Friction 12(6): 1154–1164 (2024) ISSN 2223-7690

https://doi.org/10.1007/s40544-023-0786-y CN 10-1237/TH
RESEARCH ARTICLE

Classification and spectrum optimization method of grease

based on infrared spectrum
Xin FENG1,2,*, Yanqiu XIA1,*, Peiyuan XIE1, Xiaohe LI1
1
School of Energy Power and Mechanical Engineering, North China Electric Power University, Beijing 102206, China
2
State Key Laboratory of Solid Lubrication, Lanzhou Institute of Chemical Physics, Chinese Academy of Sciences, Lanzhou 730000, China
Received: 12 March 2023 / Revised: 08 May 2023 / Accepted: 07 June 2023
© The author(s) 2023.

Abstract: The infrared (IR) absorption spectral data of 63 kinds of lubricating greases containing six different
types of thickeners were obtained using the IR spectroscopy. The Kohonen neural network algorithm was
used to identify the type of the lubricating grease. The results show that this machine learning method can
effectively eliminate the interference fringes in the IR spectrum, and complete the feature selection and
dimensionality reduction of the high-dimensional spectral data. The 63 kinds of greases exhibit spatial
clustering under certain IR spectrum recognition spectral bands, which are linked to characteristic peaks of
lubricating greases and improve the recognition accuracy of these greases. The model achieved recognition
accuracy of 100.00%, 96.08%, 94.87%, 100.00%, and 87.50% for polyurea grease, calcium sulfonate composite
grease, aluminum (Al)-based grease, bentonite grease, and lithium-based grease, respectively. Based on the
different IR absorption spectrum bands produced by each kind of lubricating grease, the three-dimensional
spatial distribution map of the lubricating grease drawn also verifies the accuracy of classification while
recognizing the accuracy. This paper demonstrates fast recognition speed and high accuracy, proving that
the Kohonen neural network algorithm has an efficient recognition ability for identifying the types of the
lubricating grease.

Keywords: grease; infrared (IR) spectroscopy; layered Kohonen network; species recognition; spectrum band
optimization

1 Introduction promoted the development of tribology. Given the

broad variety and application range of different
Grease is a two-phase lubricant composed of a thickener greases, each exhibits distinctive performance and
and additives dispersed in a base oil. It has a wide quality indices [17–19]. Therefore, it is critical to
range of applications, with many varieties and selecting a suitable grease for any given equipment.
brands, and is primarily used for bearings and gear Failure to do so can result in machine breakdowns
transmissions. and grave consequences like fatal accidents. Currently,
In recent years, several scholars have summarized recognizing various types of grease via sensory analysis
progress in machine learning-based friction and remains a challenging task. While the physicochemical
wear research [1–4], while others have made analysis of grease is objective, it entails numerous
beneficial attempts in areas such as dry friction, arduous, complex, and time-consuming steps. Hu
elastohydrodynamic lubrication, lubricating oil, and et al. [20] conducted classification research on wear
grease [5–16]. These studies have provided the ample particles for tribology classification. Similarly, despite
scope for tribological research in this field and greatly the common practice of using various greases for

* Corresponding authors: Xin FENG, E-mail: fengxinemail@ncepu.edu.cn; Yanqiu XIA, E-mail: xiayq@ncepu.edu.cn
Friction 12(6): 1154–1164 (2024) 1155

many years, there are few studies on recognizing distribution of grease types under the IR spectrum
different grease types [21]. is provided, which validates the spatial clustering of
Determining whether the proper grease has been grease under the preferred spectrum. The recognition
selected and whether it has become ineffective during model adopts hierarchical modeling [27], where each
use can be difficult to accurately judge based on designated layer recognizes only one type of grease,
the appearance alone. However, the infrared (IR) and each recognized grease is denied entry to the
spectroscopy is a powerful analytical technique that next layer. This approach enhances the accuracy of
offers numerous advantages. It is capable of detecting the model and minimizes the amount of data and
solid, liquid, or gaseous samples, and the position operating time required for modeling.
and intensities of the IR absorption peaks provide Recently, various applications of artificial intelligence
information about the molecular structure of a technology in tribology have been demonstrated. For
substance. This information can be used to identify instance, Senthil Kumar et al. [28] utilized artificial
the structural composition of unknown substances neural networks to predict the abrasive wear behavior,
or determine their chemical groups. Moreover, the while Martin-del-Campo et al. [29] collected acoustic
absorption intensities of these bands are related emission signals from rolling bearings and employed
to the contents of chemical groups, making the IR machine learning to effectively detect contaminants
spectroscopy useful for quantitative analysis and in their lubricating oil. For machine learning to
purity identification. For the type identification of become a valuable approach to predicting the friction
the lubricating grease in use as well as for quality and wear behavior, extensive exploration is necessary
inspections of products delivered from the factory, in the field of tribology, including friction, wear, and
composition analysis by the IR spectroscopy is lubrication. However, in recent years, most articles
beneficial not only for evaluating the condition of on the application of machine learning to classification
loss and denaturation of the composition but also for problems have used classification algorithms to
guiding better equipment lubrication maintenance. classify all categories at once, even if they model
A reliable elemental analysis and identification of the different types separately through multiple iterations.
lubricant composition form the basis for choosing This approach retains the already classified categories
the right grease for particular machinery. Therefore, as part of the training set for computation, leading to
the main objective of this paper is to propose a safe, accumulated training errors and ultimately affecting
cost-effective, and efficient solution for recognizing the classification accuracy. The Kohonen neural network
similar samples [21–24]. The IR spectroscopy has a algorithm is a basic stochastic kind of self-organizing
wide range of applications, but certain factors such feature network or map [30] that is trained to
as low transmittance in some IR spectral bands, distinguish certain input clusters in unsupervised
overlapping absorption bands, and multiple correlations fashion. The network weight is adjusted by self-
within the spectral data require the elimination or organized mapping that converges to a state under
minimization of redundant information to enhance the learning process and remains especially sensitive
the accuracy. To determine the relationship between to only one input mode for a particular state. In
the variable of the spectral feature band and the Ref. [21], a cuckoo search algorithm was proposed to
corresponding substance, a mathematical model is optimize the IR spectral bands for identifying three
constructed based on the unique characteristic of the types of lubricants. However, it suffered from low
IR spectral variable of the spectral feature band and recognition efficiency and limited ability to identify
its specific corresponding component, which differs different types of lubricants. As an extension of
from the principal component analysis (PCA) feature our previous work [12, 13], we propose a hierarchical
extraction method [25, 26]. This study specifically classification structure using the layered Kohonen
selects an original variable, establishes the relationship networks to classify different types of lubricants based
between an IR spectral band and grease type, and on their input pattern sensitivity. This approach
clarifies the physical meaning of the selected variable. reduces error accumulation, resulting in more accurate
To better express this relationship, a three-dimensional classification.

www.Springer.com/journal/40544 | Friction
1156 Friction 12(6): 1154–1164 (2024)

In this paper, we propose a method to solve the an input layer, a competition layer, and an output
problem of effective spectral band recognition and layer. The input layer has one neuron for each
optimization of oils and greases. First, we obtained the selected feature spectrum, corresponding to the
IR spectra of various types of grease and determined number of input variables. The competition layer
the order of grease and category recognition. Then, presents a geometrically ordered two-dimensional
we established a hierarchical recognition model using array distribution. The output layer predicts the
the Kohonen network and derived several recognition type of grease being considered. During the training
models by changing the input spectral bands of the process, the model automatically adjusts the weight
model. We tested the recognition accuracy of various vectors of the winning node in the competition layer
derived models and selected the most suitable one and between the winning node of the competition
among them. The optimal grease recognition model layer and the output node. In the testing process, the
is based on the obtained IR spectrum of each grease, unknown samples are clustered by identifying the
which integrates the IR spectrum analysis, neural competition layer node closest to the unknown sample
network, and spectroscopic theoretical methods. This as the winning node. The corresponding category
combination collectively facilitates resolving the issue represented by the output layer node with the largest
of grease type recognition, and hence formulates a connection weight value to the winning node is
novel idea for optimizing the characteristic IR bands. considered to be the category of the unknown input
sample. A schematic diagram of the Kohonen network
structure is shown in Fig. 1.
2 Experimental The specific training steps of the recognition model
are outlined as follows:
2.1 Analytical instruments and samples
1) Network initialization: The weight between the
Sixty-three types of grease were collected, including initialization input layer and the competition layer
five varieties such as calcium sulfonate composite is recorded as wij , while the weight between the
grease, polyurea grease, lithium-based grease, lithium competition layer and the output layer is denoted
complex grease, and aluminum (Al)-based grease as as w jk . The competition layer consists of m nodes,
well as three types of bentonite grease (12 × 5 + 3). and the learning rate, number of iterations, and
The used experimental instrument was a Fourier learning radius are represented by  , maxgen, and r,
transform infrared (FTIR) spectrometer (Nicolet iS5, respectively.
Thermo Scientific) with a spectral range of 350–7,800 cm−1 2) The IR spectral band X(  ( x1 , x2 , xn )) is inputted,
and a KBr window. The transmission spectral feature and the Euclidean distance between the competition
band was 400–7,800 cm−1, and the transmission rate layer and the input sample is calculated as
was more than 92% for a data interval of 1.928 cm−1. m
Acquisition settings included 16 scans at a resolution d j  ( xi  wij )2 j  1, 2, n (1)
of 4 and a scanning speed of 0.10−2 cm/s. The device i 1

utilized a wireless air-cooled grow light source to

3) Comparing the d j : The competition layer neurons
distribute the optimum energy in the fingerprint
area, which has the highest number of peaks of the
compounds, enhancing the optical energy utilization
efficiency. The wireless design ensured that the optical
platform remained sealed and dry, while the obtained
IR spectral data and images were automatically
displayed on the computer.

2.2 Structure of Kohonen network recognition model

The Kohonen neural network consists of three layers: Fig. 1 Kohonen network structure.

| https://mc03.manuscriptcentral.com/friction
Friction 12(6): 1154–1164 (2024) 1157

with the smallest distance are selected as the best not, go back to step 2) and save the recognition model
matching or winning neurons. in case it is over.
4) The weight adjustment, adjusting the node C
and the node weight coefficient present in its domain 2.3 Level network model
Nc(t) is Due to the issue of error accumulation, the recognition
accuracy of the model is limited to only one category
N c (t )   t|find(norm(pos t , posc )  r )  (2)
in each layer. Figures 2(a) and 2(b) illustrate
wij  wij  1 ( Xi  wij ) (3) the structural sketches of the non-hierarchical and
hierarchical network models, respectively. The
where norm is a neuron, and pos t and posc are the non-hierarchical model recognizes only one type
locations of neurons t and c, respectively. For adjusting of grease in each layer, with differentiated grease
the weights of the competition layer and the output data entering the next layer. In comparison to the
node level, the adjustment equation is hierarchical recognition model, the non-hierarchical
network model proves to be more efficient in terms
w jk  w jk   2 (Yk  w jk ) (4)
of general accuracy, as it curtails the amount of data
where Yk is the sample category. The learning rate entering the next level and enhances the training
adjustment equation is adaptively adjusted to speed of the model. In this paper, for each classification
layer, the lubricating grease type with the highest
i classification accuracy in that layer will be selected as
1  1max  (5)
maxgen(1max  1min ) the output of the classification result of that layer.
Additionally, the spectral information of accurately
i
 2   2max  (6) identified samples will be filtered out to prevent
maxgen( 2max   2min )
them from entering the next layer.
i Table 1 demonstrates that the primary difference
r  rmax  (7) between the non-hierarchical and hierarchical network
maxgen(rmax  rmin )
models lies in the adopted different objective functions.
where r is the radius of the field. Among them, η1 and In Table 1, 1 stands for determining the grease
η2 are the learning rates, and rmax and rmin represent category, and 0 stands for determining the non-grease
the upper and lower bounds of r, respectively. category. In the hierarchical network model, the
5) To determine whether the algorithm is over or categories that have been determined in the upper

Fig. 2 Comparison of ideal diagrams of two models.

www.Springer.com/journal/40544 | Friction
1158 Friction 12(6): 1154–1164 (2024)

Table 1 Class number design of model.

Type Non-hierarchical category number Hierarchical category number
Calcium sulfonate 1 0 0 0 0 0 1 — — — — —
Polyurea 0 1 0 0 0 0 0 1 — — — —
Lithium-based 0 0 1 0 0 0 0 0 1 0 — —
Lithium complex 0 0 0 1 0 0 0 0 0 1 — —
Aluminum-based 0 0 0 0 1 0 0 0 0 0 1 —
Bentonite 0 0 0 0 0 1 0 0 0 0 0 1

layer will be removed and will no longer appear and it is also vital to controlling the noise within the
in the data of the next layer to participate in the data within a certain range since excessive noise can
recognition. overshadow required category information needed
for correctly establishing lubricant categories and
2.4 Recognizing band optimization
specific IR absorbance bands, thereby reducing the
Currently, there is no specialized IR spectrometer accuracy of the recognition model. Overall, our
available for analyzing the lubricating grease. The proposed method of analyzing the IR spectra of the
spectrum obtained from the general FTIR contains a lubricating grease by dividing them into sections and
large amount of unnecessary data that are irrelevant analyzing them step by step provides an effective
to the compositions and properties of the grease means for achieving accurate classification of grease
samples being analyzed. Utilizing this redundant data samples while reducing the computational workload
for modeling not only increases the computational and improving the model accuracy.
workload but also prolongs the calculation time
while perturbing the accuracy of the model. Although 3 Results and discussion
compressing spectral data and extracting representative
components can reduce the amount of data, this 3.1 Recognition model training and testing
method lacks significant physical meaning and fails
to establish a correlation between the type of the The original data for the six types of greases are
used grease and the corresponding obtained spectral shown in Fig. 3. All 1,868 spectral points were
band. To address these issues, we adopt a method of entered in the manner, as shown in Table 1, while
dividing the IR spectra into sections and analyzing hierarchical and non-hierarchical training were
them step by step: performed. The parameters used for training are
1) Input the X of the grease one by one in order. as follows: maxgen = 1,000, 1max  0.05, 1min  0.005,
2) Use the established grease recognition model to rmax  1.4, and rmin  0.2.
achieve the recognition accuracy of the model in the The test results of the non-hierarchical recognition
input layer. model are presented in Fig. 4, and the test results
3) Select the model with the highest accuracy for of the first hierarchical recognition model are shown
the grease category recognition. The input IR spectrum in Fig. 5. Both Figs. 4 and 5 plot the 63 data samples
band corresponding to that particular grease type is of grease types on the x-axis and six grease categories
utilized as the characteristic spectrum for that type. (1: calcium sulfonate composite grease, 2: polyurea
The accuracy of the recognition model generally grease, 3: lithium-based grease, 4: lithium complex
depends on the selected data and recognition model. grease, 5: aluminum-based grease, and 6: bentonite
A stable model can properly classify a category even grease) on the y-axis. The model test results depicted
when the input data are disturbed by noise. Therefore, in Figs. 4 and 5 correspond to these categories.
it is critical to selecting a model that can accurately By comparing the recognition results of polyurea
classify categories under normal spectra conditions, grease in Figs. 4 and 5, it is evident that only three

| https://mc03.manuscriptcentral.com/friction
Friction 12(6): 1154–1164 (2024) 1159

Fig. 3 Original spectral data.

and three types of non-polyurea grease were mistakenly

categorized as polyurea grease. These findings confirm
that the hierarchical recognition model significantly
improves the accuracy of the recognition process.

3.2 Band optimization

The accuracy of the recognition model depends on

various IR bands provided by the lubricating grease
inputs. In the first layer, the layered Kohonen network
model recognizes polyurea grease. However, the IR
Fig. 4 Test results of non-hierarchical model. spectroscopic summit for grease fluctuates under the
influence of refining temperature, refining time, types
of thickeners, and technological settings, leading to
a large deviation in distinguishing points if only
spectroscopic points are utilized. To address this issue,
a recognition model of polyurea grease with varied
input bands was created by selecting 100 spectral
data points. Figure 6 shows the test accuracy of the
polyurea grease model under different input bands
after parameter optimization. It demonstrates that the
accuracy of the corresponding coordinates is 100%
within the range of 1,562.0–1,772.2 cm−1 in the abscissa,
Fig. 5 Test results of hierarchical model. which fits with the C=O stretching vibration in the
ureide of the polyurea grease, indicating that the finest
types of polyurea grease were misclassified under recognition model can be achieved within the band
the hierarchical recognition model with complete of 1,562–1,772.2 cm−1 as the characteristic spectrum.
spectrum training. In contrast, the non-hierarchical After polyurea grease is removed, the recognition
model misclassified four types of polyurea grease, process moves to the second layer, where it now

www.Springer.com/journal/40544 | Friction
1160 Friction 12(6): 1154–1164 (2024)

grease, was selected, and the accuracy of the model

corresponding to the IR spectrum range of 1,420–
1,630.2 cm−1 was then obtained, which showed the
highest accuracy of the recognition model. After
undergoing a similar process in the fourth layer,
bentonite grease was recognized. Based on the values
of the horizontal and vertical coordinates in Fig. 9,
the recognition model achieved 100% accuracy when
the spectral regions of 983.5–1,083.8, 1,351.8–1,562.0,
and 2,865.7–3,075.9 cm−1 were selected.
Fig. 6 Classification accuracy of first level. The lithium-based grease was identified in the
fifth layer of analysis. As indicated by the horizontal
recognizes calcium sulfonate composite grease. As and vertical coordinates in Fig. 10, the model achieved
shown in Fig. 7, the test accuracy of the model accuracy of 83.3% in the transverse coordinate
is established for various input spectral bands. band of 2,565–2,775.2 cm−1 and 87.5% in the band of
The spectral bands of the abscissa are found to be 1,550.4–1,760.6 cm−1. After filtering through five
863–1,073.2 cm−1 and 1,186–1,396.2 cm−1, with ordinate layers of recognition, only one unidentified sample
accuracy of 98.04%, indicating that this range produces (a lithium complex grease) remained. However, it
the optimum recognition accuracy. In the spectral is relatively problematic to accurately distinguish
bands of 1,675.8–1,886 cm−1 in the abscissa, the between lithium-based grease and lithium complex
recognition accuracy is 96.08%. grease, given that the accuracy of the fifth layer cannot
The polyurea and calcium sulfonate composite exceed 90%.
greases were eliminated, and the recognition process
progressed to the third layer, where the aluminum-
based grease was identified. Based on the values of
the horizontal and vertical coordinates in Fig. 8, the
accuracy of the model was determined to be 94.87%
when the spectral range of 923.75–1,133.95 cm−1,
which fits with the Al–O stretching vibration of the
aluminum-based grease, was selected. The accuracy
of the model was determined to be 97.4% when the
range of 1,465.6–1,675.8 cm−1, which fit with the
carbonyl stretching triple band of the aluminum-based
Fig. 8 Classification accuracy of third level.

Fig. 7 Classification accuracy of second level. Fig. 9 Classification accuracy of fourth level.

| https://mc03.manuscriptcentral.com/friction
Friction 12(6): 1154–1164 (2024) 1161

transmittance value corresponding to three spectral

points selected from the optimal frequency band
serves as the spatial coordinate value. This structure
establishes a spatial recognition map of various greases
by adjusting the position and orientation of two or
three vertical-dimensional axes. In this observation
space, the clustering behavior of similar greases is
directly observed, and relatively appropriate results
are obtained through the physical explanation of the
concept.
Fig. 10 Classification accuracy of fifth level.
In this study, we selected three spectral points
(1,562.1, 1,631.5, and 1,579.4 cm−1) from the optimized
Based on the aforementioned results, it is evident spectral feature band of polyurea grease. We used the
that during the continuous small band model transmittance of each spectral line corresponding to
recognition test for each layer (using 100 spectral these points to draw the three-dimensional distributions
data points in this study), multiple high recognition of polyurea and other greases, as illustrated in
rate bands are often identified. However, expressing Fig. 11(a). Although the absorption peak information
the discrete and fragmentary optimal bands using of 100 consecutive characteristic spectral points was
a simple and intuitive method or illustrating their selected when establishing the model, only the
clustering effect on a particular type of grease poses transmittance spatial distribution of grease in three
a challenge. Therefore, this paper proposes and tests spectral points was expressed in the constructed
a three-dimensional spatial structure, where the three-dimensional space, indicating the absence of

Fig. 11 Three-dimensional distributions of different greases under IR spectra.

www.Springer.com/journal/40544 | Friction
1162 Friction 12(6): 1154–1164 (2024)

valuable information. Figure 11(b) presents the spatial bentonite grease, and lithium-based grease are as
distributions of calcium sulfonate composite grease follows: 100.00%, 96.08%, 94.87%, 100.00%, and 87.50%,
and other greases at 1795.4, 1187.9 and 833.1 cm−1 respectively.
after eliminating polyurea grease from the layer. The 3) Drawing on selected spectral points from
direction of this axis is different from those of the optimized characteristic bands produces a three-
others due to the more evident effect at this viewing dimensional distribution of various grease types.
angle. Calcium sulfonate composite demonstrated a This process validates the fact that the characteristic
good clustering phenomenon. Furthermore, Fig. 11(c) spectrum retains information on different grease
shows the spatial distributions of aluminum-based categories.
and other greases in the bands at 1,081.9, 1,577.9, and
1606.4 cm−1, excluding calcium sulfonate composite
Acknowledgements
grease. Figure 11(d) demonstrates the spatial
distributions of bentonite and other greases in the The authors would like to appreciate the financial
bands at 2977.6, 1033.7 and 1471.4 cm−1. Finally, Fig. 11(e) support extended for this academic work by the Beijing
represents the spatial distributions of lithium-based Natural Science Foundation (Grant No. 2232066) and
and lithium complex grease in the bands at 2790.5, the Open Project Foundation of State Key Laboratory
1587.1 and 1722.1 cm−1. Our results showed that the of Solid Lubrication (Grant No. LSL-2212)
Kohonen neural network algorithm combined with
the three-dimensional spatial aggregation phenomenon
Declaration of competing interest
of the thickener type improved the accuracy and
speed of grease recognition. However, it was observed The authors have no competing interests to declare
that the clustering phenomenon exhibited by lithium- that are relevant to the content of this article.
based grease and lithium complex grease in the given
spectral space appeared to be ambiguous. The IR Open Access This article is licensed under a Creative
spectrum failed to essentially distinguish the category Commons Attribution 4.0 International License, which
information for these two types of grease. Therefore, permits use, sharing, adaptation, distribution and
future studies should explore other methods to reproduction in any medium or format, as long as you
effectively identify lithium-based and lithium complex give appropriate credit to the original author(s) and
greases. the source, provide a link to the Creative Commons
licence, and indicate if changes were made.
4 Conclusions The images or other third party material in this
article are included in the article’s Creative Commons
After hierarchically training the Kohonen network licence, unless indicated otherwise in a credit line to
model, the recognition accuracy of the model is the material. If material is not included in the article’s
compared under different spectral bands. The model Creative Commons licence and your intended use is
shows the highest accuracy for the spectral bands in not permitted by statutory regulation or exceeds the
the current design. permitted use, you will need to obtain permission
1) By employing the models obtained from both directly from the copyright holder.
non-hierarchical and hierarchical training, the accuracy To view a copy of this licence, visit
of distinguishing polyurea grease is found to be 88.8% http://creativecommons.org/licenses/by/4.0/.
and 95.2%, respectively. These results reveal that the
hierarchical model design can significantly improve
References
the recognition accuracy of the model.
2) In terms of accuracy, the accuracy of the model [1] Meng Y G, Xu J, Ma L R, Jin Z M, Prakash B, Ma T B,
for distinguishing among polyurea grease, calcium Wang W Z. A review of advances in tribology in 2020–2021.
sulfonate composite grease, aluminum-based grease, Friction 10(10): 1443–1595 (2022)

| https://mc03.manuscriptcentral.com/friction
Friction 12(6): 1154–1164 (2024) 1163

[2] Marian M, Almqvist A, Rosenkranz A, Fillon M. Numerical short castor oil fiber reinforced epoxy composites. Friction
micro-texture optimization for lubricated contacts—A critical 9(7): 250–272 (2021)
discussion. Friction 10(11): 1772–1809 (2022) [16] Hasan M S, Kordijazi A, Pohatgi P K, Nosonovsky M.
[3] Rosenkranz A, Marian M, Profito F J, Aragon N, Shah R. Triboinformatic modeling of dry friction and wear of aluminum
The use of artificial intelligence in tribology—A perspective. base alloys using machine learning algorithms. Tribology
Lubricants 9(1): 2 (2021) International 161: 107065 (2021)
[4] Yin N, Xing Z G, He K, Zhang Z N. Tribo-informatics [17] Heirendt L, Liu H H T, Wang P. Aircraft landing gear greased
approaches in tribology research: A review. Friction 11(1): slider bearing steady-state thermo-elastohydrodynamic
1–22 (2023) concept model. Tribol Int 82: 453–463 (2015)
[5] Pandiyan V, Akeddar M, Prost J, Vorlaufer G, Varga M, [18] De Laurentis N, Kadiric A, Lugt P, Cann P. The influence
Wasmer K. Long short-term memory based semi-supervised of bearing grease composition on friction in rolling/sliding
encoder–decoder for early prediction of failures in self- concentrated contacts. Tribol Int 94: 624–632 (2016)
lubricating bearings. Friction 11(1): 109–124 (2023) [19] Cao Z F, Xia Y Q, Ge X Y. Conductive capacity and
[6] Perčić M, Zelenika S, Mezić I. Artificial intelligence-based tribological properties of several carbon materials in
predictive model of nanoscale friction using experimental conductive greases. Ind Lubr Tribol 68(5): 577–585 (2016)
data. Friction 9(6): 1726–1748 (2021) [20] Hu X B, Song J, Liao Z H, Liu Y H, Gao J, Menze B, Liu
[7] Ulas M, Altay O, Gurgenc T, Özel C. A new approach for W Q. Morphological residual convolutional neural network
prediction of the wear loss of PTA surface coatings using (M-RCNN) for intelligent recognition of wear particles
artificial neural network and basic, kernel-based, and from artificial joints. Friction 10(4): 560–572 (2022)
weighted extreme learning machine. Friction 8(6): 1102–1116 [21] Li X H, Feng X, Xia Y Q. IR spectra of grease optimization
(2020) based on cuckoo search. Spectrosc Spect Anal 37(12):
[8] Altay O, Gurgenc T, Ulas M, Özel C. Prediction of wear 3703–3708 (2017) (in Chinese)
loss quantities of ferro-alloy coating using different machine [22] Barbin D F, de Souza Madureira Felicio A L, Sun D W,
learning algorithms. Friction 8(1): 107–114 (2020) Nixdorf S L, Hirooka E Y. Application of infrared spectral
[9] König F, Sous C, Jacobs G. Numerical prediction of the techniques on quality and compositional attributes of coffee:
frictional losses in sliding bearings during start–stop operation. An overview. Food Res Int 61: 23–32 (2014)
Friction 9(3): 583–597 (2021) [23] Leong Y S, Ker P J, Jamaludin M Z, Nomanbhay S M,
[10] Gao X L, Dai K, Wang Z, Wang T T, He J B. Establishing Ismail A, Abdullah F, Looe H M, Shukri C N S M. New
quantitative structure tribo-ability relationship model using near-infrared absorbance peak for inhibitor content detection
Bayesian regularization neural network. Friction 4(2): in transformer insulating oil. Sensor Actuat B-Chem 266:
105–115 (2016) 577–582 (2018)
[11] Parikh H H, Gohil P P. Experimental investigation and [24] Pasquini C. Near infrared spectroscopy: A mature analytical
prediction of wear behavior of cotton fiber polyester technique with new perspectives—A review. Anal Chim
composites. Friction 5(2): 183–193 (2017) Acta 1026: 8–36 (2018)
[12] Xia Y Q, Xu D Y, Feng X, Cai M R. Identification and [25] Geminale A, Grassi D, Altieri F, Serventi G, Carli C,
content prediction of lubricating oil additives based on Carrozzo F G, Sgavetti M, Orosei R, D’Aversa E, Bellucci G,
extreme learning machine. Tribology 40(1): 97–106 (2020) et al. Removal of atmospheric features in near infrared
(in Chinese) spectra by means of principal component analysis and target
[13] Xia Y Q, Wang C, Feng X. GA–BPSO hybrid optimization transformation on Mars: I. Method. Icarus 253: 51–65 (2015)
of middle infrared spectrum feature band selection of [26] Li Z F, Xu G J, Wang J J, Du G R, Cai W S, Shao X G.
lubricating oil additive type identification technology. Outlier detection for multivariate calibration in near infrared
Tribology 42(1): 142–152 (2022) (in Chinese) spectroscopic analysis by model diagnostics. Chin J Anal
[14] Xia Y Q, Wang Y X, Feng X, Cai M R. Optimization Chem 44(2): 305–309 (2016)
efficiency of swarm intelligence search in base oil performance [27] Zhang G, Liu Y, Guo J F, Cheng X Q. A hierarchical search
prediction model. Tribology 43(4): 429–438 (2023) (in result clustering method. J Comput Res Dev 45(3): 542–547
Chinese) (2008) (in Chinese)
[15] Egala R, Jagadeesh G V, Setti S G. Experimental investigation [28] Senthil Kumar P, Manisekar K, Narayanasamy R.
and prediction of tribological behavior of unidirectional Experimental and prediction of abrasive wear behavior of

www.Springer.com/journal/40544 | Friction
1164 Friction 12(6): 1154–1164 (2024)

sintered Cu–SiC composites containing graphite by using with unsupervised acoustic emission feature learning. Tribol
artificial neural networks. Tribol Trans 57(3): 455–471 Int 132: 30–38 (2019)
(2014) [30] Zupan J, Novič M, Ruisánchez I. Kohonen and
[29] Martin-del-Campo S, Schnabel S, Sandin F, Marklund P. counterpropagation artificial neural networks in analytical
Detection of particle contaminants in rolling element bearings chemistry. Chemometr Intell Lab 38(1): 1–23 (1997)

Xin FENG. She received her Ph.D. Engineering, North China Electric Power University,
degree in systems engineering China, in 2010. She is currently an associate
from Northeastern University, professor. Her research areas cover tribology and
China, in 2007. She joined School artificial intelligence.
of Energy Power and Mechanical

Yanqiu XIA. He received his Power and Mechanical Engineering, North

Ph.D. degree in mechanical China Electric Power University, China, in 2010.
engineering from Northeastern His current position is a professor. His research
University, China, in 1999, and areas cover the tribology of mechanical and
was selected as a professor in electrical equipment, oil monitoring, and artificial
2007. He joined School of Energy intelligence.

Peiyuan XIE. He got his bachelor’s China Electric Power University, China, in 2026,
degree in 2020 from North China with a Ph.D. degree in engineering. His research
Electric Power University, China. interests include the tribology of liquid lubrication
He will graduate from North combined with artificial intelligence.

Xiaohe LI. He received his graduate student at North China Electric Power
bachelor’s degree from North University, China. His research interests focus on
China Electric Power University, the research areas covering tribology and artificial
China, in 2018. He is currently a intelligence.

| https://mc03.manuscriptcentral.com/friction