Professional Documents
Culture Documents
Deep Learning-Based Damage Detection of Mining Conveyor Belt
Deep Learning-Based Damage Detection of Mining Conveyor Belt
Measurement
journal homepage: www.elsevier.com/locate/measurement
Keywords: The mining conveyor belt is an important component of the coal mine belt conveyor, which plays the role
Belt conveyor of carrying materials and transmitting power. Aiming at the problem that mining conveyor belts are easily
Conveyor belt damaged under severe working conditions, based on the reclassification and definition of conveyor belt damage
Deep learning
types, a special data set for conveyor belt damage was established and a new detection method that can
Machine vision
simultaneously detect multiple faults based on improved Yolov3 algorithm was proposed. The EfficientNet
Damage detection
was adopted as the backbone feature extraction network instead of Darknet53 in the improved algorithm,
comprehensively considers the balance between network depth, width, and image resolution for network
scaling to improve the accuracy of the algorithm in limited computing resources. Experiments have proved
that the improved algorithm in this paper takes into account both detection speed and detection accuracy.
The detection speed can reach 42 FPS, and the mean average precision can reach 97.26%. Compared with the
original Yolov3 algorithm, the accuracy is increased by 10.4%, with the speed 45.9%, which provides new
ideas and methods for ensuring the safe and stable work of conveyor belts.
∗ Corresponding author.
∗∗ Corresponding author at: College of Mechanical and Electronic Engineering, Shandong University of Science and Technology, Qingdao, SD 266590, China.
E-mail addresses: shihao@sdust.edu.cn (H. Shi), sdustzms@163.com (M. Zhou).
https://doi.org/10.1016/j.measurement.2021.109130
Received 22 November 2020; Received in revised form 4 January 2021; Accepted 31 January 2021
Available online 8 February 2021
0263-2241/© 2021 Elsevier Ltd. All rights reserved.
M. Zhang et al. Measurement 175 (2021) 109130
mainly divided into vision-based deviation monitoring and vision-based far as we know, the application of deep learning technology in the
tearing monitoring. Christopher C. Fromme [29] at Carnegie Mellon detection of conveyor belt damage is rarely reported.
University first proposed and developed a conveyor belt monitoring In order to meet the needs of the intelligent development of belt
system based on camera and image processing technology in 2003, conveyors, the article first proposed a mining conveyor belt damage
mainly for the damage of conveyor belt joints monitoring. Yan L. detection method based on deep learning, and proposed an improved
et al. [30] proposed a fast image segmentation algorithm to simulta- Yolov3 algorithm for conveyor belt damage detection, which was veri-
neously detect longitudinal tearing and deviation of the conveyor belt fied by field experiments with better detection accuracy and real-time
from the binary belt image. Ref. [31] mainly focused on the research performance. The main contributions of this paper are summarized as
of vision-based belt conveyor deviation detection, and a conveyor belt follows: (1) Better than most current conveyor belt damage detection
edge detection algorithm based on improved Canny algorithm and methods, the detection method based on deep learning realize the
Hough transform was proposed. Liu W. [32] et al. proposed a tear detection of multiple damage types of conveyor belts, not limited to
detection method for belt conveyors based on visible light and infrared. belt tear. (2) The method proposed in this paper is a more direct
This technology fusion method greatly improves the accuracy of tear detection method, without the aid of other devices or complicated
detection. Xianguo L. et al. [33] explained the application of the line image process. (3) The improved Yolov3 algorithm proposed in this
laser assisted technology in the tear detection of belt conveyors. They paper comprehensively considers the impact of model scaling on the
judged the belt tearing by judging the slope of the line laser or whether detection results, and achieves the best speed-accuracy trade-off on the
there are abnormal pixel points. Yu B. et al. [34] proposed a Dual Band dataset, also could be applied to the detection of other targets in other
Infrared Detection (DBID) method for longitudinal tear detection on fields.
conveyor belts. The method combines mid-infrared and long-infrared
vision, is not limited by visual conditions, and is more accurate and 2. Methodology and raw data
reliable than the integrated binocular visual detection (IBVD) method.
A new early warning method for longitudinal tearing of conveyor belts 2.1. The principle and structure of the improved algorithm
was proposed in Refs. [35,36]. An infrared camera was required to
collect the image, and then the binarization automatic threshold selec- As mentioned earlier, the essence of conveyor belt damage detection
tion algorithm was used to analyze the infrared image to determine is target detection. Supervised learning-based target detection methods
whether the conveyor belt tear warning should be issued. A multi- at this stage can be divided into two categories: Two-stage target
spectrum-based conveyor belt tear detection method was proposed detection algorithms based on candidate regions, which first generate
by Chengcheng Hou et al. [37]. This method combines visible light, candidate frames through regional proposal networks(RPN), and then
mid-infrared light, and far-infrared light, and has high environmental use convolutional neural networks for fractionation and border correc-
adaptability. Hao X.L. et al. [38] proposed a damage detection method tion, which are more accurate but slower and difficult to meet real-time
for belt conveyors based on SVM(support vector machines), which more requirements due to more candidate frames, such as R-CNN series [40–
considered the effectiveness of the method under both dry and wet 42]; And single-stage target detection algorithms based on regression,
conditions. Starting from the cause of the tearing of the conveyor belt, represented by SSD (Single Shot MultiBox Detector) [43] and YOLO
Wang, Y. et al. [39] proposed a method for identifying foreign objects (You only look once) [44,45]. The research in this paper is carried out
on the conveyor belt based on an improved SSD algorithm to prevent based on Yolov3.
the tearing problem from the root cause of the tearing. The target detection network could be divided into two parts, one is
In general, the current damage detection methods for belt con- the backbone feature extraction network, and the other is the prediction
veyors based on machine vision can be divided into three types: 1. network. The quality of the features extracted by the feature extrac-
The method based on image processing, that is, the surface image of tion network directly affects the prediction effect of the prediction
the conveyor belt is acquired by the camera, and the relevant image network. This means if better features, maybe better prediction results.
processing algorithm is used to realize the elimination of noise and Darknet53, as the backbone feature extraction network of Yolov3,
the enhancement of tear characteristics, mostly judged by binarized draws on the design idea of residual neural network (RNN) [46]. It
images; 2. The method for detecting damage of conveyor belts based is composed of 5 residual blocks, each residual block contains multiple
on line laser assistance. The line laser assisted detection method to residual units, and the structure of each residual unit is shown in Fig. 1.
achieve the fracture features extracted from the image to extract the The use of residual structure effectively solves the gradient explosion
characteristics of the laser line transformation, especially the detection problem caused by network deepening. A deeper network also helps to
of angular point, by highlighting the characteristics of laser, make the extract deeper features and improve the accuracy of the network. At
information accessible to extraction, effectively reduce the interference the same time, the Yolov3 draws on the idea of the feature pyramid
of dust; 3. A method based on infrared cameras or thermal imagers. The network (FPN) [47], the multi-scale detection mechanism integrates
method fully considers the reasons for the damage of the conveyor belt, three feature layers of different scales for target detection, improves
and considers the local heating problem of the conveyor belt caused by the network’s ability to detect targets of different scales, and better
the violent friction between foreign objects and the conveyor belt. The compensates for Yolov1 and Yolov2 detection of defects with poor
infrared CCD camera is used to capture and analyze the heat source accuracy for small targets.
distribution on the conveyor belt surface, and then determine whether Compared with Yolov2, the backbone feature extraction network in
any foreign objects are scratching the conveyor belt. Yolov3 is upgraded from Darknet19 to Darknet53. The deepening of
The essence of conveyor belt damage detection is target detection. the network depth brings a 9.8% improvement in accuracy. However,
Deep learning is currently the mainstream research method in the in a limited computing unit, it is very likely to increase the amount of
field of target detection. Compared with traditional target detection calculation as the network depth increases, thereby reducing the speed
methods, deep learning omits manual design features and the selection of target detection.
of classifiers. By simulating a hierarchical structure similar to the In general, three methods usually used for strength and expand the
human brain, it establishes a structure from low-level semantic signals backbone feature extraction network to better extract image features
to high-level semantics, mapping has better a fit for big data, and it and information: increasing the network width (Fig. 2(b)), increasing
is easier to meet the requirements of image and video big data for the network depth (Fig. 2(c)), and increasing the resolution of the
algorithm processing efficiency, processing accuracy, and intelligence. input image (Fig. 2(d)). However, it does not mean that the network
It has been used in machine vision, speech recognition and natural expansion will necessarily achieve better results in terms of accuracy
language processing with a good performance at present. However, as or speed.
2
M. Zhang et al. Measurement 175 (2021) 109130
The EfficientNet network [48] explores whether there is a balance convolutional network could also be expressed as:
between the depth, width, or resolution of the input image to maximize
𝑁 = ⊙ 𝐹𝑖𝐿𝑖 (𝑋[𝐻𝑖 ,𝑊𝑖 ,𝐶𝑖 ] )
the performance of the neural network with limited resources, and for 𝑖=1...𝑠
the first time proposes and verifies the effect of the integrated scaling 𝐿 [
where 𝐹𝑖 𝑖 denotes layer 𝐹𝑖 is repeated 𝐿𝑖 times in stage i, 𝐻𝑖 , 𝑊𝑖 ,
of the model (Fig. 2(e)) on the effect of the neural network. ]
𝐶𝑖 denotes the shape of input tensor X of layer i.
This problem is similar to the multi-objective optimization problem
Scaling of the model only changes the depth, width, and resolution
in mathematics, where we try to improve the computational accu-
of the network, not the predefined 𝐹𝑖 in the Baseline.
racy as much as possible by appropriately adjusting the depth, width,
Therefore, the mathematical model of objective optimization can be
and resolution of the input image of the network, given the limited expressed as:
computational resources.
A convolutional network composed of k convolutional layers is 𝑇 𝑎𝑟𝑔𝑒𝑡 ∶ max 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦(𝑁(𝑑, 𝑤, 𝑟))
𝑑,𝑤,𝑟
represented as N, and each convolutional layer can be represented as:
̂
𝑌𝑖 = 𝐹𝑖 (𝑋𝑖 ). Where 𝑌𝑖 is the output tensor, 𝐹𝑖 is the operator, 𝑋𝑖 is the
[ ] 𝑠.𝑡.𝑁(𝑑, 𝑤, 𝑟) = ⊙ 𝐹̂𝑖𝑑⋅𝐿𝑖 (𝑋<𝑟⋅𝐻̂ ,𝑟⋅𝑊̂ ,𝑟⋅𝐶̂ > )
𝑖=1...𝑠 𝑖 𝑖 𝑖
input tensor with tensor shape 𝐻𝑖 , 𝑊𝑖 , 𝐶𝑖 , where 𝐻𝑖 and 𝑊𝑖 are spatial
dimension and 𝐶𝑖 is the channel dimension. Memory(N) ≤ target_memory
The structure of the entire convolutional network can be expressed Flops(N) ≤ target_f lops
as:
where 𝑤, 𝑑, 𝑟 are coefficients for scaling network width, depth, and res-
𝑁 = 𝐹𝑘 ⊙ ... ⊙ 𝐹2 ⊙ 𝐹1 (𝑋1 ) = ⊙ 𝐹𝑗 (𝑋1 ) olution; 𝐹̂𝑖 , 𝐿̂ 𝑖 , 𝐻̂ 𝑖 , 𝑊̂ 𝑖 , 𝐶̂𝑖 are predefined parameters in baseline network.
𝑗=1,2,…,𝑘
The results shown in Fig. 3 can be obtained through a large number
In practice, ConvNet layers are often partitioned into multiple stages of repeated experiments (respectively changing the depth, width, and
and all layers in each stage share the same architecture, then the image resolution of the network, and comparing the accuracy of the
3
M. Zhang et al. Measurement 175 (2021) 109130
4
M. Zhang et al. Measurement 175 (2021) 109130
5
M. Zhang et al. Measurement 175 (2021) 109130
Fig. 6. The test results comparison between improved-Yolov3 and the original-Yolov3: Image (a)–(d) belongs to the result of original Yolov3, (e)–(h) belongs to the improved-Yolov3.
89.12%, while the average test accuracy of original Yolov3 with Dark- the further increase of FLOPS will inevitably slow down the detection
net53 as the baseline is 88.12%, which shows the improved algorithm speed.
has a certain accuracy improvement effect, the specific improvement is
mainly manifested in the detection of tears. But it is obvious that the 3.2. Hardware usage comparison of multiple algorithms
improved algorithm has greatly improved in the prediction speed, and
the FPS has been increased from 28 to 42. In addition, the hardware usage of different algorithms was also dis-
As the above explanation of the structure of EfficientNet, the Effi- cussed in the context of this paper, from the aspects of CPU, GPU(cuda),
cientNet exists 8 structures, B0–B7. Among them, EfficientNet B1–B7 RAM, etc. The specific comparison results are shown in Fig. 5. The
comparison takes the hardware cost into consideration, which has
were obtained on the basis of EfficientNet-B0 by increasing the coeffi-
guiding significance for the field staff to choose a suitable algorithm.
cient (𝜑). With the increase of (𝜑), the depth and width of network as
But there is no doubt that the high-quality hardware resources will only
well as the resolution of the input image are expanding. In this paper,
increase the detection speed, not the detection accuracy.
we constraint 𝛼 ⋅ 𝛽 2 ⋅ 𝛾 2 ≈ 2 such that for any new (𝜑), the total FLOPS
It can be seen from Fig. 5 that the GPU and RAM usage of Improved-
will approximately increase by 2𝜑 .
Yolov3 (The backbone feature extraction network of here Improved-
EfficientNet B0–B7 were all tried as the backbone of Improved- Yolov3 is EfficientNet-B0) is relatively less than the original Yolov3,
Yolov3, and the test performance was shown in Table 3. The algorithm and the CPU usage of the two is basically the same. Relatively, Faster
achieves the best speed-accuracy trade-off on the conveyor belt damage R-CNN not only takes up a lot of GPU, but also CPU, which is caused by
dataset, with 97.26% AP at 26.28 FPS, 89.12% AP at 41.91 FPS. its huge FLOPS. Thus we can infer that with the further expansion of
Due to the limitations of the experimental environment and rapid the model, the improved algorithm will take up more GPU and CPU
increase of the FLOPS, the problem of graphics memory overflow occurs resources, even exceeding Faster R-CNN. Besides, it is more clearly
when trying B5. But according to the trend shown in Table 3, as the that the proposed improved-Yolov3 algorithm based on the EfficientNet
increase of coefficient (𝜑), the model will further expand, the prediction feature extraction network has a higher timeliness while ensuring the
accuracy of the neural network will be further improved. However, measurement accuracy.
6
M. Zhang et al. Measurement 175 (2021) 109130
Table 3
The performance of different backbone feature extraction networks on the dataset.
Model-Name mAP/% FPS Wear Damage Breakdown Tear
EfficientNet-B0-Yolov3 89.12 41.91 100 81 82 94
EfficientNet-B1-Yolov3 90.57 36 100 83 85 94
EfficientNet-B2-Yolov3 91.38 33.53 100 85 85 95
EfficientNet-B3-Yolov3 93.16 30.97 100 89 89 95
EfficientNet-B4-Yolov3 97.26 26.28 100 99 93 96
3.3. Performance comparison between Yolov3 and improved Yolov3 digest new data and make accurate predictions. A model’s ability to
generalize is central to the success of a model.
The performance of improved-Yolov3 and the original Yolov3 on In order to verify the generalization ability of the proposed model,
the dataset is compared in Fig. 6. It can be seen from the figure we conducted field tests in Libo Heavy Industry Technology Co., Ltd.
that both algorithms can better realize the identification and detection Due to the high cost of the conveyor belt, we chose to paste a damaged
of various scheduled conveyor belt damage defects. Different from belt on the normally running conveyor belt as the experimental test
existing image processing-based methods, deep learning-based target
material, as shown in Fig. 7. The camera is installed directly above
detection methods allow for direct target localization and classification.
the transmission roller of the machine head. The camera frame rate
The position information of the target object is shown in the rectangular
is 25 frames and the maximum resolution is 700 × 480: the width of
frame in the figure. The numbers on the rectangular frame are the
conveyor belt is 1.6 m, and the rated operating speed is 2.5 m/s. And
category and probability of the category of the target object. At the
same time, the improved algorithm is slightly higher in accuracy than what we want to emphasize here is that the experimental platform is
the original algorithm by comparing Fig. 6(d) and Fig. 6(h), which is only used for testing. In fact, when using machine vision technology
basically consistent with the results shown in Table 1, and the improved to detect conveyor belt damage, the camera should be placed on the
algorithm has a higher recognition accuracy for tearing detection. lower branch and return section of the conveyor. At the same time, it
is necessary to comprehensively consider the effects of light and surface
3.4. Verification of model generalization ability residues on the test results.
As shown in Fig. 7(d), the algorithm accurately identifies the tearing
Generalization is a term used to describe a model’s ability to react of the conveyor belt with the accuracy 81% , and the real-time frame
to new data. That is, after being trained on a training set, a model can rate is 35.43 FPS, which meets the actual needs of the project.
7
M. Zhang et al. Measurement 175 (2021) 109130
[1] M. Zhang, H. Shi, Y. Yu, M. Zhou, A computer vision based conveyor de-
As one of the main directions of the current belt conveyor devel-
viation detection system, Appl. Sci. 10 (2020) 2402, https://doi.org/10.3390/
opment, the intelligent energization of belt conveyors is the only way app10072402.
for the construction of smart mines. As the main bearing and traction [2] L.I. Zhanli, C. Jiaying, L.I. Hongan, et al., Research on intelligent monitoring
component of the belt conveyor, the reliability of the operation state and warning method of belt conveyor, J. Graph. (2017).
[3] Y. Gao, T. Qiao, H. Zhang, et al., A contactless measuring speed system of
of the belt seriously affects the safe and efficient operation of the con-
belt conveyor based on optical flow techniques, IEEE Access PP (99) (2019)
veyor. Compared with the traditional detection methods, the proposed 1, https://doi.org/10.1109/ACCESS.2019.2937317.
deep-learning based detection method has the following improvements [4] Pytlik Andrzej, Klaudiusz, et al., Research on tightness loss of belt conveyor’s
and advantages: idlers and its impact on the temperature increase of the bearing assemblies, J.
Sustain. Min. (2016) https://doi.org/10.1016/j.jsm.2016.07.001.
(1) A conveyor belt damage dataset has been created based on [5] K. Nienhaus, M. Warcholik, Christoph Büschgens, Belt positioning and skewing
a redefinition and reclassification of the damage types of conveyor prevention in lignite mining using long-wavelength infrared cameras, in: Proceed-
belts. Four types of damage are included: surface wear, surface dam- ings of the 12th International Symposium Continuous Surface Mining - Aachen
age, breakdown and tear. The more detailed classification of damage 2014, Springer International Publishing, 2015.
[6] M. Yang, W. Zhou, T. Song, Audio-based fault diagnosis for belt conveyor rollers,
allows for more accurate monitoring of the conveyor belt’s operating Neurocomputing (2020) 397, https://doi.org/10.1016/j.neucom.2019.09.109.
conditions and prevents unnecessary operations. [7] S. Kato, T. Nagayama, D. Su, et al., Identification of a cross-sectional vibration
(2) A deep learning based conveyor belt damage detection method mode of a belt conveyor support structure using acoustic loading, J. Struct. Eng.
A (2018) https://doi.org/10.11532/structcivil.64A.354.
was designed and proposed in this paper. The detection process is more
[8] Grzegorz Peruń, Attempt to evaluate the technical condition of the rollers of the
direct, without the assistance of other equipment or complicated image belt conveyor by vibration measurements, Vibroeng. Proc. 3 (2014) 296–299.
processing operations. It enables the detection of four conveyor belt [9] R. Li, X. Yang, T. Liu, et al., Application of fiber optic sensors for vibration
damage states and can give confidence levels while classifying their and ignition monitoring of a belt conveyor system, in: Optical Fiber Sensors and
Communication, 2019.
damage type.
[10] F. Klimenda, J. Soukup, J. Sterba, Noise and vibration analysis of conveyor
(3) The Yolov3 algorithm is improved in this paper to compre- belt, Manuf. Technol. 19 (4) (2019) 604–608, https://doi.org/10.21062/ujep/
hensively considers the balance between the width, depth and input 341.2019/a/1213-2489/MT/19/4/604.
image resolution of the network during model scaling and its impact [11] Y. Pang, G. Lodewijks, Simulation-based knowledge acquisition for intelligent
belt conveyor monitoring, Part. Part. Syst. Charact. 24 (4–5) (2007) 360–364,
on prediction accuracy.
https://doi.org/10.1002/ppsc.200601117.
(4) While comparing the accuracy and speed of algorithm operation, [12] P.D.I.G. Lodewijks, D.I.J.A. Ottjes, D.I.Y. Pang, Intelligent belt conveyor mon-
the hardware usage of the algorithm during operation is also taken into itoring and control, Des. Eng. (2010) http://resolver.tudelft.nl/uuid:585579aa-
account to provide a reference for engineers to select an algorithm. 7406-4e57-af8e-66c7df9b83bd.
[13] S.G. Sakharwade, S. Nagpal, Analysis of transient belt stretch for horizontal and
The improved algorithm brings a double improvement in detection inclined belt conveyor system, Int. J. Math. Eng. Manage. Sci. 4 (5) (2019)
accuracy and speed. It has a test accuracy of up to 97% on the conveyor 1169–1179, https://doi.org/10.33889/IJMEMS.2019.4.5-092.
belt damage dataset, and the fastest test speed is 42 FPS. At the same [14] L.P. Alviari, M.F. Anggamawarti, Y. Sanjiwani, et al., Classification of impact
time, it can be seen from the test platform that the model we proposed damage on a rubber-textile conveyor belt: A review, Int. J. Mech. Eng. Technol.
Appl. 1 (1) (2020) 21–27.
has a high generalization ability. [15] V. Redko, V. Khandetskyy, P. Novak, et al., Method and apparatus for
In the next work, we will focus on the lightweight research on the eddy current-based quality inspection of dry electrode structure: U.S. Patent
model to better adapt to the needs of edge computing. At the same time, 7355395[P]. 2008-4-8.
[16] Tiezhu Qiao, Xiaoyu Lu, Laiqing Yan, Research on the signal feature extraction
we will further expand the conveyor belt damage data set to improve
method in steel-cord conveyor belt with metal magnetic memory testing, J.
the robustness and generalization ability of the neural network. Comput. Theoret. Nanoence 11 (1) (2012) 489–492.
[17] Y. Guan, J. Zhang, Y. Shang, et al., Embedded sensor of forecast conveyer belt
breaks, in: International Conference on Fuzzy Systems and Knowledge Discovery,
CRediT authorship contribution statement
IEEE Computer Society, 2008, https://doi.org/10.1109/FSKD.2008.643.
[18] Y. Pang, G. Lodewijks, A novel embedded conductive detection system for
Mengchao Zhang: Methodology, Writing - original draft, Software. intelligent conveyor belt monitoring, in: 2006 IEEE International Conference on
Service Operations and Logistics, and Informatics, Shanghai, 2006, pp. 803–808,
Hao Shi: Data curation, Writing - review & editing. Yuan Zhang:
https://doi.org/10.1109/SOLI.2006.328958.
Supervision. Yan Yu: Investigation, Supervision. Manshan Zhou: Con- [19] H.I. Bozma, H. Yalçın, Visual processing and classification of items on a moving
ceptualization, Writing - review & editing. conveyor: a selective perception approach, Robot. Comput.-Integr. Manuf. 18 (2)
(2002) 125–133, https://doi.org/10.1016/S0736-5845(01)00035-7.
[20] Michał Kontny, Machine vision methods for estimation of size distribution of
Declaration of competing interest aggregate transported on conveyor belts, Vibroeng. Proc. (2017) https://doi.org/
10.21595/vp.2017.19151.
[21] Jayson Tessier, Carl Duchesne, Gianni Bartolacci, A machine vision approach to
The authors declare that they have no known competing finan- on-line estimation of run-of-mine ore composition on conveyor belts, Min. Eng.
cial interests or personal relationships that could have appeared to (2007) https://doi.org/10.1016/j.mineng.2007.04.009.
influence the work reported in this paper. [22] Z. Zhang, Y. Liu, Q. Hu, et al., Multi-information online detection of coal quality
based on machine vision, Powder Technol. (2020) 374, https://doi.org/10.1016/
j.powtec.2020.07.040.
Acknowledgment [23] C. Aldrich, G.T. Jemwa, J.C. Van Dyk, et al., Online analysis of coal on a
conveyor belt by use of machine vision and kernel methods, Int. J. Coal Prep.
Util. 30 (6) (2010) 331–348, https://doi.org/10.1080/19392699.2010.517486.
The authors wish to acknowledge Yibing Shang at University of [24] M. Kistner, G.T. Jemwa, C. Aldrich, Monitoring of mineral processing systems
Exeter, United Kingdom and Zhengxu Zhang, postgraduate of Shandong by using textural image analysis, Min. Eng. 52 (2013) 169–177, https://doi.org/
University of Science and Technology, China, for their help in collecting 10.1016/j.mineng.2013.05.022.
[25] T. Andersson, M.J. Thurley, J.E. Carlson, A machine vision system for estimation
the original data and labeling.
of size distributions by weight of limestone particles, Min. Eng. 25 (1) (2012)
38–46, https://doi.org/10.1016/j.mineng.2011.10.001.
Funding [26] M.H. Zhao, S. Ma, D.X. Zhao, Image processing based on gray information
in sorting system of coal gangue, in: 2018 10th International Conference on
Intelligent Human-Machine Systems and Cybernetics (IHMSC), 2018, https://doi.
This research received no external funding. org/10.1109/IHMSC.2018.10125.
8
M. Zhang et al. Measurement 175 (2021) 109130
[27] D. Dou, W. Wu, J. Yang, et al., Classification of coal and gangue under multiple [38] X.L. Hao, H. Liang, A multi-class support vector machine real-time detection sys-
surface conditions via machine vision and relief-SVM, Powder Technol. 356 tem for surface damage of conveyor belts based on visual saliency, Measurement
(2019) 1024–1028, https://doi.org/10.1016/j.powtec.2019.09.007. 146 (2019) 125–132, https://doi.org/10.1016/j.measurement.2019.06.025.
[28] Y.D. Zhao, M.F. Sun, Image processing and recognition system based on DaVinci [39] Y. Wang, Y. Wang, L. Dang, Video detection of foreign objects on the surface
technology for coal and gangue, Appl. Mech. Mater. 130134 (2012) 2107–2110, of belt conveyor underground coal mine based on improved SSD, J. Amb. Intell.
https://doi.org/10.4028/www.scientific.net/AMM.130-134.2107. Hum. Comput. (2020) https://doi.org/10.1007/s12652-020-02495-w.
[29] Christopher C. Fromme, David J. Stager, Thomas E. Pilarski, et al., Conveyor [40] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for
belt inspection system and method: US, 2006. accurate object detection and semantic segmentation, in: 2014 IEEE Conference
[30] Y. Yang, C. Miao, X. Li, et al., On-line conveyor belts inspection based on on Computer Vision and Pattern Recognition (CVPR), Columbus, OH, USA, 2014,
machine vision, Optik 125 (19) (2014) 5803–5807, https://doi.org/10.1016/j. pp. 580–587, https://doi.org/10.1109/CVPR.2014.81.
ijleo.2014.07.070. [41] R. Girshick, Fast R-CNN, in: 2015 IEEE International Conference on Computer
[31] Z. Aichun, H. Gang, W. Yongxing, The research on the detection method of Vision (ICCV), Santiago, 2015, pp. 1440–1448, https://doi.org/10.1109/ICCV.
belt deviation by video in coal mine, in: 2011 International Conference on 2015.169.
Mechatronic Science, Electric Engineering and Computer (MEC), IEEE, 2011, pp. [42] S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time object
430–433, https://doi.org/10.1109/MEC.2011.6025493. detection with region proposal networks, IEEE Trans. Pattern Anal. Mach. Intell.
[32] W. Liu, T. Qiao, G. Yan, et al., Research on visible light and infrared vision 39 (6) (2017) 1137–1149, https://doi.org/10.1109/TPAMI.2016.2577031.
real-time detection system for conveyor belt longitudinal tear, Iet ence Meas. [43] W. Liu, D. Anguelov, D. Erhan, et al., Ssd: Single shot multibox detector, in:
Technol. 10 (6) (2016) 577–584, https://doi.org/10.1049/iet-smt.2015.0297. European Conference on Computer Vision, Springer, Cham, 2016, pp. 21–37.
[33] L. Xianguo, S. Lifang, M. Zixu, et al., Laser-based on-line machine vision [44] J. Redmon, S. Divvala, R. Girshick, et al., You only look once: Unified, real-time
detection for longitudinal rip of conveyor belt, Optik 168 (2018) 360–369, object detection, in: Proceedings of the IEEE Conference on Computer Vision and
https://doi.org/10.1016/j.ijleo.2018.04.053. Pattern Recognition, 2016, pp. 779–788.
[34] B. Yu, T. Qiao, H. Zhang, et al., Dual band infrared detection method based on [45] J. Redmon, A. Farhadi, Yolov3: An incremental improvement, 2018, arXiv
mid-infrared and long infrared vision for conveyor belts longitudinal tear, Mea- preprint arXiv:1804.02767.
surement 120 (2018) 140–149, https://doi.org/10.1016/j.measurement.2018.02. [46] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
029. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las
[35] Y. Yang, C. Hou, T. Qiao, et al., Longitudinal tear early-warning method for Vegas, NV, USA, 2016, pp. 770–778, https://doi.org/10.1109/CVPR.2016.90.
conveyor belt based on infrared vision, Measurement 147 (2019) 106817, https: [47] T.Y. Lin, P. Dollár, R. Girshick, et al., Feature pyramid networks for object
//doi.org/10.1016/j.measurement.2019.07.045. detection, in: Proceedings of the IEEE Conference on Computer Vision and
[36] R. Yang, T. Qiao, Y. Pang, et al., Infrared spectrum analysis method for detection Pattern Recognition, 2017, pp. 2117–2125.
and early warning of longitudinal tear of mine conveyor belt, Measurement [48] M. Tan, Q.V. Le, Efficientnet: Rethinking model scaling for convolutional neural
(2020) 107856, https://doi.org/10.1016/j.measurement.2020.107856. networks, 2019, arXiv preprint arXiv:1905.11946.
[37] C. Hou, T. Qiao, H. Zhang, et al., Multispectral visual detection method for [49] M. Everingham, L.V. Gool, C.K.I. Williams, et al., The pascal Visual Object Classes
conveyor belt longitudinal tear, Measurement 143 (2019) 246–257, https://doi. (VOC) challenge, Int. J. Comput. Vis. 88 (2) (2010) 303–338.
org/10.1016/j.measurement.2019.05.010.