Sylvain Létourneau

Data mining to predict aircraft component replacement

Aircraft sensors generate vast amounts of data, much of which languishes in storage after its initial analysis. The authors have developed an approach for using this data to build models for predicting aircraft component failure. Their approach addresses several key data-mining issues.

Learning to predict train wheel failures

This paper describes a successful but challenging application of data mining in the railway industry. The objective is to optimize maintenance and operation of trains through prognostics of wheel failures. In addition to reducing maintenance costs, the proposed technology will help improve railway safety and augment throughput. Building on established techniques from data mining and machine learning, we present a methodology to learn models to predict train wheel failures from readily available operational and maintenance data. This methodology addresses various data mining tasks such as automatic labeling, feature extraction, model building, model fusion, and evaluation. After a detailed description of the methodology, we report results from large-scale experiments. These results clearly show the great potential of this innovative application of data mining in the railway industry.

Application of Rough Sets Algorithms to Prediction of Aircraft Component Failure

This paper presents application of Rough Sets algorithms to prediction of component failures in aerospace domain. To achieve this we first introduce a data preprocessing approach that consists of case selection, data labeling and attribute reduction. We also introduce a weight function to represent the importance of predictions as a function of time before the actual failure. We then build several models using rough set algorithms and reduce these models through a postprocessing phase. End results for failure prediction of a specific aircraft component are presented.

Developing Data Mining-Based Prognostic Models for CF-18 Aircraft

The CF-18 (CF denotes Canadian Forces) aircraft is a complex system for which a variety of data are systematically being recorded: flight data from sensors, built-in test equipment data, and maintenance data. Without proper analytical and statistical tools, these data resources are of limited use to the operating organization. Focusing on data mining-based modeling, this paper investigates the use of readily available CF-18 data to support the development of prognostics and health management systems. A generic data mining methodology has been developed to build prognostic models from operational and maintenance data. This paper introduces the methodology and elaborates on challenges specific to the use of CF-18 data from the Canadian Forces. A number of key data mining tasks are examined including data gathering, information fusion, data preprocessing, model building, and model evaluation. The solutions developed to address these tasks are described. A software tool developed to automate the model development process is also presented. Finally, this paper discusses preliminary results on the creation of models to predict F404 no. 4 bearing and main fuel control failures on the CF-18.

Two-stage classifications for improving time-to-failure estimates: a case study in prognostic of train wheels

In order to meet the need for higher equipment availability and lower maintenance cost, much attention is being paid to the development of prognostic systems. Such systems support a proactive maintenance strategy by continuously monitoring the components of interest and predicting their failures sufficiently in advance to avoid disruptions during operation. Recent research demonstrated the potential of a comprehensive data mining methodology for building prognostic models from readily available operational and maintenance data. This approach builds a binary classifier that can determine the likelihood of a failure within a broad target window but cannot provide precise time to failure (TTF) estimations. This paper introduces a two-stage classification approach that helps improve the precision of TTF estimations. The new approach uses the initial methodology to learn a variety of base classifiers and then relies on meta-learning to integrate them. The paper details the model building process and demonstrates the usefulness of the proposed approach through a real-world prognostic application.

Improving preciseness of time to failure predictions: Application to APU starter

Despite the availability of huge amounts of data and a variety of powerful data analysis methods, prognostic models are still often failing to provide accurate and precise time to failure estimations. This paper addresses this problem by integrating several machine learning algorithms. The approach proposed relies on a classification system to determine the likelihood of component failures and to provide rough indications of remaining life. It then introduces clustering and SVM-based local regression to refine the time to failure estimations provided by the classification system. The paper illustrates the applicability of the proposed approach through a real world aerospace application and discusses data pre-processing requirements. The preliminary results show that the proposed method can reduce uncertainty in time to failure estimates, which in turn helps augment the usefulness of prognostics.

Machine learning-based methods for TTF estimation with application to APU prognostics

Machine learning-based predictive modeling is to develop machine learning-based or data-driven models to predict failures before they occur and estimate the remaining useful life or time to failure (TTF) accurately. Accurate TTF estimation plays a vital role in predictive maintenance or PHM (Prognostic and Health Management). Despite the availability of large amounts of data and a variety of powerful data analysis methods, predictive models developed for PHM still fail to provide accurate and precise TTF estimations. This paper addresses this problem by integrating machine learning algorithms such as classification, regression and clustering. A classification system is used to determine the likelihood of component failures such that rough indications of TTF are provided. Clustering and SVM-based local regression are then introduced to refine the time to failure estimations provided by the classification system. The paper illustrates the applicability of the proposed approach through a real world aerospace application with details on data pre-processing requirements. The results demonstrate that the proposed method can reduce uncertainty in estimating time to failure, which in turn helps augment the usefulness of predictive maintenance.

Comparison of metrics to monitor and compensate for piezoceramic debonding in structural health monitoring

This article investigates metrics to assess and compensate for the degradation of the adhesive layer of surface-bonded piezoceramic transducers for structural health-monitoring applications. Capacitance, resonance frequency, and modal damping parameters are derived from admittance curves using a lumped parameter model to monitor the degradation of the transducer adhesive layer. A pitch-catch configuration is then used to discriminate the effect of bonding degradation on actuation and sensing. It is shown that below the first mechanical resonance frequency of the piezoceramic transducers, the degradation causes a decrease in the amplitude of the transmitted and received signals, while above resonance, in addition to a decrease in the amplitude of the transmitted and received signals, a linear phase shift is observed. A signal-correction factor is proposed to adjust signals based on adhesive degradation evaluated using the measured modal damping. The benefits of the signal-correction factor are demonstrated in the frequency domain for both the A0 and S0 modes.

Model evaluation for prognostics: estimating cost saving for the end users

Unexpected failures of complex equipment such as trains or aircraft introduce superfluous costs, disrupt operation, have an effect on consumers satisfaction, and potentially decrease safety in practice. One of the objectives of prognostics and health management (PHM) systems is to help reduce the number of unexpected failures by continuously monitoring the components of interest and predicting their failures sufficiently in advance to allow for proper planning. In other words, PHM systems may help turn unexpected failures into expected ones. Recent research has demonstrated the usefulness of data mining to help build prognostic models for PHM but also has identified the need for new model evaluation methods that take into account the specificities of prognostic applications. This paper investigates this problem. First, it reviews classical and recent methods to evaluate data mining models and it explains their deficiencies with respect to prognostic applications. The paper then proposes a novel approach that overcomes these deficiencies. This approach integrates the various costs and benefits involved in prognostics to quantify the cost saving expected from a given prognostic model. From the end users perspective, the formula is practical as it is easy to understand and requires realistic inputs. The paper illustrates the usefulness of the methods through a real-world case study involving data-mining prognostic models and realistic costs/benefits information. The results show the feasibility of the approach and its applicability to various prognostic applications.We introduce a polynomial-time algorithm to learn Bayesian networks whose structure is restricted to nodes with in-degree at most k and to edges consistent with the optimal branching, that we call consistent k-graphs (CkG). The optimal branching is used as an heuristic for a primary causality order between network variables, which is subsequently refined, according to a certain score, into an optimal CkG Bayesian network. This approach augments the search space exponentially, in the number of nodes, relatively to trees, yet keeping a polynomial-time bound. The proposed algorithm can be applied to scores that decompose over the network structure, such as the well known LL, MDL, AIC, BIC, K2, BD, BDe, BDeu and MIT scores. We tested the proposed algorithm in a classification task. We show that the induced classifier always score better than or the same as the Naive Bayes and Tree Augmented Naive Bayes classifiers. Experiments on the UCI repository show that, in many cases, the improved scores translate into increased classification accuracy.

Data mining to detect abnormal behavior in aerospace data

The operation and maintenance of todays aircraft is a complex task. It requires use of some state-of-the-art data mining facilites that are not currently available. This paper is about dev elopment and use of data mining techniques to detect abnormal situations in aircraft operation. Using historical sensor data, that is normally generated during the operation of aircraft, w e induce models to predict abonormal situations in aircraft engines. The method involv es creating new features from raw data and identifying trends in particular parameters of interest. We describe how models generated from individual aircraft with abnormal situations can be combined to generate a single model. We evaluate our approach using over 5 y ears of historical data from the operation of engines of 34 Airbus A-320s.