for predicting remaining useful life of rolling bearing : Case of OCP company. 1. University Mohammed VI Polytechnic, BenGuerir 43150, Morocco. Abdelhak ...
Using M ulti-L ayer Perceptron neur al network for pr edicting r emaining useful life of rolling bear ing : Case of OCP company Abdelhak ELIDRISSI 1,2,* , Saad BENJELLOUN1, Janah SAADI 1 , Imad BENTTALEB3, Noureddine AJIM 3 1 2
University Mohammed VI Polytechnic, BenGuerir 43150, Morocco
Ecole Mohammadia d?Ingénieurs, Avenue Ibn sina, Rabat 10000, Morocco 3
OCP Group, 2-4, Rue Al Abtal, Hay Erraha, 20200, Casablanca, Morocco
Abst r act In this work, we are focused on the concept of prognosis [1][2] which represents a key process of predictive maintenance. The scientific literature has several different classifications of prognostic approaches [3] [4], namely the prognosis based on models, the prognosis guided by data and the prognosis based on experience. In this work we will focus on the prognosis guided by data mainly with machine learning methods (multi-layer perceptron neural network and multiple linear regression) to predict the remaining useful life of a strategic equipment (Grinding mill) of the OCP?s company mine. An exploration of prediction capabilities was conducted based only on operating conditions data (tonnage, phosphate layer, phosphate quality, hours of operation) to predict the remaining life of critical grinding mill components such as rolling bearing, contrarily to classical approaches based primarily on vibration and temperature monitoring. The two proposed models show a good fit with historical testing data, and can be used for predictive maintenance planning.
Descr ipt ion of t h e st u died syst em Our studied system is the grinding mill of the Benguerir mine, it's a critical system in the mine because his eventual stoppage conditioned the totality of the Rock removal site of Benguerir mine. The representation of the rock removal site of BenGuerir is as follows : The grinding mill is composed of : 2 drive motors, 6 belts pulleys, 2 connecting rods and 4 bearings. The most critical component in the grinding mill is the rolling bearing because for changing 1 rolling bearing, we must spent 84 hours of stoppage.
M et h ods Data descr iption : The database used in this work contains five independent variables values which correspond to the operating conditions of the grinding mill: Poste, Phosphate Layer, Quality of phosphate, hours of operation (HO), tonnage, and one dependent variable: RUL. We first, add two additional data which consist on nonlinear transformation of the last data, namely using the cumulative working hours and processed tonnage, Cumulative HO, Cumulative Tonnage. The data was recorded during the period from 28/01/2014 to 26/09/2016, The 2014-2015 data is used as learning/test set and 2016 data for prediction. M ultiple linear r egr ession model : The general form of multiple linear regression model can be written as following : Y = a0 + a1X 1 + a2X 2 + ..... + apX p - Y is the dependent or explained variable. - X 1 , X 2, X 3 : are independent or explanatory variables measured. - a0, a1 , a2, a3 : are the model parameters. Ar tificial neur al network : An artificial neural network is a computational model whose original inspiration was a biological model, computing and mathematical representation of biological neuron in the artificial neural network is called formal neuron. We propose in this study to work with multi-layer perceptron (MLP) neural network (type of neural network ). The structure of our proposed neural network is a as follow :
Resu lt s Multiple Linear Regression using the R language give the following results: Historical data vs Model (Learning set)
RUL predictions for 2016
2
The R on the testing set is 0,786. By performing Multi-Layer Perceptron Neural network, with the architecture given above, the model output and predictions are as follows :
The R2 on the testing set is 0,998. By our proposed approach, we predict that the failure of rolling bearing of grinding mill may occur between : 01/01/2017 and 28/02/2017. The neural network prediction seems to be more in accordance with recent data gathered from OCP.
Neur al Network
M ulti-linear r egr ession
R-squared
99,87%
78,64%
Pr edicted date of failur e
Between: 01/01/2017 and 28/02/2017
01/06/2016
Con clu sion In this work we propose a new approach based on the exploitation of the historical operating conditions for the predictive maintenance of the OCP grinding mill. The model can be used to perform better scheduling of maintenance (pre-ordering of components) and hence to increase the availability of the grinding mill. For the perspectives, we propose to generalize our approach to various type of defects and other strategic machines of the OCP mine.
Ref er en ces [1] Mobley, R,K, (2002), An introduction predictive maintenance : Second Edition, Elsevier Science. [2] ISO 13381-1, Condition Monitoring and Diagnostics of Machines - Prognostics - Part 1 : General Guidlines : International Standards Organization 2004. [3] Jardine, A,K,S, Lin, D, Banjevic, D, (2006). A review on machinery diagnostics and prognostics implementing condition based maintenance. Mechanical Systems and signal Processing 20, 1483-1510. [4] Kandukuri, S, T, klausen, A, (2016). A review of diagnostics and prognostics of low speed machinery towards wind turbine farm-level health management . renewable and Sustainable Energy Reviews 53 697-708.