Keywords
Instituto de Investigación Tecnológica (IIT), anomaly detection, pattern discovering, normal behavior characterization, self-organizing maps, k-means,probability density functions, engine diagnosis
Method for condition characterization of industrial components by dynamic discovering of their pattern behaviour
Abstract
This paper presents a method to assess the behavior of an industrial component by a set of typical patterns which characterize the normal behavior of the component. Once such patterns are defined they can be used both for anomaly detection and diagnosis, and suggestion of maintenance re-scheduling. A main novelty introduced in the method presented is that the behavior pattern of an industrial component is defined progressively by multiple clusters discovered from characteristic feature values registered during the period of observation of the component. Each feature cluster is made up of two main elements: a centroid, which represents the most representative feature values within the same cluster (the pattern itself); and the probability density distributions (PDFs) of the feature values that belong to each pattern cluster (the domain of the pattern discovered). Clusters are obtained by unsupervised clustering algorithms such as Self-Organizing Maps (SOM) and/or K-means. The method presented includes the definition of two new indicators for the behavior assessment based on the patterns discovered. The first indicator (similarity) is obtained through the location of each new observation within the PDF of the cluster to which it belongs. The second indicator (distance) is based on the Euclidean feature distance between observations and their nearest centroids. These two indicators are combined to obtain a single Health Index (HI) used to estimate the Remaining Useful Life (RUL) of the component. In addition, this paper includes a novel approach that has been experimented focused on the prognosis and forecasting of the RUL of the component once its behavior patterns are discovered. Such approach has been implemented using Recurrent Neural Networks (RNN). All these concepts are applied to a real example of industrial process including comments about the results obtained.
This paper presents a method to assess the behavior of an industrial component by a set of typical patterns which characterize the normal behavior of the component. Once such patterns are defined they can be used both for anomaly detection and diagnosis, and suggestion of maintenance re-scheduling. A main novelty introduced in the method presented is that the behavior pattern of an industrial component is defined progressively by multiple clusters discovered from characteristic feature values registered during the period of observation of the component. Each feature cluster is made up of two main elements: a centroid, which represents the most representative feature values within the same cluster (the pattern itself); and the probability density distributions (PDFs) of the feature values that belong to each pattern cluster (the domain of the pattern discovered). Clusters are obtained by unsupervised clustering algorithms such as Self-Organizing Maps (SOM) and/or K-means. The method presented includes the definition of two new indicators for the behavior assessment based on the patterns discovered. The first indicator (similarity) is obtained through the location of each new observation within the PDF of the cluster to which it belongs. The second indicator (distance) is based on the Euclidean feature distance between observations and their nearest centroids. These two indicators are combined to obtain a single Health Index (HI) used to estimate the Remaining Useful Life (RUL) of the component. In addition, this paper includes a novel approach that has been experimented focused on the prognosis and forecasting of the RUL of the component once its behavior patterns are discovered. Such approach has been implemented using Recurrent Neural Networks (RNN). All these concepts are applied to a real example of industrial process including comments about the results obtained.
This paper presents a method to assess the behavior of an industrial component by a set of typical patterns which characterize the normal behavior of the component. Once such patterns are defined they can be used both for anomaly detection and diagnosis, and suggestion of maintenance re-scheduling. A main novelty introduced in the method presented is that the behavior pattern of an industrial component is defined progressively by multiple clusters discovered from characteristic feature values registered during the period of observation of the component. Each feature cluster is made up of two main elements: a centroid, which represents the most representative feature values within the same cluster (the pattern itself); and the probability density distributions (PDFs) of the feature values that belong to each pattern cluster (the domain of the pattern discovered). Clusters are obtained by unsupervised clustering algorithms such as Self-Organizing Maps (SOM) and/or K-means. The method presented includes the definition of two new indicators for the behavior assessment based on the patterns discovered. The first indicator (similarity) is obtained through the location of each new observation within the PDF of the cluster to which it belongs. The second indicator (distance) is based on the Euclidean feature distance between observations and their nearest centroids. These two indicators are combined to obtain a single Health Index (HI) used to estimate the Remaining Useful Life (RUL) of the component. In addition, this paper includes a novel approach that has been experimented focused on the prognosis and forecasting of the RUL of the component once its behavior patterns are discovered. Such approach has been implemented using Recurrent Neural Networks (RNN). All these concepts are applied to a real example of industrial process including comments about the results obtained.
Información detallada
Materias, derechos, colecciones e identificadores