Optimization automating monitoring based on classification for rolling bearing

1. Introduction

Nowadays, both efficiency and the sustainability of the production processes of high-tech manufacturing products depend on the implementation of advanced manufacturing strategies in the production processes. In particular, the development of advanced solutions for zero-defect application metrology is considered a robust technology capable of delivering a wide competitive advantage to manufacturing industries, which is the motor of the 4th industrial revolution (Industry 4.0). Rotary machines are presently prevalent in applications involving power transmission systems. As demand increases, the importance of timely repair of defects increases due to the importance of reducing losses due to interruption of work, so it was necessary to use several methods of detecting defects, including mention of vibration analysis [1]. The vibration analysis of these types of equipment is an accurate method because they give remarkable results in maintenance programs. At present, vibration analysis plays a role in manufactures to maintain and increase the condition of available machines. The classification has used recently to detect defects in the rolling bearing [2] based on the extracted vibratory signals. The OPTICS (Ordering points to identify the clustering structure) [3] is an efficient method utilized to classify data, has a weakness in the determination of parameters epsilon the radius of the cluster, and Minpts the minimum number of data points needed for clustering. Both have a considerable effect on the clustering results because the parameter values affect the detection of the defect. Besides, to the parameters epsilon and Minpts, the Euclidean distance shows the limitation of the use of large amounts of data. To solve these problems, we suggested the optimization and automating of the classification. The procedure used in this methodology leads to reach the parameter automatically. The automating consist of the calculation of parameters Minpts and Epsilon. The optimization is the applying of the Relief method and the distance Manhattan. Then the Relief method for calculating the correlation between features and extracting the most important, to use in the detection of the defect. It also replaced the Euclidean distance by the Manhattan distance [4]. The dimension reduction of features by method t-SNE [5] to optimize and thus make reliable real-time detection and the tracking of rolling defects. In particular, the methodology applied to numerical and experimental data in the outer race of the rolling bearing.

2. Optics method

OPTICS (Ordering Points to Identify Clustering Structure) is an algorithm founded by [6], it has been used in many fields of biology, astronomy, topology, and recently in the mechanical area especially for the detection of defects in the rolling bearings in rotary machines by [7].

OPTICS is a density-based cluster algorithm that can classify the objects, by their density, the dense regions separated from the noises which are the less dense. The two concepts of OPTICS are:

– Core distance. The database $D$ has a point object $p$, and distance $\epsilon$, the $N\epsilon \left(p\right)$ is distance-neighbor of $p$, MinPts is a numeric number, and $MinPt{s}_{distance}$($p$) is the distance from $p$ to its MinPts neighbor. The core-distance of $p$ is defined in Eq. (1):

– Reachability distance. The database $D$ has two points object $p$, and $o$, $N\epsilon \left(o\right)$ be the $\epsilon$ distance-neighbor of $o$, MinPts is a numeric number. The reachability-distance of $p$ concerning o defined in Eq. (2):

3. Proposed method: AOM-OPTICS

4. Experimental validation

The experimental validation was carried out at the University of Reims Champagne-Ardenne in France. The testing bench in Fig. 1(a). has an electric motor which delivers a maximum power of 10 kW across a shaft held with two bearings: a healthy 6206 ball bearing and a N.206.E. G15 rolling bearing, which is being damaged. The rotation is regulated by a speed controller. A steel cable actuation is used to load the shaft radially. The defects are generated by an engraving pen on the outer race of the roller bearing. The defects are slightly varying and their sizes are measured using PlastiformTM paste. The defect sizes are shown in Table 1. The collected database for each defective measurement size, using 5 different loads varying from 100 to 220 daN, with a 30 daN step, and with 5 different rotational speeds varying from 1405 to 1560 rpm and with a step of 50 rpm. A Piezo-electric sensor positioned radially on the bearing is taken as the best measuring point. We have used the OROS OR34 acquisition system. The database that was created (9×10×8), composed of 9 states, one healthy and eight have different defect sizes, the single state has five different speeds as well as five different loads, which results to give us a set of 10 signals, the signal split into eight parts. The database has 720 values. The same methodology with the simulated data used here in the experimental. The Relief method and $t$-SNE used for the reduction of dimension, then the AOM-OPTICS methodology applied.

Fig. 1a) Test bench, b) defected N.206.E. G15 rolling bearing, c) N.206.E. G15 rolling bearing

a)

b)

c)

The healthy state in rolling bearing, which is represented by one cluster with the dispersed points surround, shown in Fig. 2(a). The detection of defect lead to appear the second cluster that correspond to defect number one observed in the Fig. 2(b). The growth in the defect, with eighteen amplitudes, corresponding to defect number eight, can show in the Fig. 2(c).

Fig. 2a) #10 Healthy, b) #11 faulty, c) #18 faulty

a)

b)

c)

4.1. Follow-up

Follow-up is the final step in monitoring the development of defects over time. The second cluster evolution is obtained from the characteristic parameters (Calinski index, Davies-Bouldin index). The experimental results of follow-up, for the Calinski index, the curve grows with the number of the signal. Otherwise, the Davies Bouldin increase with the signals, that means the defect growing with the signals value the results shown in Fig. 3(a-b).

Fig. 3a) Calinski-index, b) Davies-Bouldin index

a)

b)

5. Conclusions

Based on a dynamic classification by OAM-OPTICS methodology, an unsupervised classification method developed to monitor bearings by the vibration signals has been proposed. The proposed methodology validated through the simulated and experimental signals. This automatic implantation and optimized methodology used in the classification, allows us to detect the defects, with a high accuracy equal to 100 %. The performance of the method is very efficient; it can detect small variations. The influence of the load and speed cannot lead to discovering the defect, which improved numerically and experimentally, by the variety of load and speed. The characteristic parameters facilitate observing the evolution of errors with time. The follow-up makes it easier to monitor the influence of the defect growth over time.