which denotes the current change of soft sensors, η is the accumulation of changes of the scheduling vector for the past sampling time 1, 2, . . . k−1, which denotes the past change of soft sensor, σ is the accumulation covariance of scheduling vector for the past sampling time 1, 2, . . . k−1, which denotes the known knowledge of soft sensors φ is the schedule vector, γ(k)=φ(k)Tφ(k),
η
⁡
(
k
)
=
∑
j
=
1
k
-
1
⁢
ϕ
⁡
(
j
)
T
=
η
⁡
(
k
-
1
)
+
ϕ
⁡
(
k
-
1
)
T
,
⁢
σ
⁡
(
k
)
=
∑
j
=
1
k
-
1
⁢
ϕ
⁡
(
j
)
T
⁢
ϕ
⁡
(
j
)
=
σ
⁡
(
k
-
1
)
+
ϕ
⁡
(
k
-
1
)
T
⁢
ϕ
⁡
(
k
-
1
)
;
merging two nearest clusters, updating existing clusters or adding a new cluster to eliminate the recording of the redundant online data for saving memory space of the distributed control system, wherein merging two nearest clusters, updating existing clusters or adding the new cluster further comprise:
merging two nearest clusters into a new cluster when the two cluster centers are closer than a threshold;
updating existing clusters;
while obtaining new online data;
and
adding the new cluster when the new online data is far away from the plurality of clusters already exist;
operating the recursive regression algorithm to update a plurality of local models in real-time according to the linear model of the local regions;
and
obtaining a predicted output value by using each of the linear models.

147 more elements