Research Article

Multi‑condition identification of thermal process data based on mixed

constraints semi‑supervised clustering
Yue Zhang1,2 · Xiaona Gao1 · Yingjun Bai1 · Mengxue Wang1 · Qing Tian1

Received: 7 June 2021 / Accepted: 25 May 2022

© The Author(s) 2022  OPEN

Abstract
The multi-model method is used in complex system modeling or industrial data monitoring, and its goal is to establish
the sub-models corresponding to different conditions. How to divide the modeling data into datasets corresponding
to different conditions in the case of insufficient prior knowledge of the process is an important problem to be solved.
Machine learning provides many excellent algorithms for condition identification. However, many algorithms are easily
affected by abnormal data, but making full use of prior knowledge can effectively solve this problem. Based on the semi-
supervised clustering, mixed constraints which include composite distance and the pairwise constraint are introduced
to distinguish strongly and weakly dependent data and clarify boundary data to realize a multi-model condition division
of thermal process data. Radial Basis Function neural network is used to realize feature learning, and an online condition
identification is constructed. The influence of network structure and parameters on the generalization ability of the rec-
ognizer is analyzed. On the premise of not significantly increasing the amount of calculation, the generalization ability
is improved by adjusting the weight coefficient in the composite distance, and the weight coefficient under low error
rate is given by particle swarm optimization. Compared with classic the methods, such as Sliding Windows, Bottom-Up
and Top-Down, the proposed method has better performance in segmentation results.

Article Highlights component, is given to distinguish different types of

data.
• The thermal process data in a segment is divided into • The semi-supervised clustering with mixed constraints,
strongly and weakly dependent data, and different seg- superimposing composite distance constraint and pair-
mentation methods are given. wise constraint, is given to segment the thermal pro-
• The definition of composite distance parameter, includ- cess data.
ing time component, value component and velocity

Keywords Semi-supervised clustering · Mixed constraints · Composite distance · Polynomial fit · Online condition
recognizer · RBFNN

* Yue Zhang, [email protected] | 1Department of Automation, North China Electric Power University,
Lianchi District, Baoding 071003, China. 2Hebei Technology Innovation Center of Simulation and Optimized Control for Power Generation,
North China Electric Power University, Baoding 071003, Hebei Province, China.

SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

1 Introduction [4]. With the development of machine learning, many

practical methods for condition division or modal iden-
There are two main types of complex system models, the tification have been developed, such as Sliding Windows
first-principle models and the data-driven models. The (SW), Top-Down, Bottom-Up, Sliding Window and Bottom-
first-principle models require strong professional back- Up (SWAB) [5], Feasible Space Window (FSW) [6]. Ge et al.
ground knowledge and a deep understanding of mode- [7] divided the historical data into several independent
ling objects. Also, many approximations and assumptions groups using the fuzzy C-means clustering method, and
are needed in the modeling process. Therefore, when the then established the sub-models corresponding to the
object structure is complex, a first-principle model can- data subsets using feature extraction, and finally, inte-
not be obtained. In contrast, the data-driven models are grated the monitoring results of each sub-model with
more robust to determine the relationship between the Bayesian reasoning, the proposed method can handle
input and output data. These models are based on a small non-Gaussian information in each operation mode. Zhu
amount of background knowledge and have a strong data et al. [8] proposed a two-layer clustering method based on
processing capability. the global moving window strategy and global clustering
The establishment process of data-driven models can solution strategy and successfully divided the data into
be divided into two categories according to whether the different modes using this method. The method allows
model structure needs to be determined in advance. the different ICA–PCA models to overlap. Lu [9] and Zhao
The first category determines the model structure first, et al. [10] applied the K-means clustering algorithm to the
and then identifies the model parameters, developing a period division of batch processes, the performance of the
parameter model. The second category determines model methods in continuous process needs to be discussed. In
structure and model parameters simultaneously dur- [11], a fuzzy segmentation of time series based on the core
ing the modeling process, developing a non-parameter principal component analysis (KPCA) and Gath-Geva (G-G)
model. The latter methods are more computationally com- clustering, where the window division of multivariate time
plex, but the model accuracy is higher than the former. series is used, is proposed, the advantage of the method
However, with the development of deep learning, many is that time dimension attributes are introduced as extra
efficient algorithms have emerged in these methods, such variables. Zhang Yue [12] introduced the principal com-
as stacked autoencoder, deep belief networks (DBN), and ponent analysis into the design method of sub-window
long short-term memory (LSTM). time span division, and demarcation points of the time
According to the number of models, a data-driven span of the sub-windows are determined by the piecewise
model can be a single model or a multi-model. Commonly, analysis, rolling merging, and cyclic validation. The final
a single model is used to describe objects with obvious result is obtained by multi-step cyclic iteration, and the cal-
linear characteristics, and a multi-model is applied to culation is heavy. Song et al. [13] used the recursive local
complex system modeling and complex industrial data outlier factor algorithm to divide the multimode chemical
monitoring [1–3]. The basic idea of the multi-model is to process into the stable mode and conversion mode and
establish sub-models corresponding to different operat- established the corresponding models, in the algorithm,
ing conditions or modes with different characteristics to not only the number of modes does not need to be deter-
form the model database of conditions or models, which mined in advance, but also details of mode switching can
is used to identify the current data status during the online be acquired. Lv [14] proposed a feature extraction method
monitoring and select the appropriate sub-model for data based on the weighted kernel Fisher criterion to improve
monitoring. the clustering accuracy, where the feature mapping is
The multi-model method needs to establish different adopted to bring the edge classes and outliers closer
monitoring models for multiple modalities with obvious to other normal subclasses. LI Wei [15] used the fuzzy
differences, so how to divide the offline historical data into C-means clustering algorithm based on the conditional
different modes when the prior knowledge of the process positive definite kernel to realize the clustering division of
is insufficient is an important problem to be solved. This a dataset, and then the least square support vector regres-
problem is embodied in the working point identification sion is conducted for each cluster. The method realized
in the multi-model modeling process and the modal iden- the clustering of irregular data. Zhang Shu Mei [16] pro-
tification in the industrial monitoring process. posed an automatic offline modal recognition method for
The traditional methods for multi-condition identifica- multi-modal processes based on an improved K-means
tion include artificial judgment based on prior knowledge clustering algorithm. The method avoided the influence
and machine assistance based on the recursive method of manual identification on the results. Among the exist-
ing analysis methods, clustering based on unsupervised

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

learning has been most widely used, and its main advan- where ds,t (0 ≤ s < t ≤ T ) is the segmentation error of seg-
tage is that it does not rely on prior knowledge, but it also ment [s, t]. It is local error, and
{ it is determined
} by the data
has certain shortcomings, especially in the case of window in the time series segment xs , xs+1 , … xt ,
or modal boundary data. t
This paper considers the time-sequence relationship ∑ ( )
ds,t = (x𝜏 − x̂ 𝜏 )T x𝜏 − x̂ 𝜏 (2)
of the object under study from the perspective of semi- T =s
supervised clustering and proposes a hybrid constraint
that combines pairwise constraints and time constraints where x̂ 𝜏 is estimated value.
to improve the identification accuracy of working condi- It is not a simple single objective optimization problem
tions or mode in the boundary area. The simulation results in the process of condition partition or segmentation. On
show that the semi-supervised clustering based on mixed the basis of ensuring the overall segmentation cost, it is
constraints has higher accuracy in condition identification, necessary to make the local segmentation cost as close
especially in the boundary data. as possible and at a lower level. As mentioned above, the
This paper is organized into three major sections. multivariate time series segmentation problem is trans-
Firstly, the background knowledge of the paper is intro- formed into a constrained optimization problem or a
duced, including the cost of condition division and tradi- multi-step optimization problem.
tional methods. Secondly, the semi-supervised clustering
with mixed constraints is introduced in detail. Thirdly, an 2.2 The traditional methods for multi‑condition
online recognizer based on RBFNN is designed. Fourthly, identification
the method in this paper is compared with the traditional
method by simulation. The final section presents the key The traditional methods for multi-condition identifica-
conclusions and limitations of this work, while offering tion include artificial judgment based on prior knowledge
future directions for research that could advance the cur- and machine assistance based on the recursive method.
rent body of knowledge on this subject. Among them, the more famous methods include Sliding
Windows (SW), Top-Down, Bottom-Up, Sliding Window
and Bottom-Up (SWAB), Feasible Space Window (FSW).
2 Background knowledge Sliding Windows (SW) [5] the algorithm determines
the width of the potential segment by recursive method.
2.1 Cost of condition division Anchor the left point at the first data point, then attempt
to approximate the data to the right with increasing
Process data condition partition is equivalent to the prob- longer segments. At point i, the error is greater than the
lem of multivariate time series segmentation. In essence, user specified threshold, so the subsequence from the
it means that for anchor to i − 1is transformed into a segment. The anchor
{ } a given ( k-dimensional
)T time series
X = x1 , x2 , … xT , xT = x1t , x2t , … xkt , the time domain is moved to location i , and the process repeats until the
is divided according to the change law of the data and the entire time series has been transformed into a piecewise
correlation relationship before and after. linear approximation.
Assuming that the time series is divided into N seg- Top-Down [5] the algorithm works by considering every
ments, the boundary possible partitioning and splitting it at one location. Both
{ time label}of the segmentation result
is defined as t = t1 , t2 , … tN , then the segmentation subsections are tested to see if their approximation error is
result t satisfies 0 < t1 < t2 < ⋯ < tN = T . below some user specified threshold. If not, the algorithm
In the problem of time series segmentation, t1 , t2 , … , tN recursively continues to split the subsequences until all the
segments have approximation errors below the threshold.
[ called segmentation boundary
are ] [ or mutation
] points,
t1 , t2 ], [t2 + 1, t3 ], [t3 + 1, t4 , … tN−1 + 1, tN are called Bottom-Up [5] firstly, the algorithm creates the finest
segmentation segments, and the number of segments N possible approximation of the time series, so that n∕2 seg-
is called segmentation order [17]. ments are used to approximate the n-length time series.
Thermal data condition partition or time series seg- Next, the cost of merging each pair of adjacent segments
mentation can be described as optimization problem. The is calculated, and the algorithm begins to iteratively merge
overall cost of segmentation is J(t), the lowest cost pair until a stopping criteria is met.
Sliding Window and Bottom-Up (SWAB) [5] the algorithm
N−1
∑ keeps a small buffer. Bottom-Up is applied to the data in
J(t) = dti +1,ti+1 (1) the buffer and the leftmost segment is reported. The data
i=0
corresponding to the reported segment is removed from
the buffer and more datapoints are read in. These points

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

( )
are incorporated into the buffer and Bottom-Up is applied |S| is the size of S, and ∃yi ∈ Y , so that xi , yi ∈ L, then S
again. This process of applying Bottom-Up to the buffer, is called the seed set. In particular, when the number of
reporting the leftmost segment. categories which the samples in S belong to is equal to
Feasible space window (FSW) [6] the algorithm intro- k, S can be expressed as S = Uki=1 Si , where Si is the class i
duces a point called a Candidate Segmenting Point (CSP) non-empty sample set.
which may be chosen to be the next eligible segmenting Although the semi-supervised clustering algorithm
point. The distances of all the points lying between the based on the seed sets can effectively improve the cluster-
last segmenting point and the new chosen one are all ing performance, it largely depends on the scale and qual-
within the maximum error tolerance. The key idea of FSW ity of the seed set. Guo Maozu and Deng Chao [20] intro-
is to search for the farthest CSP to make the current seg- duced a semi-supervised clustering algorithm based on
ment as long as possible under the given maximum error the tri-training and data editing, and combined it with the
tolerance. depuration data editing technology to correct and purify
The comparison of the above algorithms is shown in the mislabeled samples in the seed set while expanding
Table 1. the size of the seed set to improve its quality.
The Cop-K-means algorithm introduces the idea of
2.3 Semi‑supervised clustering method pairwise constraints to the traditional K-means algorithm.
During the data distribution process, data objects must
Unlike the traditional unsupervised clustering algorithms, meet the Must-link (ML) constraints and Cannot-Link (CL)
such as the K-means algorithm and expectation–maxi- constraints. In the ML, two selected points must belong
mization (EM) algorithm, semi-supervised clustering to the same class, and in the CL, two selected points do
combines clustering and semi-supervised learning to not have to be elements of the same class. The constraints
improve the clustering performance using a small amount have symmetry and transitivity characteristics that are
of label data and prior knowledge in massive data. The expressed as follows.
semi-supervised clustering algorithms can be divided into Symmetry:
constraint-based semi-supervised clustering algorithms, ( ) ( )
distance-based semi-supervised clustering algorithms, xi , xj ∈ ML xj , xi ∈ ML
and constraint- and distance-based semi-supervised clus-
tering algorithms. ( )
xi , xj ∈ CL
( )
xj , xi ∈ CL

2.3.1 Semi‑supervised clustering based on constraints Transitivity:

( ) ( ) ( )
The idea of such algorithms is to add constraint and xi , xj ∈ ML & xj , xk ∈ ML xi , xk ∈ ML
restriction information to the traditional clustering basis
to improve the clustering performance. The most typical ( ) ( )
xi , xj ∈ CL & xj , xk ∈ CL
( )
xi , xk ∈ CL
algorithms are the Seeded-K-means [18] and Cop-K-means
[19] algorithms, which play a crucial role in the develop- Symmetry and transitivity are crucial for pairwise con-
ment of semi-supervised clustering from different per- straints, which means that when a sample is forced to allo-
spectives of supervisory information. cate the constraint relationship, the constraint violation
The Seeded-K-means algorithm uses a small number of will occur only in the case of CL constraint, and in other
labeled samples as seed sets, and obtains the initial cluster cases, there will be no sample allocation failure.
centers according to the seed sets, improving the cluster- The quality of constraint information directly affects
ing accuracy. Basu et al. [18] define the seed set as follows: clustering results, so the clustering performance can
when S satisfies S ⊆ L, for ∀xi ∈ S , where i = 1, 2,⋯,|S|, and be improved only by obtaining better constraint infor-
mation. Zhu Yu et al. [21] proposed an improved linked
COP-K-means (LCOP-K-means) algorithm based on the
Table 1  Comparison results of the above algorithms
breadth-first search and the Cop-K-means algorithm,
Algorithm Online Complexity which improved data stability and clustering accuracy.
Sliding windows Yes O(Ln) Li Chao Ming et al. [22] proposed a cross-entropy semi-
Top-down No O(n2K) supervised clustering algorithm based on pair constraints.
Bottom-up No O(Ln) This method uses the cross-entropy of samples to express
Sliding window and bottom-up Yes O(Ln) the pairwise constraint information, thus providing higher
Feasible space window Yes O(Kn)

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

clustering accuracy and better results using a smaller condition data appears to be disorderly in an ultra short
amount of pairwise constraint information. period, and the overall change trend is to increase or
decrease in one direction. There is no instantaneous jump
2.3.2 Semi‑supervised clustering combining multiple of amplitude in both steady-state data and transient data.
methods According to the above characteristics of thermal pro-
cess data, the key to the division of steady-state condi-
Chen [23] and Chang [24] have suggested that the semi- tion and process condition is to solve three problems: first,
supervised clustering algorithm can simultaneously use the problem of working condition central point; second,
two types of supervision information, such as class label data screening problem with obvious strong connection
and pairwise constraints, for clustering, especially when relationship with working condition central point; third,
active learning is added to actively label samples, and determination of boundary data of adjacent working
higher quality supervision information and better cluster- conditions.
ing results can be obtained.
Wei et al. [25] proposed a semi-supervised clustering 3.2 The flowchart of the condition identification
method based on pairwise constraints and measures. In
the case of data marked by pairwise constraints, the semi- In this paper, the semi-supervised clustering with mixed
supervised clustering method based on constraints and constraints is used to realize condition identification, as
measures is used to generate different basic clustering shown in Fig. 1.
partitions, and then the target clustering is conducted by
integration. 3.3 Initial seed set establishment
Compared with the single clustering method, the com-
bination of multiple clustering methods can yield the best After data preprocessing, the dataset is subjected to the
use of given monitoring information, improving the algo- whitening process by referring to the idea of the PCA algo-
rithm performance. For the identification of multi-model rithm to reduce the linear correlation under the premise
working conditions, this article uses the initial seed set to that the data are as true as possible. The goal is to mini-
provide the number of working conditions and the centers mize the variance between the original data and the pre-
reference of the dataset under the same working condi- processed data.
tion. It focuses on solving the problem of a fuzzy bound- N
ary between different working conditions. By introducing 1∑
J= x − x̃n2 (3)
mixed constraints, the accuracy of the boundary division N n=1 n
of working conditions has been improved.
M D
∑ ∑
x̃n = ani ui + bi ui (4)
3 Multi‑model condition identification i=1 i=M+1
of thermal process data based
where xn is the original data, x̃n is the preprocessed data,
on mixed‑constraint semi‑supervised
N is the
{ number of data, M is principal component dimen-
clustering }
sion, ui is D dimensional unit orthogonal set.
The correlation coefficients of different dimensions are
3.1 Characteristics of thermal process data
guaranteed to be as small as possible. The correlation coef-
ficient 𝜌ij is defined as follows:
Different from batch process and random process, ther-
mal process data has its own unique characteristics, such N
( )( )
1 ∑ xni − x̃i xnj − x̃j
as strong coupling, non-linear, slow time-varying, etc. 𝜌ij = (5)
N n=1 𝜎i 𝜎j
Reflected in the data, the main performance is that there
are many factors affecting the data change, it is difficult to where 𝜎i and 𝜎j is standard deviation.
predict, the correlation between different data is strong, Next, the data is processed by the density-based clus-
and has obvious time series characteristics. tering method [26], and the initial seed sets are estab-
Thermal process is usually considered as a transition lished according to the distance from the data centre.
process from a steady state to a new steady state. There-
fore, the thermal process data can be divided into steady-
state condition data and transition condition data. The
characteristic of steady-state data is that it fluctuates in
a small range near the steady-state data. The transition

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

3.4 Clustering data screening based on distance Dij

According to the characteristics of the thermal process

data mentioned in Sect. 3.1, there are three problems in
dividing the steady state and process conditions to be
solved: (1) the problem of the central point of the work-
ing condition, which can be solved by the previous pre-
liminary seed set; (2) the data screening problem with an
obvious, strong connection relationship with the central
point of the working condition, which can be solved by
the clustering data filtering based on distance from cen-
tres; and (3) the problem of determination of the boundary
data between adjacent working conditions, which can be
clarified by the following boundary based on the hybrid
constraint to fulfill.
According to the thermal data characteristics, distance
Algorithm˖multi-condition identification
Dij is defined as a comprehensive index, including the sam-
Input: time sequence( 1 , 2 , 3 ⋯, ),hyperparameter
, , .
pling time length from the centre point of working condi-
Output: table sequence(
tion tdis , the absolute value of amplitude deviation from
1, 2, 3 ⋯, ).
Initial: = 0, = 1, = 2.
the center point of working condition eu, and the absolute
1 Data preprocessing, normalization; value of the deviation between the data change velocity
2 The initial seed set is established by the density-based at the sampling time and the data change velocity at the
clustering method; centre point 𝜕v , and it is expressed as follows:
3 While not finished segmenting time series;
Dij = 𝛼1 tdis + 𝛼2 eu + 𝛼3 ev
4 Use , to divide the data near the
| | | ( ) | | ( ) |
initial seed set into strong dependent data and weak Dij = 𝛼1 |tij − tci | + 𝛼2 |x tij − x(tci )| + 𝛼3 |𝜕 tij − 𝜕(tci )|
| | | | | |
dependent data (boundary area);
(6)
5 strong dependent data processing;
1 = 0.6, 2 = 0.3, 3 = 0.1, calculate composite
where 𝛼i denotes the coefficient of the distance compo-
distance with the following formula 6. Use nent, referring to the intensity of the component, tci
K-mean clustering based on composite distance to denotes the time label of the centre point of the ith case,
cluster the strong dependent data, The data with tij denotes the time label of the jth point of the ith case,
good performance in compactness and separation are
( ) the amplitude of the centre point of the ith
x(tci ) denotes
labeled with category labels, and the data with poor case, x tij denotes the amplitude of the jth point of the
performance are reclassified into weak dependent ith case; 𝜕(tci ) denotes the speed(of )the centre point under
data; the ith working condition, and 𝜕 tij denotes the speed of
6 weak dependent data processing;
the jth point under the(ith( working
) ( ) condition.
( ) (Consider
)) the
= 0.6, = 0.3, = 0.1, calculate composite
1 2 3
t i(m e) s e(r i e) s X( = )x t1 , x t2 , x t3 ⋯ x tn , a n d
distance with the formula 6. The pairwise
Δ tk = x tk − x tk−1 , k = 2, 3, … , n denotes the data
constraints are superimposed on the composite
distance, and the weak dependent data are clustered
increment of� �time series X from tk−1 to tk ,
1 ∑n � �
and labeled with category labels; E = n−1 �Δ ti � denotes the average absolute value of
( ) ( i=2 ) � 1 | (� )|
7 The model performance is modeled by polynomial Δ tk , 𝜕 tk = E |Δ tk | denotes the equalization of the
| |
fitting for all partitioned working condition data, and
data increment of time series X from tk−1 to tk , and it repre-
the accumulated error of the multi-model is taken as
sents the rate of change at time tk , and it is dimensionless
the stop condition;
data.
8 If the accumulated error of the multi-model is
small, end the program, otherwise adjust the value of
Therefore, the data with a strong relationship with the
the parameter , .
working condition center point can be converted into the
end clustering based on distance Dij . The objective function of
the clustering problem is given by:
Fig. 1  The flowchart and pseudocode of the condition identifica- N M
∑ ∑
tion using the semi-supervised clustering with mixed constraints J= Dij2 (7)
i=1 j=1

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

The clustering based on distance Dij can be used to jitter problems. According to the category continuity
divide the data with strong affiliation into datasets cor- characteristics of adjacent points, the pairwise constraint
responding to different working conditions. is improved by superimposing on the distance constraint,
and the results are as follows:
3.5 Clarification of class boundaries based on mixed N M
constraints J=
∑ ∑
Dij2 +
∑
wij +
∑
w ij
i=1 j=1
xi , xj ∈ M x i , xj ∈ C (8)
The boundary data between adjacent working condi-
tions refer to the data with a weaker relationship with the s.t.li ≠ li s.t.li = li
center point of a working condition. From the perspec- where M and C are the given Must-link set and Cannot-link
tive of modeling, such data may belong to two adjacent set respectively; wij , w ij is the penalty weight for violating
working conditions at the same time; specifically, the Must-link and Cannot-link constraint rules, respectively.
closer the data to the boundary point are, the more obvi- In semi-supervised clustering method with pairwise con-
ous the multi-condition attribute of the data is. However, straints, the constraint set satisfying li = lj is called the
the problem of jitter and jumping of the condition attrib- Must-link set, and the constraint set satisfying li ≠ lj is called
ute needs to be avoided. The standard K-means clustering Cannot-link set.
algorithm cannot overcome this problem. For example, a In Eq. 8, Dij is the composite distance mentioned above,
set of typical thermal data, namely the bed temperature which is composed of the change characteristics of data
data of a fluidized bed boiler, is used for clarification. It and the time span from the central point of the category.
includes 3000 sampling points with a sampling interval of It makes the category of data clearer. Pairwise constraints
5 s. The standard K-means clustering is used to classify the reduce the frequent jumps of the categories of boundary
normalized data from the dimension of value. The results data.
are shown in Fig. 2. Moreover, the boundary area data may have multi cat-
In Fig. 2, different colors represent different categories; egory attributes, which can be judged by the change rate
the red line represent the category belonging of the cor- of time series data. For example, when the change rate of
responding sampling time points. The standard K-means sampling points is less than a certain fixed value within
clustering considers only the value, so the category lines a time span, it can be approximately considered that the
are irregular and frequently jittered, and such a classifica- data within the time span is in a steady state. If the time
tion is of little significance. span is just within the boundary interval, the data can
In view of this, mixed constraints are designed to
achieve clearer working condition boundaries and avoid

Fig. 2  The k-mean clustering

results

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

have multi category labels and belong to two adjacent The connection between nodes determines the behav-
categories. ior of neural network. There are many connection meth-
ods, The most common is full connection. All nodes in a
3.6 Hyperparameter optimization layer are connected to all nodes in a higher layer, and there
are other methods such as sparse connection network [32]
In the working condition identification process, there are and direct connection between input and output nodes
three hyperparameters; the first is the time span (time dis- [33]. If the selection of radial basis function is reasonable,
tance) of data with a strong relationship with the center the hidden layer nodes in full connection mode can filter
point of clustering; the second is the time span of multi- the input data well.
category labels in the boundary area; and, the third is the Improving efficiency and ensuring quality is the direc-
time span ts that is used to calculate the change in the data tion of training method improvement. The growing RBFNN
increment at a certain time series. There are three common improves training efficiency by increasing the number of
methods for determining the hyperparameters, manual RBF units at each learning step [32]. It is an effective solu-
method based on experience, machine-assisted method, tion to ensure the training quality by k-fold cross valida-
and algorithm-based method. In this work, the machine- tion of training samples and validation samples [33].
assisted method has been selected. Therefore, there are In summary, the design of RBFNN needs to solve the
two problems to be solved, an optimization objective number of nodes, the type of hidden layer activation func-
function and an optimization method. tion, the connection mode between nodes and the weight
The model performance is modeled by polynomial fit- adjustment method.
ting under the same conditions for all partitioned working
condition data, and the accumulated error of the multi- 4.1 Structure design of RBF neural network
model is taken as an optimization objective function of
the hyperparameter optimization. RBFNN has three layers, such as input layer, output layer
[ n ] and hidden layer. The number of input layer and output
m
∑ 1 ∑( ( ) ) layer nodes is easy to determine, which is determined
E= y x , 𝜔 − yij (9)
i=i
2 j=1 ij ij according to the characteristic variables of the model. In
this paper, the number of input nodes Ninput is determined
The conventional grid-search method has been by the composite distance Dij , including the data value and
selected as the optimization method. change rate at the current sampling time, and the data
value and change rate at the adjacent sampling time,
Ninput = Nrange × Nsample (10)
4 Online condition recognizer based
on RBFNN where Nrange is the range of adjacent data and Nsample is the
number of the reference information at the sampling time.
RBFNN is considered to be one of the most promising algo- The number of hidden layer nodes Nhidden is determined
rithms. Compared with other artificial neural networks, by the dimension of input data Ninput and the number of
RBFNN is more popular because of its simple structure, conditions Ncond . The number of working conditions can be
fast learning process and appropriate approximation abil- determined by machine learning method [12],
ity [27, 28]. There are many factors that affect the perfor-
mance of RBFNN, such as the network structure, hidden Nhidden = Ninput × Ncond (11)
layer activation function, the connection between nodes,
training methods, and so on. Many researchers have given Considering the nonlinearity of the thermal process
good suggestions in these methods. Mosavi et al. analyzed data, radial basis function can describe the information
the hidden layer structure of RBF neural network and pro- of conditions (The parameter C of radial basis function
posed an efficient training method, which uses Stochastic represents the center of condition, and the closer to the
Fractal Search Algorithm (SFSA) for training RBFNN [29]. center, the greater the output value). Therefore, standard
By choosing a more reasonable radial basis function, the RBF (SRBF), Cauchy RBF (CRBF), inverse multiquadric RBF
problem of slow classification speed of RBFNN is solved and generalized inverse multi-quadric Functions can meet
[28, 30, 31]. the basic requirements of the algorithm.

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

The online condition recognizer realizes the mapping

from continuous data to discrete data. By growing the
number of hidden layer nodes, the nonlinear approxi-
mation ability of the network will be improved, and the
amount of computation will also be increased. By increas-
ing the number of adjacent sampling time and the data
information of sampling time, the recognition rate of the
network can be improved.

5 Simulation analysis

Fig. 3  Structure of RBFNN 5.1 Condition division

In this section, the typical thermal data set used above is

used for simulation. The normalized data is shown in the
4.2 Weight calculation of RBFNN
scatter in Fig. 2.
Based on what is explained, the RBFNN structure com-
5.1.1 Compare the single time constraint with mixed
prises three layers as shown in Fig. 3. The input of the ith
constraints
hidden neuron si,
[ ]
Based on the changing trend of the data, there are many
si = x1 𝜔h1,i , x2 𝜔h2,i , x3 𝜔h3,i , … xn 𝜔hn,i , … xN 𝜔hN,i (12)
working conditions. The clustering analysis is used to
divide the working conditions, and the results of standard
where n is the index of input; i is the index of hidden unit;
K-means clustering are shown in Fig. 2. The category labels
xn is the nth input and 𝜔hn,i is the input weight between nth
of the data clearly showed that in the adjacent boundary
input and ith hidden unit.
areas of different working conditions, the data category
The output oj of jth neuron is calculated as follows:
had repeated jumping process, which differed from prior
P
∑ ( ) knowledge of condition division. After adding a single
oj = 𝜑p sp 𝜔op,j + 𝜔o0,j (13) time tag constraint, the classification results shown in
p=1 Fig. 4 are obtained.
Here j is the index of output, 𝜔op,j is the output weight In Fig. 4, the method of Semi-supervised clustering
results with single time constraint is shown to have supe-
between the pth hidden neuron and output neuron j , and
rior performance compared to the standard K-mean clus-
𝜔o0,j is the bias weight of the jth output neuron.
tering, which has been shown in Fig. 2. But, When the time
The weight adjustment value is obtained by gradient tag is around 400 and 2350, there is still the phenomenon
descent method [34]: of category jump, such as the change of category label in
( ( )) ( ) the enlarged area.
Δωij = 𝜂1 yi(k) − fi x (k) 𝜑j x (k) (14) The proposed method is used to conduct semi-super-
vised clustering with mixed constraints, and the cluster-
( ) x − μj ∑m ( ( (k) )) ing results are shown in Fig. 5. Comparing with the the
(k)
Δμj = 𝜂2 𝜑j x (k) ω ij y i
− f i x (15) enlarged area in Fig. 4, the obtained category label of the
𝜎j3 i=1
data shows that the phenomenon of the repeated jump
of data type has been weakened. In Fig. 5, near the time
2 m
( ) x − μj ∑ ( ( (k) )) tags 400 and 2350, there are data classification overlaps
(k)
Δ𝜎j = 𝜂3 𝜑j x (k) ω ij y i
− f i x (16) between the two categories, which verifies the previous
𝜎j3 i=1
analysis of the thermal process data characteristics.

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Fig. 4  Semi-supervised cluster-

ing results with a single time
constraint

Fig. 5  Semi-supervised cluster-

ing with mixed constraints

5.1.2 Comparison of different number of initial seed sets of seed sets. If the number of initial seed sets is defined,
the location can be determined by the density-based clus-
In addition to the hyperparameters mentioned in Sect. 3.6, tering method. The influence of different number of initial
the information of the initial seed sets will also affect the seed sets on segmentation results is shown in Figs. 6 and 7.
segmentation results, it includes the location and number

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

Fig. 6  Semi-supervised cluster-

ing with mixed constraints and
four initial seed sets

Fig. 7  Semi-supervised cluster-

ing with mixed constraints and
six initial seed sets

Compare Figs. 5, 6 and 7, setting a smaller number of primary air flow, second air flow and fuel flow are selected
initial seed sets can make the segmentation of the origi- as an input, and the model polynomial fitting is performed.
nal data clearer, but the segmentation effect is evaluated The results of model polynomial fitting on segmented data
by the model polynomial fitting in segmented data. On are shown in Table 2. The segmentation effect of setting
this basis, the data of each working condition is taken as five initial seed sets is better than setting four or six initial
an output, the four group of data such as bed pressure, seed sets.

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Table 2  Comparison results No. of ini- Segmenting Polynomial fitting results Fitting Overall
of different number of initial tial seeds score (R2) error
seeds Regression coefficient Intercept (MSE)

4 (0, 858) [− 0.26, − 0.52, 1.14, 1.27, − 1.14, − 0.48] 0.44 0.78 0.023
(859, 1468) [− 0.23, 0.52, − 0.13] 0.33 0.76
(1469, 2270) [− 0.28, 0.51, 0.003] 0.30 0.39
(2230, 3028) [− 0.29, − 0.65, 1.22, 1.44, − 1.25, − 0.47] 0.45 0.74
5 (0, 367) [− 0.32, 0.34, − 0.51, 1.91] 0.35 0.71 0.011
(316, 858) [− 0.23, − 0.07, − 2.28, 0.34] 2.69 0.65
(859, 1468) [− 0.19, 0.42, 0.003, 0.46] 0.16 0.75
(1469, 2270) [− 0.006, 0.33, − 0.003, − 0.32] 0.46 0.55
(2253, 3028) [− 0.25, 0.16, 0.06, 0.04] 0.48 0.69
6 (0, 367) [− 0.47, 0.52, − 1.49, − 2.11, 1.51, 0.99] 1.45 0.69 0.037
(316, 858) [− 0.28, − 0.09, − 2.26] 2.84 0.59
(859, 1075) [0.008, 0.46, 2.36, 0.05, − 2.59, − 0.36] − 0.02 0.52
(1076, 1646) [1.26, 0.41, − 1.45, − 1.89, 0.15, 1.50] 0.50 0.48
(1482, 2270) [1.38, 0.33, − 1.45, − 1.94, 0.04, 1.60] 0.50 0.57
(2230, 3028) [− 0.26, 0.18, 0.05] 0.50 0.72

5.1.3 Compare semi‑supervised clustering with mixed the method in this paper. The polynomial fitting results
constraints with Sliding windows, bottom‑up, given in Table 3 show that the semi-supervised clustering
top‑down with mixed constraints can achieve the condition identifi-
cation, and compared with Sliding Windows, Bottom-Up,
The data of each working condition is taken as an out- Top-Down, the mixed constraint is better for the division
put, the four group of data such as bed pressure, primary of working conditions.
air flow, second air flow and fuel flow are selected as an
input, and the model polynomial fitting is performed. 5.2 Online condition recognizer
The comparison between the clustering results obtained
using mixed constraints, Sliding Windows, Bottom-Up, The online condition recognizer realizes the mapping
Top-Down are shown in Fig. 8, The specific comparative from continuous data to discrete data. By increasing the
data is shown in Table 3. number of hidden layer nodes, the network can perform
( ) perfectly in training data, but the network has poor nor-
u
R2 = 1 − , malization ability, as shown in Fig. 9.
v
Here, the weight coefficients 𝛼i of composite distance
( )2 Dij are equal. Increase the amount of input data informa-
n n n
∑ ( )2 ∑ 1∑ tion, such as Nrange and Nsample can improve the ability of
u= yi − ŷ i , v = yi − y (17) normalization, but the amount of computation increases
i=1 i=1
n i=1 i
significantly. In general, using PSO (particle swarm optimi-
zation) to optimize the weight coefficients 𝛼i of composite
In Table 3, the calculation formula of fitting score is distance Dij can also improve the network normalization
shown in formula 17, different methods have different ability, but it does not increase the amount of calcula-
segmentation results, Sliding Windows, Bottom-Up, and tion, as shown in Fig. 10. The optimization process of
Top-Down achieve good results in the linear fitting of weight coefficients 𝛼i is shown in Fig. 11. Therefore, other
local sub model, but the results in different segments have parameters, such as Nrange , Nsample and the characteristic
large deviation, and the overall effect is not as good as

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

Fig. 8  Polynomial fitting

on different segmentation
methods

(a) Semi-supervised clustering with mixed constraints

(b) Sliding Windows

parameters of radial basis function in hidden layer, can be numerical component of the input data and weakening
further optimized by PSO. the proportion of the sequence information can improve
On the test data set, when the error rate is low, the normalization ability of the network. On the contrary, it
𝛼1 = 0.96, 𝛼2 = 3.60, 𝛼3 = 1, the ratio of the weight coef- is necessary to enhance the proportion of sequence infor-
ficient shows that increasing the proportion of the mation to improve the ability of condition identification.

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Fig. 8  (continued)

(c) Bottom-Up

(d) Top-down

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

Table 3  Comparison results of different segmentation methods

Method Condition Segmenting Polynomial fitting results Fitting score Overall
No (R2) error
Regression coefficient Intercept (MSE)

Sliding Windows 1 (0, 863) [− 0.19, 1.15, − 1.18, 0.77] 0.94 0.71 0.013
2 (864, 1515) [− 0.23, 0.50, − 0.08, 0.33] 0.23 0.82
3 (1516, 2080) [− 0.33, 0.39, 0.18, − 0.28] 0.40 0.66
4 (2081, 2845) [0.04, 0.26, 0.03, 0.24] 0.07 0.74
5 (2846, 3028) [− 0.13, − 0.11, − 0.02, − 0.08] 0.75 0.14
Bottom-Up 1 (0, 858) [− 0.17, 0.90, − 1.09, 0.06] 1.27 0.76 0.013
2 (859, 1463) [− 0.25, 0.53, − 0.07, 0.66] 0.17 0.76
3 (1464, 1970) [− 0.38, 0.43, 0.06, − 2.01] 1.13 0.67
4 (1971, 2429) [0.16, 0.57, − 0.61, 0.10] 0.35 0.22
5 (2430, 3028) [− 0.14, 0.21, 0.09, − 0.13] 0.49 0.78
Top-Down 1 (0, 67) [− 0.49, − 0.33, 0.70, − 0.66] 0.98 0.27 0.015
2 (68, 1077) [− 0.38, 1.64, − 0.96, 1.43] 0.26 0.88
3 (1078, 1970) [− 0.18, 0.78, − 0.34, − 0.16] 0.35 0.82
4 (1971, 2429) [0.16, 0.42, − 0.17, 1.11] − 0.57 0.50
5 (2430, 3028) [− 0.14, 0.21, 0.09, − 0.14] 0.49 0.78
Semi-supervised clustering 1 (0, 367) [− 0.32, 0.34, − 0.51, 1.91] 0.35 0.71 0.011
with mixed constraints 2 (316, 858) [− 0.23, − 0.07, − 2.28, 0.34] 2.69 0.65
3 (859, 1468) [− 0.19, 0.42, 0.003, 0.46] 0.16 0.75
4 (1469, 2270) [− 0.006, 0.33, − 0.003, − 0.32] 0.46 0.55
5 (2253, 3028) [− 0.25, 0.16, 0.06, 0.04] 0.48 0.69

6 Conclusion time span of the child window. The online condition iden-
tifier is designed, and the network normalization ability
The multi-model approach has been proved to be very is improved by optimizing the weight coefficient of input
effective in describing complex processes. The overall information. The simulation results show that the pro-
model precision is affected by the sub-model form and posed method is feasible in dividing the sub-windows, and
division of multiple model sub-windows. Once the time the overall error of the sub-model established is improved.
span of the sub-window is established, the modeling pro- One limitation of this study is that the method pre-
cess within the sub-window is the same as that of a sin- sented in this paper is based on historical data and cannot
gle model. Therefore, the division of sub-windows greatly be segmented online. Although the online recognizer is
affects the overall precision of the multi-model method. designed, it is still offline in essence. It limits the applica-
In this paper, using machine learning and combining tion of the method. Improving it into an online method is
the characteristics of thermal process data, the semi- the next research content.
supervised clustering with mixed constraints is used to
realize condition segmentation and sharpen division of

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Fig. 9  RBFNN performance

before optimizing weight coef-
ficient 𝛼i

(a) Training results

(b) Validation results

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

Fig. 10  RBFNN performance

after optimizing weight coef-
ficient 𝛼i

(a) Training results

(b) Validation results

Vol.:(0123456789)
Research Article SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y

Fig. 11  Error rate curve

Funding This work is supported by the Central Universities Funda- 4. Choi SW, Martin EB, Morris AJ (2005) Fault detection based on a
mental Research Fund under Grant 2019MS098. maximum-likelihood principal component analysis (PCA) mix-
ture. Ind Eng Chem Res 44:2316–2327
Availability of data and material All the data pertaining to this study 5. Keogh E, Chu S, Hart D, Pazzani M (2002) An online algorithm for
are available upon request. segmenting time series. In: Proceedings 2001 IEEE international
conference on data mining, IEEE
6. Liu X, Lin Z, Wang H (2008) Novel online methods for time series
Declarations segmentation. IEEE Trans Knowl Data Eng 20(12):1616–1626
7. Ge ZQ, Song ZH (2009) Multimode process monitoring based
Conflict of interest The authors declare that they have no conflict of on Bayesian method. J Chemom 23(12):636–650
interest. 8. Zhu ZB, Song ZH, Palazoglu A (2012) Process pattern construc-
tion and multi-mode monitoring. J Process Control 22:247–262
Open Access This article is licensed under a Creative Commons Attri- 9. Lu NY, Gao FR, Wang FL (2004) A sub-PCA modelling and online
bution 4.0 International License, which permits use, sharing, adap- monitoring strategy for batch processes. AIChE J 50(1):255–259
tation, distribution and reproduction in any medium or format, as 10. Zhao CH, Wang FL, Lu NY, Jia MX (2007) Stage-based soft-tran-
long as you give appropriate credit to the original author(s) and the sition multiple PCA modelling and on-line monitoring strategy
source, provide a link to the Creative Commons licence, and indicate for batch processes. J Process Control 17(9):728–741
if changes were made. The images or other third party material in this 11. Ling W, Hui Z (2021) Fuzzy segmentation of multivariate
article are included in the article’s Creative Commons licence, unless time series with KPCA and G-G clustering. Control Decis
indicated otherwise in a credit line to the material. If material is not 36(1):115–124
included in the article’s Creative Commons licence and your intended 12. Zhang Y, Zhang B, Zheng Wu (2019) Multi-model modeling of
use is not permitted by statutory regulation or exceeds the permitted CFB boiler bed temperature system based on principal compo-
use, you will need to obtain permission directly from the copyright nent analysis. IEEE Access 8:389–399
holder. To view a copy of this licence, visit http://​creat​iveco​mmons.​ 13. Song B, Tan S, Shi HB (2016) Key principal components with
org/​licen​ses/​by/4.​0/. recursive local outlier factor for multimode chemical process
monitoring. J Process Control 47:136–149
14. Ye Lv, Zhong YH (2014) A multi-model approach for soft sen-
sor development based on feature extraction using weighted
References kernel fisher criterion. Chin J Chem Eng 22(2):146–152
15. Wei L, Yu-pu Y, Na W (2008) Multi-model LSSVM regression mod-
1. Hwang DH, Han C (1999) Real-time monitoring for a process eling based on kernel fuzzy clustering. Control Decis 23(5):560–
with multiple operating models. Control Eng Pract 7(7):891–902 562, 566
2. Zhao SJ, Zhang J, Xu YM (2006) Performance monitoring of 16. Shu-Mei Z, Fu-Li W, Shuai T, Shu W (2016) A fully automatic
processes with multiple operating modes through multiple PLS offline mode identification method for multi-mode processes.
models. J Process Control 16:763–772 ACTA Autom Sin 42(1):60–80
3. Natarajan S, Srinivasan R (2010) Multi-model based process con- 17. Hongyue G (2017) Multivariate time series segmentation and
dition monitoring of offshore oil and gas production process. prediction approach and application research. Dalian University
Chem Eng Res Des 88:572–591 of Technology, Dalian
18. Basu S, Banerjee A, Mooney R (2002). Semi-supervised cluster-
ing by seeding. In: Proceedings of the nineteenth international
conference on machine learning (ICML 2002), pp 19–26

Vol:.(1234567890)
SN Applied Sciences (2022) 4:194 | https://doi.org/10.1007/s42452-022-05076-y Research Article

19. Wagstaff K, Cardie C (2000). Clustering with instance-level con- 29. Mosavi MR, Khishe M, Hatam Khani Y, Shabani M (2017) Train-
straints. In: Proceedings of the 17th international conference on ing radial basis function neural network using stochastic fractal
machine learning (ICML 2000), pp 1103–1110 search algorithm to classify sonar dataset. Iran J Electr Electron
20. Deng C, Guo MZ (2008) Tri-training and data editing based semi- Eng 13(1):100–111
supervised clustering algorithm. J Softw 19(3):663–673 30. Montazer GA, Giveki D (2015) An improved radial basis func-
21. Zhu Y, Qian J, Ji Z (2015) An improved COP-Kmeans algorithm tion neural network for object image retrieval. Neurocomputing
based on BFS. Beijing: China science and technology paper 168:221–223
online. http:// ​ w ww. ​ p aper. ​ e du. ​ c n/ ​ r elea ​ s epap ​ e r/ ​ c onte ​ n t/​ 31. Thandar AM, Khine MK (2012) Radial basis function (RBF) neural
201507-​93 network classification based on consistency evaluation meas-
22. Li CM, Xu SB, Hao ZF (2017) Cross-entropy semi-supervised clus- ure. Int J Comput Appl 54(15):20–23
tering based on pairwise constraints. Pattern Recogn Artif Intell 32. Vachkov G, Stoyanov V, Christova N (2015) Growing RBF network
30(7):598–608 models for solving nonlinear approximation and classification
23. Chen ZY, Wang HJ, Hu M et al (2017) An active semi-supervised problems. In: Proceedings 29th European conference on model-
clustering algorithm based on seeds set and pairwise con- ling and simulation
straints. J Jilin Univ Sci Edn 55(3):664–672 33. Duliba KA (1991) Contrasting neural nets with regression in
24. Chang Yu, Ji-Ye L, Jia-Wei G, Jing Y (2012) A semi-supervised predicting performance in the transportation industry. In: Pro-
clustering algorithm based on seeds and pair-wise constraints. ceedings of the twenty-fourth Hawaii international conference
J Nanjing Univ Natl Sci 48(4):405–411 on system sciences. IEEE
25. Wei S, Li Z, Zhang C (2018) Combined constraint-based with 34. Diaconiţa I, Leon F (2011) A learning model for intelligent agents
metric-based in semi-supervised clustering ensemble. Int J using radial basis function neural networks with adaptive train-
Mach Learn Cybern 9(7):1085–1100 ing methods. Bul Inst Politeh Iaşi Autom Control Comput Sci
26. Rodriguez A, Laio A (2014) Clustering by fast search and find of Sect 57(61)(2):9–20
density peaks. Science 344:1492–1496
27. Laleh MS, Razaghi M, Bevrani H (2020) Modeling optical filters Publisher’s Note Springer Nature remains neutral with regard to
based on serially coupled microring resonators using radial jurisdictional claims in published maps and institutional affiliations.
basis function neural network. Soft Comput 25(1):585–598
28. Dash CSK, Behera AK, Dehuri S et al (2016) Radial basis function
neural networks: a topical state-of-the-art survey. Open Comput
Sci 6(1):33–63

Vol.:(0123456789)