روش مبتنی بر شباهت برای تشخیص و پیش بینی عملکرد فرآیند تولید
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|21828||2007||9 صفحه PDF||سفارش دهید||5197 کلمه|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Computers in Industry, Volume 58, Issue 6, August 2007, Pages 558–566
Full realization of all the potentials of predictive maintenance highly depends on the accuracy of long-term predictions of the remaining useful life of manufacturing equipments. In this paper, we propose a new method that is capable of achieving high long-term prediction accuracy by comparing signatures from any two degradation processes using measures of similarity that form a match matrix (MM). Through this concept, we can effectively include large amounts of historical information into the prediction of the current degradation process. Similarities with historical records are used to generate possible future distributions of features indicative of process performance, which are then used to predict the probabilities of failure over time by evaluating overlaps between predicted feature distributions and feature distributions related to unacceptable equipment behavior. The analysis of experimental results shows that the proposed method can yield a noticeable improvement of long-term prediction accuracy in terms of mean prediction errors over the Elman Recurrent Neural Network (ERNN) based prediction, which was shown in the past literature to predict well behavior of highly non-linear and non-stationary time series.
Reducing downtime cost and achieving near zero downtime is the ultimate goal of predictive maintenance. However, it is impossible to realize all the advantages of predictive maintenance without accurate predictions of the remaining useful life before the actual failure occurs. The inaccurate predictive information may result in either unnecessary maintenance, such as early replacement of components, or production downtime because of unexpected machine failures. Therefore, the accuracy of remaining useful life prediction, particularly the long-term prediction, which gives sufficient time to prepare for a maintenance operation, plays an essential role in the full realization of the potentials of predictive maintenance. The degradation process cannot be directly observed or measured in general. It can only be observed indirectly through the time series of features extracted from available process measurements, such as vibrations and forces. Extrapolating these time series in time can help us to predict risks of failure or unacceptable behavior of the process over time. This places great significance on one's ability to accurately and reliably predict the feature time series. A variety of techniques have been used in the past for time series modeling and prediction. Parametric linear prediction techniques, such as Auto-Regressive Moving Average (ARMA)  and  or Kalman filtering , may work well only for short-term predictions because of their assumption that the considered time series is generated from a linear process. These linear prediction techniques are well interpretable but with limited capabilities for predicting real world problems which are usually complex and non-linear. Variety of approaches for predicting non-linear time series, such as fuzzy time series and clustering  and , multi-resolution wavelet models ,  and  and neural networks , has been extensively studied in the literature. Without a priori knowledge about the time series under consideration, selecting an appropriate non-linear model and its structure is a difficult task. Among these non-linear prediction technique, neural networks, such as Radial Basis Function (RBF) networks  Multi-Layer Perceptron (MLP) neural network  and Recurrent Neural Networks (RNN) ,  and , maybe the most extensively applied techniques for complex non-linear time series predictions because of their capability to approximate non-linear functional and dynamic dependencies . Unlike feed-forward networks such as RBF and MLP, which have limitations of identifying temporal relationships in the time series, RNN takes into account temporal dependencies through local or global feedback connections in the network. As a result, RNN is able to approximate a wide class of non-linear dynamical systems . However, the commonly used gradient descent algorithms for RNN training exhibit certain problems during training, such as having difficulty dealing with long-term dependencies in the time series , which in turn limits their capability of achieving accurate long-term predictions. In addition, finding a suitable number of hidden neurons and appropriate RNN structure remains a challenging problem. In many manufacturing facilities, large amounts of historical records of past equipment behavior are available and can be used to enhance and reinforce the equipment performance prediction. The goal of the method pursued in this paper is to increase long-term prediction of signatures depicting equipment performance through incorporation of historical records into the prediction process, while at same time capturing the dynamic changes of the signatures as the process changes. The following terminology will be used throughout the paper: • The term feature vectors will be used for signatures describing the current state of the machine/process and containing a number of features which are considered to be correlated to the process degradation. The evolution/dynamics of those feature vectors over time is then essentially the characteristic of the degradation process. Once appropriate sensors and adequate features are selected, one can estimate the current degradation state and conduct prediction based on the evolution of the feature vectors. • The time interval between two consecutive maintenance cycles will be referred to as a run. The maintenance operation may be a component replacement, part repair, etc. From the set of sensor readings in each past run, a time-ordered sequence of feature vectors can be extracted to represent the degradation states of the process. The time ordering of the features is necessary in order to be able to explore the temporal evolution of the feature vectors and characterize the process degradation over time. The rest of the paper is organized as follows. In Section 2, the new match matrix (MM) based prediction method is introduced. The newly proposed approach has been tested on predicting tool degradation in a boring process and the results are shown in Section 3. Section 4 provides conclusions.
نتیجه گیری انگلیسی
A novel prediction algorithm capable of dealing with a long term prediction of non-stationary multivariate time series is presented in this paper. The method is based on the concept of match matrices for comparison of signatures describing the current degradation process with those observed in the past degradation processes of the same machine/process. It provides one with a natural way to derive the predicted feature distributions from the prediction uncertainties, which could be used to obtain information about predicted probabilities of failure or unacceptable behavior. The new method was tested in predicting signatures extracted from spindle load profiles of a boring process in an automotive manufacturing plant and results showed that the newly proposed prediction method based on match matrices yields noticeably smaller mean-squared errors of prediction, compared with ARMA based prediction and Recurrent Neural Network (RNN) based prediction. In the case that a large number of past runs are present in historical records, comparing the current run with all the past runs and conducting random sampling within each match matrix may be computationally demanding and inefficient. Significant reduction in the computational load can be achieved if one could observe groups of similar degradation dynamics expressed through similar performance features evolution, and evaluate match matrices only within the group that is the most similar to the degradation process in the currently observed run. A Hidden Markov Model (HMM) based technique as reported in ref.  can be possibly used to select the most similar runs for creating match matrices with the current run in order to accomplish the prediction of feature evolution in the current run. Nevertheless, this problem is outside the scope of this paper.