Probabilistic Reasoning for Improving the Predictive Maintenance of Vital Electrical Machine : Case Study

Nowadays, new information technologies produce new methodological approaches attempting to extract not just valid and reliable information, but more generally a particular technical and professional expertise to support the decision making. A Bayesian network was developed for fault assessment of an electrical motor. By inference, this model made it possible to calculate the probability of rotor fault of the induction motor, while de ning the weakest branch in the structure of the Bayesian network that leads to failure by determining the probabilities of intermediate events. The most likely faults were then de ned and the information system consolidated, as well as the decision-making process. The article ends with an application that shows the methodology developed and gives some results illustrated by gures.


Introduction
Induction motors are widely used (it is estimated that 80 % of the motors used in the various industrial domains in the world are of the induction motor type) because their cost is lower than that of other machines.Moreover these motors are robust.The power range covered by induction motors is very wide, ranging from 5 watts, for single-phase induction motors, to 36.8 MW, for squirrel-cage motors.They are of variable criticality.In practice, some motors are classied as vital, others as important, and others as unimportant.
Preventive maintenance is commonly associated with high power motors.Nowadays, and given the importance of these machines and the economic losses that their shutdown generates, a conceptual evolution is necessary.We need to move from preventive maintenance to predictive maintenance.Predictive maintenance ensures that the machine will be kept in good condition, extends its life cycle and, above all, improves its availability.The implementation of a predictive maintenance strategy requires a rigorous analysis of the motors, the modes of degradation, the electrical and mechanical parameters, the means c 2017 Journal of Advanced Engineering and Computation (JAEC) to be used, and the objectives in terms of availability.
In order to ensure a good mastery over time of the predictive maintenance of induction motors, the maintenance plan must contain all the necessary information.The exploitation of the history les and the experience gained through the exploitation of the information held by the experts give a good contribution.The masterpiece of the approach set out in this article is the inventory of failures, in logic of causality, while linking causes to eects.The next phase is the denition of the priorities in terms of intervention.In this case, the fundamental objective is to establish the planning of the interventions Bayesian networks are also articial intelligence tools little used for fault prediction in electrical machines [5].A new fault diagnosis method for rotating machinery based on adaptive statistic test lter and Diagnostic Bayesian Network is presented by [6].A three-layer diagnostic Bayesian network is developed to identify condition of rotation machinery based on the Bayesian theory.In another contribution, a Bayesian network of motor fault diagnosis model based on stator current signal and its Hilbert marginal spectrum feature was introduced [7].
The results found have shown that this fault diagnosis model is eective and accurate.
In this paper, a Bayesian model will be developed to build an information system that can predict failures in induction motors.The Bayesian formalism presented in this work gives a strong contribution to the prioritization of maintenance actions and the planning of preventive maintenance work.

Bayesian network tool
Basically, a Bayesian network represents a graphic transposition of knowledge.The two main elements constituting a Bayesian network are: its structure and its parameters.The structure of this type of network is simple: a graph in which the nodes, representing random variables, and the edge (so the graph is oriented), connecting these nodes, are connected to conditional probabilities.Note that the graph is acyclic, i.e. it does not contain a loop.Edges represent relationships between variables that are either deterministic or probabilistic.Thus, the observation of one or more causes does not systematically result in the eect or eects that depend on it, but only modies the probability of observing them.
The particular interest of Bayesian networks is to take into account simultaneously the a priori knowledge of experts (on the graph) and the experience contained in the data.The graphical representation of a Bayesian network is explicit, intuitive and comprehensible by a nonspecialist, which facilitates both the validation of the model, its possible evolutions and especially its use.Depending on the type of application, the practical use of a Bayesian network can be considered in the same way as other models: neural networks, expert systems, decision trees, data analysis models, tree failures, logic models, etc. a BN (Fig. 1) is dened by a directed acyclic graph (DAG) as: Where C(V i ) is the set of parents (or causes) of v i in the graph (Fig. 1).
In recent last years, Bayesian networks have been widely used in the eld of diagnosis and prediction [8,9].The central element of any Bayesian approach is information.Complete or incomplete, these tools nd solutions for reasoning under uncertainty.They are also powerful decision support tools widely used in the modeling of complex and dynamic systems [10].

Bayesian network development
A Bayesian network was constructed from peer-reviewed technical literature, census data, and fault statistics reports.When required probability data were unavailable or the sample size was too small, an expert in maintenance engineering provided subjective estimates of the probabilities.

Construction of the model structure
In any case, in fault analysis the eect is a feared event, whose causes we attempt to iden-

Application and discussion
In order to improve the availability of the  The inference rules are given by the conditional probability table.Some readings are as follows: • If the cause 132 takes the state True then the fault 14 exists.
• If the cause 141 takes the false state and the cause 132 takes the false state then the fault 14 exists if and only if the cause 142 takes the state True.
• If the causes 132, 141, and 142 take the state false, there is no fault in the rotor.
In the Bayesian formulations, as soon as the number of variables becomes important, computations become complicated.To relax these constraints a program was written in Matlab environment.After inference in the network of Fig. 2, the a posteriori probabilities for each fault can be found, as well as for the rotor fault (Tab.4).
From the results given in Tab. 4, it is possible to make decisions on the corrective actions that should be carried out in a certain environment.
Also these prediction results give the possibility to organize the actions in a specic order of pri- In the Bayesian networks, the predicting models are universal to the all power range of mo- "This is an Open Access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0)." which allows representing, in a global and synthetic way, the activity of the predictive maintenance on these motors.With the required high productivity levels at industrial plants, any urgent corrective maintenance action or unscheduled shutdown can cause signicant economic losses.In industries like petrochemical, some techniques, as vibration analysis, are able to detect the fault's early onset which could avoid more serious problems.In this sense, there are many studies focused on early fault detection.In complementarities, in recent years several articial intelligence techniques have been developed and applied in the monitoring processes of faults, among them, the Fuzzy Logic, Articial Neural Networks, and Support Vector Machines [1, 2, 3].An ecient supervised Articial Neural Network (ANN) learning technique that is able to identify fault type when situation of diagnosis is uncertain was presented by [4].Signicant features are taken out from the electric current which are based on the dierent frequency points and associated amplitude values with fault type.

Tab. 1 :
tify.Tab. 1 presents basic and intermediate events causing faults in the rotor of induction motor.Also in the same Table a codication is adopted, each code representing a variable in the BN developed in the rest of this article.The causal relationship determines the structure of the network.A research in the literature made it possible to dene the parents of each node in a global way then a questionnaire addressed to the operators made it possible to give details on the causes and the aects and the relations between them.Census of causes causing rotor faults.rotor, the faults are limited to the breaking of rods or the breakage of short-circuit rings.Thus, for the purpose of a detailed presentation faults are classied.The most intuitive graphical representation of the inuence of an event, a fact, or a variable on another is to represent causality by linking the cause to the eect by an oriented edge.The structure of the BNs, developed in this paper, focuses on the causality principle which is the basis of Bayesian reasoning.Identied faults are thus illustrated in the BN given in Fig. 2. A global analysis of the fault causes organized on a single induction motor is carried out.After grouping relevant and meaningful information by categories (cause and eect), and presenting cause-and-eect links (after conrmation of the fault) with edges.Such a network allows a complete and visual operation, relatively easy, of each induction motor.
induction motor, we might have to establish a motor failure model that represents the rotor's faults as an eect function of some input variables (causes).Having done that, prediction of a rotor fault could be accomplished by calculating a posteriori probabilities.For example, when considering the availability of an induction motor, as we do in this paper, the controllable variables can be the electrical measurements, vibration, and the number of spare parts of dierent types.To validate the proposed methodology an important induction motor used in SONATRACH / SKIKDA / GL1K / LNG plant is assigned as the study motor with a power of 288 kW.The installations of the motor were started up in 2000.The information for dening the probability of fault will be evaluated only on the basis of experience feedback and the history le of the mac 2017 Journal of Advanced Engineering and Computation (JAEC) VOLUME: 2 | ISSUE: 1 | 2018 | March chine.While the structure of the network will be dened by an expert, in view of its complexity of denition, in this case the role of the operator is primordial.Referring to the history le of the studied induction motor, the results of the conducted diagnostics, it's possible to dene the a priori probabilities for each cause (Tab.3).Information concerning Thermal fatigue fault does not exist in the historical data, an estimate of an expert can regularize the a priori probability.

5 .
ority.It is to be noticed that the probability of having a fault in the rotor (73.70 %) is signicant, and the main faults in decreasing order are respectively: Bearing fault and rotor misalignment, mechanical imbalance, bearing misalignment, magnetic circuit fault, breaking bars.The Bayesian network of Fig. 2 is causal network.From a practical point of view, to connect causality with Bayes' theorem, the joint law of a cause C(V i) and its consequence V i is represented by the factorization P (C(V i).P (V i|C(V i)).To show that the graph of Fig.2satises the Markov condition, the calculation of the probability of node 13 with two c 2017 Journal of Advanced Engineering and Computation (JAEC) VOLUME: 2 | ISSUE: 1 | 2018 | March states (true and false) is taken as an example: P (13 = T ) = P (13 = T /122 = T, 132 = T ) × P (122 = T ) × P (132 = T ) + P (13 = T /122 = T, 132 = F ) × P (122 = T ) × P (132 = F ) + P (13 = T /122 = F, 132 = T ) × P (122 = F ) × P (132 = T ) + P (13 = T /122 = F, 132 = F ) × P (122 = F ) × P (132 = F ) = (1 × 0.042 × 0.095) + (1 × 0.042 × 0.905) + (1 × 0.958 × 0.095) + (0 × 0.958 × 0.905) = 0.00399 + 0.03801 + 0.09101 = 0.13301 In the same way it is possible to check the node 11, 12, 14, 15, 16 and nally the node 1. Tab. 4: A Posteriori Probabilities.current signature analysis are widely used for fault diagnosis in electric motors.In the case where several faults appear simultaneously, the fault diagnosis becomes dicult with the vibration analysis technique.A Rotor misalignment fault and Magnetic circuit fault can occur at the same moment and with neighboring frequencies (100 Hz for 3000 rot/min).Also, a poorly calibrated sensor may give false alarm.The current signature analysis is sensitive to load variations and supply oscillations.However, the reliability of these precursors is not sucient to identify the pos-sible fault.The same go with temperature, partial discharge, gas analysis, surge test, magnetic ux, the air-gap torque, and power.All these methods diagnostic a single fault.In this situation the exploitation of the information held by the operators remains the only way to make decisions on the faults prediction.The Bayesian network developed in this study gives the probability of occurrence of fault on the basis of the behavior of the machine and with very precise values which makes it possible to take decisions in all certainty.In the network of Fig. 2 some information modies the beliefs that we have on the machine.In this Bayesian network the notion of probability used is a subjective notion of belief.Results of Tab. 4 measure the belief that a cause attributes to the occurrence of a given fault.Operators operate a vital induction motor even with alarming vibration levels.On the other hand face a strong probability that the rotor breaks down (73%) the operator does not take the risk.One reason for attributing a stronger belief to one fault rather than another is given by the a posteriori probabilities.These results give the possibility to anticipate the faults; this also saves time earlier wasted in responding to false alarms, and consequently makes reliable the prediction.It should be mentioned that the structure of the model remains the same for all induction electrical machines.But the network parameters (a priori probabilities) change from one machine to another.The basis for prioritizing the maintenance actions of motor for each fault is the Bayesian network of Fig. 2. Tab. 4 shows the a posteriori probabilities of the basic failures in the rotor of motor which were obtained by inference in the Bayesian network.The impact of the bearing fault of all faults is jointly seen in the results for the predictive maintenance plan of the studied motor.Considering the values given by Tab. 4, the prioritizing failure within the framework of the predictive maintenance is: Bearing fault and rotor misalignment, mechanical imbalance, bearing misalignment, all faults are mechanical.It is seen that in the nal scenario the failures that have a higher priority should be corrected rst.Also, the prioritization depends on the motor and the operating conditions, as shown in Fig. 2. In this study, fault assessment is stated in terms of probabilities.Also the proposed Bayesian network is simple and intuitive.By comparing it by other tools (expert system, neural network) it can be used by a non-specialist.Now, to practically proof the eectiveness of the proposed work, it is necessary to eliminate the causes that have high priority and to re-evaluate again the availability rate of the motor.The proposed tool is a systematic tool for decision making in the eld of diagnostic and prediction of faults, formalizing the interpretation of information and measurement results, and in particular the recurrent logical links that can be made, has the signicant advantage of facilitating the work of the expert, or even of replacing it in simple or classic cases of diagnosis / prediction.Indeed, discharged from a systematic work that he realizes each time, the expert can focus then on higher-level work.Advantages of the bayesian network tool over other articial intelligence techniques In the Bayesian formulation used in this paper the belief calculation amount for each fault candidate is small regarding to others techniques such as: articial neural networks (ANN), Fuzzy logic, Expert systems.The second common advantage of Bayesian network regarding the articial intelligence techniques listed above is that the graphic representation of a Bayesian network is simple, intuitive and understandable by a nonspecialist. it should be noted that the articial intelligence tools listed above are more suitable and more ecient for online fault detection in induction motor compared to the Bayesian networks, but do not give a justied and precise information on the forecast occurrence of the fault.Two types of ANNs have given a strong contribution in diagnostic and prediction of faults in induction motors: supervised and unsuper-vised ANNs.Unsupervised ANNs do not need to train before, and their training is carried out online continuously.However, implementation of ANN based method needs a spacious practical data of the motor in dierent operating conditions.Also, change of the motor working conditions in dierent environments may aect the fault detection, and the trained ANN for each motor is unique [11].Also expert knowledge cannot be previously represented in the ANN.By contrast, in the Bayesian network the structure represents knowledge and the relationships between the variables and the parameters of the network can be updated easily.The learning ability is moderate regarding Bayesian network.The ANN not allow a very large work amount in the fault prediction, their performance is small, while the Bayesian networks are very ecient on this.The calculations of the probability distribution associated with a Bayesian network are falls within the inference.For the methods based on fuzzy logic technique the accuracy of predicting results is bad in deep inference regarding with Bayesian networks tools which is excellent in error tolerance and in deep inference.Also, if the domain experience representation is good for the induction motor fault prediction based on fuzzy logic, it is moderate relatively to Bayesian network in the learning ability.The wide range of size and power of induction motors that can be covered by prediction methods based on Bayesian networks and those based on fuzzy logic remains a common advantage.
TACHI received his State Engineer degree in Automatic Control from the Department of Electrical Engineering, University of Badji Mokhtar, Annaba, Algeria, and the MSc. in Control Systems Engineering from the Department of Automatic Control and Systems Engineering, University of Sheeld, Sheeld, U.K., in 1990 and 2004, respectively.He then moved to Algerian Petroleum Institute, School of Skikda, Algeria, where he is currently Lecturer Researcher in Control Engineering.
Fouad c 2017 Journal of Advanced Engineering and Computation (JAEC) 17