Computers in Industry 57 (2006) 552–568 www.elsevier.com/locate/compind SIMAP: Intelligent System for Predictive Maintenance Application to the health condition monitoring of a windturbine gearbox Mari Cruz Garcia a, Miguel A. Sanz-Bobi a,*, Javier del Pico b a Universidad Pontificia Comillas, Instituto de Investigación Tecnológica, IIT, Santa Cruz de Marcenado 26, 28015 Madrid, Spain b Molinos del Ebro S.A. Paseo de la Independencia 21, 38 50001 Zaragoza, Spain Accepted 9 February 2006 Available online 13 June 2006 Abstract SIMAP is the abbreviated name for the Intelligent System for Predictive Maintenance. It is a software application addressed to the diagnosis in real-time of industrial processes. It takes into account the information coming in real-time from different sensors and other information sources and tries to detect possible anomalies in the normal behaviour expected of the industrial components. The incipient detection of anomalies allows for an early diagnosis and the possibility to plan effective maintenance actions. Also, the continuous monitoring performed allows for an estimation in a qualitative form of the health condition of the components. SIMAP is a general tool oriented to the diagnosis and maintenance of industrial processes, however the first experience of its application has been at a windfarm. In this real case, SIMAP is able to optimize and to dynamically adapt a maintenance calendar for a monitored windturbine according to the real needs and operating life of it as well as other technical and economical criteria. In particular this paper presents the application of SIMAP to the health condition monitoring of a windturbine gearbox as an example of its capabilities and main features. # 2006 Elsevier B.V. All rights reserved. Keywords: Predictive maintenance; Maintenance effectiveness; Health condition; Diagnosis; Artificial intelligence 1. Introduction The use of wind is one of the most attractive new sources of energy at the present moment, as can be seen by the growing installation of windfarms all over the world. Windturbines are relatively young machines where the application of a correct maintenance strategy would be very important for the protection of their future life, productivity and profitability [1]. The current practice of maintenance applied to the existing aerogenerators is based on periodical or preventive maintenance actions recommended by their manufacturers. These are good and general guidelines for the maintenance of aerogenerators, however they do not focus on the specific characteristics of the real and local life of them such as: weather conditions at the location, stress by over-load, hours continuously working, etc. These factors determine the particular life and health of each aerogenerator and for this reason the maintenance applied has to also take them into account. In order to do this, a predictive maintenance plan is * Corresponding author. Tel.: +34 91 542 28 00x4240; fax: +34 91 542 31 76. E-mail address: masanz@upcomillas.es (M.A. Sanz-Bobi). 0166-3615/$ – see front matter # 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.compind.2006.02.011 the best option to guarantee the long life of the new investments in aerogenerators due to the maintenance actions which are applied according to the real and specific health conditions of every aerogenerator during its life and not only based on general guidelines. When thinking about a predictive maintenance strategy for aerogenerators, it is important to remark that windturbines are quite new machines using an important number of sensors able to supply information to different controllers in order to perform the best control and efficient operation of them. The information collected by the sensors of aerogenerators for control purposes can also be used for monitoring the health condition of their different components and to apply a predictive maintenance plan. According to this, no new investment in sensors is required in order to perform an effective windturbine predictive maintenance strategy because all the aerogenerators include a set of sensors from the manufacturer for different aspects of the control of their elements. The information from these sensors can also be used as main information source for a predictive maintenance plan. This paper presents the architecture of a new predictive maintenance system, called SIMAP, based on artificial M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 intelligent techniques. Its predictive maintenance strategy can be applied to any industrial system or equipment and its main goal is to find the most appropriate time to carry out the needed maintenance actions both from a component health condition and an incipient failure diagnosis perspectives. The new and positive aspects of this predictive maintenance methodology have been tested in windturbines. SIMAP is able to create and dynamically adapt a maintenance calendar for the windturbine that it is monitoring. The criteria followed are set up according to the real needs and operating life of the windturbine. This process is performed on-line and is different from the traditional scheduled maintenance plan based on fixed time intervals following the manufacturer criteria which do not focus on the real operation conditions of the aerogenerators. According to this, SIMAP implements the main aspects of an e-maintenance approach in a computer network such as local and remote continuous monitoring and diagnosis of the main components of the aerogenerators, maintenance actions planned according to the current and historical information collected, distribution of the on-line diagnosis and maintenance workload in different modules interconnected through a computer network and finally, different levels of warnings for the operator. This predictive maintenance system has been developed and applied to a windfarm owned by a Spanish wind energy company called Molinos del Ebro, S.A. This paper provides in Section 2 an overview of the main features and architecture of SIMAP and, in Sections 3–9, presents the capabilities of SIMAP applied to a particular case of a windturbine, which is the possible failures in a gearbox and how SIMAP works in real-time to detect and diagnose anomalies in this component, to evaluate its health condition and to plan predictive maintenance actions. 2. SIMAP: objectives and architecture The principal tasks performed by SIMAP are the following:
Continuous collection of data coming from different sensors installed in the aerogenerator and meteorological towers.
Continuous processing of the information collected in order to evaluate on-line the health condition of the aerogenerator components and also to detect if some symptoms of degradation or anomalies are present or could become present [3]. Both health condition evaluation and incipient fault detection are based on normal behaviour modelling (that is, in absence of failures) of the aerogenerator components. Thus, previously normal behaviour models were obtained using real data in order to characterize the normal dynamics of the representative variables of each component without any failure, taking into account both the different operation conditions of the components. SIMAP is working on-line taking current measurements from the process and evaluating the prediction of values from the models. The comparison between measured and predicted values of particular variables permits the incipient fault detection and the health condition evaluation for: a. Diagnosis of the root causes of the symptoms detected. 553 b. Failure risk forecasting of the aerogenerator components according to their actual health condition. c. Dynamical maintenance scheduling based on the machine condition, its environmental conditions and the aerogenerator production plan. Maintenance scheduling pursues to interfere the least possible with the production plan in order to maximize the aerogenerator availability and, also, to minimize the maintenance costs required. Other technical criteria considered are: - the failure risk of the aerogenerator components, estimated on-line based on their health condition; - the criticality of the components; - the maintenance actions efficiency to solve or mitigate the failure or degradation diagnosed; - the variable maintenance resources as well as the different relations among maintenance actions (precedence relations, compatibility relations, etc.). Effectiveness of the maintenance actions applied according to the change observed in the health condition and degradation of the components affected by these maintenance actions. This measurement will allow for both technical and economical comparisons of possible different maintenance strategies to be applied, as well as maintenance actions performed during different time periods. These tasks are organized in a modular architecture presented in Fig. 1 around the following six main modules: - Normal Behaviour Models. These models are able to predict on-line the normal behaviour (or reference behaviour) expected for each windturbine component, according to its current working and environmental conditions. These models are created mainly by means of artificial neural networks due to their ability to model dynamic non-linear industrial processes [4,5]. - Anomalies Detection Module. Its main goal is to detect possible anomalies in components by means of the results given by the normal behaviour models. Thus, by comparing for each component, its normal behaviour estimation with its real behaviour, both a normal behaviour deviation degree as well as an estimation certainty degree are obtained. These are used to recognize an anomaly present and the certainty of it [6–10]. - Health Condition Assessment Module. Its function is to evaluate on-line the health condition of each windturbine component as well as the general windturbine health condition. This function is performed by means of the results given by the normal behaviour models [11]. - Diagnosis Expert Module. Its main goal is to identify the possible failure modes that are present or developing in a windturbine component before this component faults in an irreversible way (for this reason, these detection and diagnosis tasks are called incipient) [12]. In order to reach this objective, this module employs a fuzzy expert system [14,15] able to represent in a flexible way both the knowledge and the uncertainty involved in this reasoning process, that is, mainly 554 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 Fig. 1. SIMAP architecture. the windturbine component relations between symptoms (anomalies and degradations) and failure modes. - Predictive Maintenance Scheduling Module. Its main goal is to optimally schedule windturbine maintenance actions, suitable to avoid or mitigate the detected incipient failures as well as the measured component degradation. This scheduling is optimal according to several technical and economical criteria which were previously mentioned. For the scheduling task, this module employs a fuzzy genetic algorithm [16–18] due to its ability to perform real and large-scale dynamical multi-objective non-linear optimizations with variable constraints. Fuzzy logic is employed in order to adequately represent maintenance tasks costs and duration uncertainties [19,20]. - Maintenance Effectiveness Assessment Module. Its main function is to obtain an effectiveness measurement for each maintenance action applied. Consequently, it allows for the assessment of the maintenance convenience from a technical and economical viewpoint. As an example, this index is calculated by means of measuring the aerogenerator gearbox health condition change before and after applying a maintenance action (see Fig. 2): efficiency ¼  through the corresponding models of normal behaviour in order to estimate its predicted normal behaviour according to the current working conditions. Once this is completed, both predicted and observed new sets of values are passed in parallel through the anomalies detection module and the health con- T ðtanh ðai2 Þ  tanh ðai1 ÞÞ 2T max According to Fig. 1, every new set of real measurements taken by the data acquisition system of an aerogenerator, is passed Fig. 2. Maintenance task efficiency calculation. M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 555
Optimization of aerogenerators life cycle by applying a maintenance strategy that pretends to delay or reduce components degradation.
Actual effectiveness measurement of maintenance actions applied, information that is important for getting the best scheduling of maintenance actions that have been proved to effectively solve or mitigate particular components degradations or incipient failures. All of the SIMAP modules have been developed in C++ language for MS-Windows operative system. Different data files contain the information required from various modules of SIMAP. The historical information collected by SIMAP is also stored in a set of files. The neural network models are based on multiplayer perceptrons. They were trained and tested using the neural network toolbox of MATLAB and finally they were integrated with the rest of the C++ code of SIMAP. Fig. 3. Gearbox and oil cooler system. dition assessment module. The first module is in charge of discovering if symptoms of an anomaly of possible failure mode is present because something does not correspond to the normal behaviour expected. At the same time and in any case, the health condition assessment module evaluates the health condition of each windturbine component. If one or several anomalies are detected, the Diagnosis Expert Module tries to identify the causes using experience stored in its knowledge base. According to the anomalies and causes detected and/or the health condition of the windturbine components, the predictive maintenance scheduling module will re-plan the maintenance actions to be executed. The aerogenerator maintenance calendar will be taken into account before being adapted to the new requirements of e-maintenance. Finally, and after the application of the predictive maintenance actions, the maintenance effectiveness assessment module will analyse the effectiveness of the maintenance performed from a qualitative point of view. Another important issue implemented in SIMAP is related to the automatic self-learning and refinement of the knowledge and functionality of their previous tasks. This method is based on the dynamical analysis of the effectiveness measurements of the maintenance actions applied as well as the health condition or degradation paths registered along the whole life of the aerogenerator components. This last information lets SIMAP characterize the health condition dynamics under degradation and failure processes and according to this knowledge, to forecast on-line the components failure time and to plan the best predictive maintenance strategy to avoid or delay this situation. The advantages of applying this predictive maintenance strategy may be stated as follows:
Maintenance intervals are frequently better adapted to the real needs of the windturbine than when using a preventive maintenance strategy with fixed maintenance intervals, because the life of the aerogenerator is taken into account.
It is more cost effective, and provides the most availability, reliability and security effectiveness. 3. Application of SIMAP to determine the health condition of a windturbine gearbox This section, and those that immediately follow, describe a real application of SIMAP focused on knowing the health condition of a windturbine gearbox in order to apply a predictive maintenance scheduling. SIMAP can analyse more aspects and also, can be applied to more complex components, however it was preferred to only focus on the evaluation of the health condition of this simple case as an example. First the physical system will be briefly described. Fig. 3 shows a diagram of a windturbine gearbox and its oil cooler system. It includes the physical layout of each component and the available sensors: gearbox bearing temperature, tank oil temperature and two digital signals informing if the cooler fan is operating on a slow or a fast speed mode. The main purpose of the windturbine gearbox is the conversion from the rotor slow speed to the electrical generator fast speed. Furthermore, the generated power of the windturbine is proportional to the wind speed and consequently to the rotor speed. Thus, the gearbox bearing temperature depends mainly on several working and environmental conditions: windturbine generated power, the nacelle temperature and the cooling phenomenon produced by the cooling system, which can be measured by the two digital signals of the cooler fans previously mentioned. Fig. 4 shows the typical temporal evolution of all these variables monitored on-line in a windturbine over a period of 2 weeks. The gearbox is one of the critical components in a windturbine. It is responsible for around 15–20% of its maintenance costs and also its downtime [2,24,13]. It is difficult to inspect and in case of replacement, it takes a great deal of time for a crane on top of the aerogenerator to dismount and mount the gearbox. In order to reduce these costs and downtime to the minimum, the strategy of emaintenance based on SIMAP has been applied. According to the previous information, the behaviour and health condition of the windturbine gearbox can be described by three characteristics: gearbox bearing temperature, cooling oil temperature of the gearbox and the difference between the 556 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 Fig. 4. Temporal evolution of gearbox main variables. hot and cold temperatures of the oil circulating through the gearbox cooling circuit. These three characteristics supply good arguments to do a qualitative evaluation about the behaviour and health condition of the windturbine. All the characteristics will be monitored online by three normal behaviour models. In general, in SIMAP the normal behaviour models are used to predict the evolution expected for key variables representative of the health conditions of the components monitored. The models are fitted before they are used as references of normal behaviour. They allow for the characterization of the typical relationships between a set of input variables and one or several output variables for all the possible working conditions considered as normal behaviour of the equipment or process monitored. Once the models are fitted, they can be used in real-time to compare their estimation to the real values of the variables predicted and, if a relevant deviation is observed, the normal performance expected is violated and an anomaly is discovered. The work scheme of the normal behaviour models is represented in Fig. 5. Three normal behaviour models are created for diagnosis and health monitoring of the windturbine gearbox: - Gearbox bearing temperature model; - Gearbox thermal difference model; - Cooling oil temperature model. The next section describes the process followed to obtain the normal behaviour model corresponding to the gearbox bearing temperature. The two models related to the gearbox, and other models developed in SIMAP for other components, follow the same procedure to be created. 4. Normal behaviour model for the windturbine gearbox bearing temperature In order to develop a normal behaviour model for the windturbine gearbox bearing temperature, a selection of accessible on-line variables that can explain such temperature must be done. The process of selection can be different depending on the type of model, but normally it is based on the physical symptoms that characterize the anomaly to be detected. In the case of the model proposed it is necessary to make the following considerations. The incidental wind mechanical power over the aerogenerator is converted in electrical power plus losses: Pwind ¼ Pgenerated þ Plosses The wind mechanical power is proportional to the air density, to the area covered by the tooth of the aerogenerator and to three times the wind speed: Pwind ¼ 12  r  A  v3 Fig. 5. Normal behaviour model. Work scheme. where r is the air density, A the area covered by the aerogenerator and v is the wind speed. According to these equations the turn speed of the windturbine gearbox and its work stress will depend mainly on the power generated. Also, the temperature of the main gearbox bearing will depend on the power generated and the temperature of the environment in the nacelle of the windturbine. The main gearbox bearing is cooled by oil that is circulating continuously. The oil is cooled in a heat exchanger by air 557 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 Table 1 Gearbox normal behaviour models Model Type Inputs Gearbox bearing temperature model Multilayer perceptron Gearbox bearing temperature (t  1, t  2) Generated power (t  3) Nacelle temperature (t) Cooler fan slow run (t  2) Cooler fan fast run (t  2) Gearbox thermal difference model Multilayer perceptron Gearbox thermal difference (t  1) Generated power (t  2) Nacelle temperature (t) Cooler fan slow run (t  2) Cooler fan fast run (t  2) Cooling oil temperature model Multilayer perceptron Cooling oil temperature (t  1) Generated power (t  2) Nacelle temperature (t) Cooler fan slow run (t  2) Cooler fan fast run (t  2) Fig. 6. Normal behaviour model for the gearbox bearing temperature of the windturbine. impulsed by a fan with two speeds: low and fast. Its operation also has to be taken into account. Once the a priori important variables for the health condition of the windturbine gearbox have been identified, the process of variable selection is completed by a statistical linear analysis in order to determine the importance of the influence of these variables in the gearbox bearing temperature. In order to do this, an analysis of cross-correlation and impulsional response between the explicative variables and the bearing temperature is done [21]. The information obtained is used to know the influence of the explicative variables on the bearing temperature and if their influence is present at the same instant or if it is delayed. The results of both analysis: cross-correlation and impulsional response, show the important influence of the explicative variables and also a delay of the maximum time of influence. Finally the model proposed for the normal behaviour of the gearbox bearing temperature is presented in Fig. 6. Fig. 7 shows an example of the performance of the model once it was fitted when it is receiving real information about its inputs. The figure shows the real bearing temperature of the gearbox, its estimated value from the model and the upper and lower confidence bands for the estimation along the time presented at a confidence level of 95%. The conclusion is that the prediction of the evolution of the gearbox bearing temperature corresponds well to the real evolution of this variable for any typical working condition of the aerogenerator. This model will make it possible to analyse if the gearbox is working under healthy conditions or it is under some stress that could produce a more serious failure. If the real evolution is going outside the bands of confidence, the gearbox is suffering some stress process that, depending on the severity and on the length of the time interval suffered, could evolve from an abnormal behaviour with minor consequences to a severe anomaly or fault. Normal behaviour models can be developed for characterization of several types of possible anomalies. These are very important tools for a qualitative characterization of the component health of an industrial process (in this case an aerogenerator), prevention of possible failures and planning of maintenance in a context of predictive maintenance. In particular for the case of the gearbox health condition monitoring three normal behaviour models are created and presented in Table 1. The first one on the table was previously presented. 5. Detection of incipient anomalies in the gearbox windturbine of an aerogenerator Fig. 7. Example of the normal behaviour model for the gearbox bearing temperature working in real-time. Once the normal behaviour models are obtained, it is possible to use them for the detection of possible anomalies and for refitting the maintenance planned according to the real health of the physical components. In order to do this the input and output variables of the normal behaviour models are taken in real-time and a prediction of the output is obtained in realtime too. The comparison between the values of the real and estimated output variables is used to detect possible anomalies to be diagnosed and to be mitigated by the corresponding maintenance action. The case of anomaly detection in a gearbox windturbine is analysed in the next paragraphs as an example of how this method works. Fig. 8 shows a real evolution of the gearbox bearing temperature using the normal behaviour model 558 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 rules as a method for knowledge representation and uncertainty based on fuzzy sets. This is represented in Fig. 1 as Diagnosis Expert Module. In order to demonstrate how SIMAP works in real-time diagnosing anomalies, the example of an anomaly from the previous section will be continued here. The anomalies detected in the gearbox of the A3 windturbine are passed on to the Diagnosis Expert Module of SIMAP. Its knowledge base contains information about the following failure modes in the windturbine gearbox: FM1: failure in the gearbox main bearing; FM2: unbalance of the gearbox main shaft; FM3: failure in the gearbox cooling circuit. The knowledge base of the expert system contains the following two rules related to these failure modes: Fig. 8. Real and estimated gearbox bearing temperature when an anomaly is present. obtained in the previous section. This figure covers a time period of 4 days (March 5–8 all inclusive). It is clear that around March 6, the real value of the bearing temperature is going outside the upper band of the normal value predicted for this temperature. The residual or difference between the real and estimated values of the bearing temperature is growing from March 6 on, but the gearbox is still working. Finally, on March 8 the gearbox fails and the windturbine is unavailable to produce energy. Also, Fig. 8 demonstrates the ability of a normal behaviour to detect anomalies before a catastrophic situation is present. Fig. 9 shows the gearbox thermal difference normal behaviour model during the same period of time shown in Fig. 8. Fig. 10 shows the cooling oil temperature model during the same period of time. It is easy to conclude from Figs. 8 and 9 that an anomaly is present on March 6 that does not correspond to the normal behaviour expected, however from Fig. 10 it is possible to deduce that the anomaly does not affect the cooling circuit of the gearbox. Therefore, the anomaly is on the side of the inner part of the gearbox. Once again these examples demonstrate that the normal behaviour models obtained are very useful for two different reasons. First, they can detect anomalies that do not correspond to normal behaviour and that can evolve to catastrophic failures. Second, they can be used for a qualitative estimation of the health condition of a component based on the stress or residual obtained from the difference between real and estimated values. Both aspects can be used for re-planning the predictive maintenance according to the real situation of a component. 6. Diagnosis of anomalies related to the gearbox of an aerogenerator Once anomalies have been detected it is necessary to find the causes so that SIMAP can try to diagnose them. For this purpose SIMAP uses a fuzzy expert system based on production According to the values that SIMAP took in real-time on March 6 (see Section 5) and once they were fuzzified using typical methods of fuzzyfication, the Diagnosis Expert Module had the information about the anomalies detected and also the following: - Gearbox main bearing temperature is HIGH with membership degree 1.0. - Gearbox thermal difference is HIGH with a membership degree 0.9. - Cooling oil temperature is NORMAL with a membership degree 1.0. The expert system put all this information in its facts database and found in its knowledge base the knowledge useful to diagnose the situation. Here, it found the previous two rules, but only the rule r_m1 satisfies its conditions in a degree of 0.9 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 559 Fig. 11. Fuzzy certainty degree of diagnostics concluded. Fig. 9. Real and estimated gearbox thermal difference when an anomaly is present. (minimum of the three conditions) and as a consequence, fired its conclusions using the fuzzy inference rule specified in the rule. This diagnosis rule concludes that the two possible failures exist along with their associated fuzzy certainty degrees. They are presented in Fig. 11. The inference process issued the following diagnoses: - Failure in the gearbox main bearing is certain in a degree 0.9752. - Unbalance of the gearbox main shaft is quite certain in a degree 0.75. 7. Diagnosis of anomalies related to the gearbox of an aerogenerator According to Fig. 1 in parallel to the Anomalies Detection Module and also using as inputs the normal behaviour models Fig. 10. Real and estimated cooling oil temperature when an anomaly is present. results, the Health Condition Assessment Module is able to conclude the gearbox health condition on-line. This module uses knowledge obtained from real evolutions of anomalies that finished in failures. The similarity of these histories of failures is compared and a pattern is obtained that can be used as reference of the health condition of the components. If the evolution of the life of a component is close to the pattern of some failure mode previously developed, it is possible to estimate the current health condition by reference to the end point of this failure pattern. As in previous sections, the case of the gearbox windturbine will be analysed as an example that demonstrates how the Health Condition Assessment Module works in real-time. Before this module can work in SIMAP, the failure patterns have to be developed, and for that, the history of failures in the gearbox has to be analysed. As an example the history of six past failures corresponding to the failure mode FM1 defined in Section 6 are analysed. Fig. 12 shows the evolution of the residual (distance between real and estimated values) corresponding to the three normal behaviour models available for the detection of the failure modes described in Table 1 in the windturbine gearbox. The residuals of all the histories are fuzzified and all the fuzzy sets are aggregated using the T-conorm maximum. Thus a common pattern can be obtained of typical residual values for all the histories along their evolutions to the failure mode FM1. Fig. 13 shows the fuzzy patterns resulting for the three models of normal behaviour analysed. Fig. 13a shows three fuzzy sets (from left to right): normal, high and very high for the gearbox bearing temperature. Fig. 13b shows three other fuzzy sets (from left to right): normal, high and very high for the gearbox thermal difference. Fig. 13c shows two fuzzy sets (from left to right): normal and high for cooling oil temperature of the gearbox. Once the fuzzy patterns for the failure mode FM1 have been obtained, it is possible to estimate their sensibility for identification of this failure mode. The fuzzy pattern corresponding to the gearbox bearing temperature has a sensibility of 1 (maximum of the scale), 0.8522 is the 560 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 sensibility of the fuzzy pattern corresponding to the gearbox difference thermal and 0 is the sensibility of the fuzzy pattern of cooling oil temperature for the failure mode FM1. Making similar analysis for the other two failure modes of the gearbox, it is possible to obtain a set of knowledge rules based on the real histories of failures that will be used for the estimation of the health condition of the gearbox. These rules include a certainty factor that depends on the sensibility of the fuzzy patterns to the different failure modes. For example, from the fuzzy pattern of residuals coming from the normal behaviour model of the gearbox bearing temperature it is possible to deduce the following three rules about the health condition of the gearbox: schedules a maintenance plan which contains the actions capable of avoiding or mitigating the failures present in the gearbox, according to several technical and economical criteria while maximizing windturbine availability and minimizing maintenance costs. In order to reach this functionality, this module performs several steps: - Gearbox failure time forecasting. This task employs gearbox failure histories for comparison and current time failure forecasting. - Preventive maintenance actions recommendations which try to avoid or mitigate diagnosed gearbox failures. - Maintenance tasks optimal time evaluation. This step compares the preventive maintenance plan with the corrective maintenance plan, both suitable to the diagnosed failures, and concludes the best maintenance moment for preventive tasks according to technical and economical criteria. - Finally, maintenance tasks scheduling according to their priority and optimal application times as well as other criteria such as: - Aerogenerator production plan. Maintenance scheduling pursues to interfere the least possible with the production plan for maximizing windturbine availability. - Maintenance tasks costs. - Variable maintenance resources (mainly personal, machines and non-renewal material) as well as different relations among maintenance actions: precedence relations, compatibility relations, etc. All these steps will be described in the next subsections. 8.1. Predictive maintenance to be applied to the gearbox of an aerogenerator These rules and the other corresponding to the different fuzzy models are used by the Health Condition Assessment Module in real-time to estimate the health condition of the different components. In the example developed here concerning the gearbox, its health condition is evaluated in real-time according to the information available in SIMAP and Fig. 14 shows its evolution. It is possible to see that the health of the gearbox passed from normal to bad condition on March 6, and finally on March 8 passed to very bad condition. The catastrophic failure in the gearbox happened some hours later. 8. Predictive maintenance to be applied to the gearbox of an aerogenerator Using both sources of information (failure diagnoses and gearbox health condition), the Predictive Maintenance Module The dynamic of the residuals coming from the normal behaviour models when a failure has occurred includes important information about the failure. It is similar to a signature of the failure. The residuals obtained from the normal behaviour models do not have information when a normal behaviour is present. The residuals have a probability distribution very close to the typical white noise. However, when an anomaly appears and it finishes in a failure, the proper residual has the information of the failure dynamic that does not correspond to normal behaviour. If this happens, it is possible to develop models for the residuals representing failures occurred. The following real case can demonstrate how SIMAP can predict the time remaining till failure of the windturbine gearbox for the example analysed in the previous sections. Fig. 7 presented the estimated and predicted values for the gearbox main bearing temperature. Their difference is the residual for this normal behaviour model. This evolution finished in a catastrophic failure of the gearbox, however suppose that this is not known and our time is set 2 days before March 6. Would it be possible to predict the time remaining till the failure? The response is yes. Taking as pattern of reference the historical dynamic evolution of residuals corresponding to failure modes in the M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 561 Fig. 12. Evolutions for six histories of the failure mode FM1 of the residuals corresponding to the normal behaviour models of: (a) gearbox bearing temperature, (b) gearbox thermal difference, (c) cooling oil temperature. gearbox, it is possible to know how close the current evolution is to the residual in an abnormal behaviour to the pattern of historical residuals for fault situations. Fig. 15a shows the evolution of the residuals in a historic case that finished in a gearbox failure (higher trend) and the current evolution of the residual for the normal behaviour model corresponding to the gearbox main bearing temperature. First, in order to use the historical residual pattern of failure of this model, it is necessary to test the similarity with the current evolution of the residual. Some pre-processing work has to be done before comparing both evolutions of residuals. It is necessary to make a translation of time origin to fit the evolutions of the abnormal behaviour to similar start time. This is obtained by trying different common starting points of anomalies and taking the situation where the distance between both series is minimum. Also, it is necessary to ensure that both evolutions of residuals to be compared correspond to similar conditions. In order to do this, a model about the dynamical of the residuals for the historical case that finished in the gearbox failure is created. This model represents how the failure appears added to the normal behaviour model and characterises the failure. The model found for this case is the following: This model shows that the evolution of the residual when a failure is present depends on the working conditions of the gearbox and, also, from the previous residual of the gearbox health. Once the model is obtained, the current conditions of the gearbox are passed as inputs through the model. This allows for verification if the current working conditions of the gearbox stimulates the failure model in a similar way to how the residuals evolve now. The similarity is observed between the 562 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 Fig. 13. (a–c) Fuzzy patterns for the residuals of the three normal behaviour models aggregating information of six histories that finished in the failure mode FM1. current evolution of the residual of the normal behaviour model for the gearbox main bearing temperature and the output of the residual model built with historic data of gearbox failures but excited with current conditions as inputs. Finally, it is possible Fig. 14. Gearbox health condition during failure period. to obtain a possibility distribution from the different distances and this can be approximated by a triangular fuzzy set like this presented in Fig. 15b. Using the model of the residuals previously mentioned and the current evolution of the gearbox life, it is possible to predict the evolution of the residual and also its uncertainty. The prediction is done from an instant of time till the moment where the historical pattern of failure finished. In this case if the date is March 6 with the anomaly present, it is possible to predict the time remaining till failure, if the current conditions correspond to a failure of the gearbox that can be detected by the normal behaviour model of the main bearing temperature. This is shown in Fig. 15c. Finally, Fig. 15d presents the fuzzy time to failure. The centre of gravity of this triangle is 26.5 h. This means that from now (March 6) the most important possibility of failure will occur in 26.5 h and a probable interval of failure between 24.5 and 28.5 h. It is easy to observe that the prediction of failure time corresponds very well with to reality because on March 8 this gearbox had a catastrophic failure. M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 563 Fig. 15. (a–d) Gearbox failure time forecasting. This entire process is done for every series of residuals coming from different models of normal behaviour. Also this process is repeated when new information is coming to SIMAP about anomalies and the prediction of times to failure are updated according to the real life of the aerogenerator for each moment. Fig. 16 shows an example of similarity between residuals of a current and pre-processed historical evolution of failure conditions. They fit very well. 8.2. Predictive maintenance to be applied to the gearbox of an aerogenerator In Section 6, the Diagnostic Expert Module of SIMAP diagnosed two possible causes of the anomaly detected. They were: - Failure in the gearbox main bearing is certain in a degree 0.9752. - Unbalance of the gearbox main shaft is quite certain in a degree 0.75. According to these diagnoses, the maintenance expert system included in the Predictive Maintenance Module (Fig. 1), selects the suitable knowledge rules to recommend maintenance actions. They can be corrective maintenance actions or preventive maintenance actions. The following two rules were selected from the knowledge base for this case. The numbers are the certainty factors. The priorities of the maintenance actions are elaborated using the criticality of the components and the certainty of the diagnoses issued. In this case, the maintenance actions recommended would be: PMA1: gearbox bearing repair, priority 0.89; CMA1: gearbox replacement, priority 0.9; 564 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 distributions is the best moment to execute a preventive maintenance action according to the health of the gearbox and the costs involved. This cross-point is the moment where the preventive maintenance plan reaches a similar profile to the corrective maintenance plan. 8.4. Optimal time to apply the maintenance actions to correct the gearbox anomalies Fig. 16. Similarity between residuals of current and pre-processed historical evolution of failure conditions. PMA2: shaft alignment, priority 0.89; CMA2: shaft replacement, priority 0.9. 8.3. Optimal time to apply the maintenance actions to correct the gearbox anomalies According to all the previous information about the anomalies detected in the gearbox, some maintenance actions can be executed. Some of them are of preventive nature and others are corrective. SIMAP evaluates in realtime if it is possible to perform preventive maintenance actions at the current moment or if it is better to perform a corrective plan. In order to do this, several considerations are performed around both plans of maintenance. As an example, next the optimum time will be estimated to perform a preventive maintenance action for the gearbox bearing replacement. The idea is to compare from different items both plans: preventive and corrective and to find the crosspoint of both. This point will be the optimum moment to perform a preventive maintenance action taking into account the health condition of the component and the economical impact. The optimal time to perform a maintenance action is calculated taking into account the risk of failure, estimated from the prediction of the time to failure of the gearbox (see Section 8.1), the criticality of the component to maintain, the index of convenience for apply the maintenance action from a technical point of view and the costs of both types of maintenance plans: preventive and corrective. All the processes for the case analysed are presented in Fig. 17. It is based on a combination of the six different fuzzy sets that represent the mentioned factors on the left side of Fig. 17. The conclusions of these combinations are two fuzzy distributions: one for the preventive maintenance plan and another one for the corrective maintenance plan. They are represented at the top right of Fig. 17. The cross-point of both possibility The scheduling of maintenance actions is a dynamical fuzzy maintenance scheduling task [16,22,23]. Fuzzy scheduling is used because the maintenance cost and duration uncertainties are represented by means of fuzzy sets. The maintenance is dynamical because each time any of the scheduling inputs change (production plan, maintenance actions to schedule, maintenance resources availabilities, etc.) this module reschedules on-line a new optimal maintenance plan, according to this new situation. Next a real case of rescheduling is briefly described in order to present the main features of the scheduling of maintenance tasks. In Section 8.3 two maintenance actions were recommended in the windturbine AE3. They were: PMA1: gearbox bearing repair, priority 0.89; PMA2: shaft alignment, priority 0.89. Both tasks, coded as MA3 and MA8 in this case, must be done when a previous maintenance has been done MA2 (generator shutdown and gearbox opening and cleaning). Furthermore, at the same time SIMAP recommended to perform the following maintenance actions in other aerogenerators: - two maintenance actions, MA0 and MA5, for fitting the control system of the aerogeneretors AE24 and AE38; - one maintenance action, MA7, for cleaning the hydraulic system in the aerogenerator AE40; - two maintenance actions, MA1 and MA6, for cleaning the cooling systems of the aerogenerators, AE11 and AE12; - two maintenance actions, MA4 and MA9, for testing communications with two points from the windfarm. The information about each maintenance action includes: duration, priority, required conditions to be done, costs, personnel, equipment and material required, compatibility to be executed with other maintenance actions and previous actions required. In this example all this information about the ten maintenance actions mentioned is presented in Table 2. The information found in Table 2 plus that which is related to the production planning and availability of the different resources required are the input to an genetic algorithm based on fuzzy sets. This genetic algorithm tries to minimize time and overall costs of the maintenance actions planned keeping the production plan and using the available resources. The results obtained for the scheduling of 565 Fig. 17. Optimal time to apply a maintenance action. M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 566 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 Table 2 Maintenance action information 9. Evaluation of the maintenance effectiveness and cost Finally, a real case is presented showing the estimation of the maintenance effectiveness of two different maintenance tasks: - 24-month preventive gearbox maintenance set (MP24); - 36-month preventive gearbox maintenance set (MP36). maintenance actions are presented by SIMAP in a format similar to that presented in Fig. 18. If a rescheduling is performed when a maintenance planning is in execution, the genetic algorithm takes into account this fact. There are six available maintenance histories corresponding to similar aerogenerators: two for the MP36 maintenance action set and four for the MP24 set. Fig. 19 shows the calculated efficiency of these maintenance tasks in respect to the measured gearbox health condition at the moment that each maintenance action was applied. It is possible to observe that in the case of the MP24 maintenance action set, there is a clear inverse relation between the gearbox health condition and maintenance efficiency and from here the conclusion is that when the gearbox health condition is worst (that is, the most degraded gearbox), much lower is the efficiency of the MP24 maintenance action and vice verse. In the case of the MP36 maintenance set, no clear relation is observed and therefore no conclusion can be determined between maintenance efficiency Fig. 18. Maintenance scheduling. M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 567 - perform a dynamical multi-objective non-linear optimization with constraints, by means of genetic algorithms; - represent the uncertainty inherent to the knowledge issued, by means of fuzzy logic. An example of how SIMAP works has been presented, focusing on the on-line health condition monitoring of a windturbine gearbox. Future works are oriented to the monitoring and experience derived from the new maintenance plan implemented in different windfarms. References Fig. 19. MP36 and MP24 maintenance sets efficiency evaluation. and the gearbox health condition. Another conclusion is that the MP36 set is more efficient than the MP24 set, this is reasonable because the MP36 set comprises more maintenance actions than the MP24 set. 10. Conclusions The main new features of SIMAP in the field of diagnosis and maintenance introduced in this paper have been the following: - the integration and cooperation of every task involved in a formal predictive maintenance strategy, that is, mainly: continuous monitoring, incipient failure detection and diagnosis, health condition evaluation, predictive maintenance scheduling and effectiveness measure of maintenance actions performed; - on-line and automatic components health condition evaluation, based on a degradation perspective; - a maintenance scheduling method which considers both technical and economical criteria; - on-line, direct and automatic measurement of applied maintenance actions effectiveness, by means of the change observed in the health condition and degradation of the components affected by these maintenance actions. Furthermore, this study concludes that artificial intelligence and modelling techniques are adequate for reaching the main goals of this predictive maintenance strategy, due mainly to their ability to: - model dynamic non-linear industrial processes, by means of artificial neural networks; - characterize and represent both quantitative knowledge coming from historical data (by means of artificial neural networks) as well as qualitative knowledge coming from maintenance and operation experts (by means of expert systems); [1] D.N.V. Riso, Guidelines for Design of Wind Turbines, 2001. [2] Danish Wind Turbine Manufacturers Association, Operation and maintenance costs for wind turbines, http://www.windpower.dk, 2000. [3] A. Davies, Handbook of Condition Monitoring, Chapman & Hall, 1998. [4] S. Haykin, Neural Networks: A Comprehensive Foundation, IEEE Press, 1994. [5] A. Muñoz, M.A. Sanz-Bobi, An incipient fault detection system based on the probabilistic radial basis function network. Application to the diagnosis of the condenser of a coal power plant, Neurocomputing 23 (1998) 177–194. [6] R. Iserman, Process fault detection based on modelling and estimation methods, Automatica 20 (1984) 387–404. [7] R.J. Patton, P.M. Frank, R.N. Clark, Issues of Fault Diagnosis for Dynamic Systems, Springer-Verlag, 2000. [8] M.A. Sanz-Bobi, J. Gilmartı́n, Condition monitoring of a hydraulic pump using a self-organising map, in: Proceedings of the Scientific Conference Artificial Intelligence in Industry, High Tatras, Slovakia, (1998), pp. 267– 274. [9] M.A. Sanz-Bobi, M.A. Donaire Toribio, Diagnosis of electrical motors using artificial neural networks, in: IEEE International Symposium on Diagnostics for Electrical Machines, Power Electronics and Drives (SDEMPED), Gijón, Spain, (1999), pp. 369–374. [10] M.A. Sanz-Bobi, R. Palacios, A. Muñoz, R. Garcı́a-Escudero, M. Pérez, A.L. Matesanz, ISPMAT: Intelligent System for Predictive Maintenance Applied to Trains, EUROMAINTENANCE, Helsinki, Finland, 2002 . [11] H. Kumamoto, E. Henley, Probabilistic Risk Assessment and Management for Engineers and Scientists, Institute of Electrical and Electronics Engineers Press, 1996. [12] J. Rasmusen, Diagnostic reasoning in action, IEEE Transactions on Systems Man and Cybernetics 23–24 (1993) 981–992. [13] L. Rademakers, H. Braam, M.B. Zaaijer, G. van Bussel, Assessment and optimisation of operation and maintenance of offshore wind turbines. ECN Report: ECN-RX-03-044, June 2003. [14] R. Yager, L. Zadeh, An Introduction to Fuzzy Logic Applications in Intelligent Systems, Kluwer Academic Publishers, 1992. [15] J. Villar, M.A. Sanz-Bobi, Semantic analysis of fuzzy models, application to Yager models, in: Nikos Mostorakis (Ed.), Advances in Fuzzy Systems and Evolutionary Computation, World Scientific and Engineering Society Press, 2001, pp. 82–87. [16] M.C. Garcı́a, M.A. Sanz-Bobi, Dynamic Scheduling of Industrial Maintenance Using Genetic Algorithms, EUROMAINTENANCE, Helsinki, Finland, 2002. [17] D.E. Godlberg, Genetics Algorithms in Search, Optimization & Machine Learning, Addison-Wesley, 1989. [18] D.A. Van Veldhuizen, G.B. Lamont, Multi-Objective Evolutionary Algorithms: Analyzing the State-of-the-Art. Evolutionary Computation, MIT Press Journals 8 (2) (2000). [19] S. Martorell, A. Muñoz, V. Serradell, Age-dependent models for evaluating risks and costs of surveillance and maintenance of components, IEEE Transactions on Reliability 45–3 (1996) 433–442. 568 M.C. Garcia et al. / Computers in Industry 57 (2006) 552–568 [20] A. Chande, R. Tokekar, Expert-based maintenance: a study of its effectiveness, IEEE Transactions on Reliability 47 (1998) 53–58. [21] G.P. Box, G.M. Jenkins, Time series analysis, in: Forecasting and Control, Holden Day Inc., 1976. [22] K. Deb, Multi-Objective Genetic Algorithms: Problem Difficulties and Construction of Test Problems. Evolutionary Computation, MIT Press Journals 7 (3) (1999). [23] M. Jahangirian, Intelligent dynamic scheduling system: the application of genetic algorithms, Integrated Manufacturing Systems (2000) 11–14. [24] G. Munson, The Maintenance Challenge in Wind Power, North American WINDPOWER, Dec 2004. Dr. Miguel A. Sanz Bobi is professor at the Computer Science Department and also researcher at the Institute for Research and Technology (IIT) both inside the Engineering School of the Pontificia Comillas University, Madrid (Spain). He divides his time between teaching and research in the artificial intelligence field applied to diagnosis and maintenance of industrial processes. He has been the main researcher in more than 35 industrial projects over the last 20 years related to the diagnosis in real-time of industrial processes, incipient detection of anomalies based on models, knowledge acquisition and representation, reliability and predictive maintenance. All these projects have been based on a combination of artificial intelligence, new information technologies and data mining techniques. Maria Cruz Garcı́a is PhD in industrial engineering at University Pontificia Comillas, Spain. She worked at National Grid Company, UK, in 1997, developing a real-time monitoring and analysis system of the electrical network stability and robustness. Afterwards, she worked at the Technology Investigation Institute in the University Pontificia Comillas as a researcher involved in artificial intelligence projects in collaboration with Spanish utility and wind-power generation companies (Repsol, Unión Fenosa, Molinos del Ebro). Her PhD thesis was related to the planning and effectiveness assessment of predictive maintenance applied to industrial processes, using for that purpose dynamic modelling techniques as well as artificial intelligence. In 2004, she joined the Spanish Savings Bank Caja Madrid, working in a research project for developing real-time automatic trading systems for fixed income futures, equity and forex products. These trading systems use neural networks, quantitative and technical analysis techniques. Currently, she is a quantitative credit analyst at BSCH. Her areas of interest are: artificial intelligence, data mining, quantitative techniques applied to financial markets, process modelling and signal processing techniques. Javier del Pico-Aznar is a mechanical engineer at the University of Zaragoza in Spain. During his years of engineering studies, he worked at various different power plants. At present, he is the Director of the energy area of SAMCA, a Spanish company working in different industrial sectors including energy generation. He has participated in the development of several projects concerning the optimization of processes, cogeneration plants and windfarms.