Data-driven Prognosis Research Articles

Fault Detection and Prognosis of Time Series Data with Random Projection Filter Bank

We introduce Random Projection Filter Bank (RPFB) as a general framework for feature extraction from time series data. RPFB is a set of randomly generated stable autoregressive filters that are convolved with the input time series. Filters in RPFB extract different aspects of the time series, and together they provide a reasonably good summary of the time series. These features can then be used by any conventional machine learning algorithm for solving tasks such as time series prediction, and fault detection and prognosis with time series data. RPFB is easy to implement, fast tocompute, and parallelizable. Through a series of experiments we show that RPFB alongside conventional machine learning algorithms can be effective in solving data-driven fault detection and prognosis problems.

Bearing Health Condition Prediction Using Deep Belief Network

Bearings play a critical role in maintaining safety and reliability of rotating machinery. Bearings health condition prediction aims to prevent unexpected failures and minimize overall maintenance costs since it provides decision making information for condition-based maintenance. This paper proposes a Deep Belief Network (DBN)-based data-driven health condition prediction method for bearings. In this prediction method, a DBN is used as the predictor, which includes stacked RBMs and regression output. Our main contributions include development of a deep leaning-based data-driven prognosis solution that does not rely on explicit model equations and prognostic expertise, and providing comprehensive prediction results on five representative run-to-failure bearings. The IEEE PHM 2012 challenge dataset is used to demonstrate the effectiveness of the proposed method, and the results are compared with two existing methods. The results show that the proposed method has promising performance in terms of short-term health condition prediction and remaining useful life prediction for bearings.

Data-Driven Based Battery Health Prognosis with Diagnosis Uncertainties and Insufficient Training Data Sets

This paper investigates data-driven based battery prognosis with diagnosis uncertainties and insufficient training data sets. Four types of data-driven prognosis methods are investigated including the neural network, similarity-based approach, relevance vector machine, and a recently developed copula-based approach. The remaining useful life (RUL) predictions of lithium-ion battery capacity are compared with capacity estimation error due to the fact that onboard lithium-ion battery capacity estimation is difficult and almost always contains estimation errors. Thus, robustness of each prognosis methods can be studied for real time capacity RUL estimation. Furthermore, collection of sufficient run-to-failure training data sets for lithium-ion batteries is almost impossible even though it is desirable for all data-driven based methods. Therefore, robustness of these methods in terms of the insufficient training data sets is also studied. These insightful results will helpdesigners choose appropriate prognosis algorithms in designing battery management systems (BMS) for lithium-ion batteries.

Intelligent Prognostic Framework for Degradation Assessment and Remaining Useful Life Estimation of Photovoltaic Module

All industrial systems and machines are subjected to degradation processes, which can be related to the operating conditions. This degradation can cause unwanted stops at any time and major maintenance work sometimes. The accurate prediction of the remaining useful life (RUL) is an important challenge in condition-based maintenance. Prognostic activity allows estimating the RUL before failure occurs and triggering actions to mitigate faults in time when needed. In this study, a new smart prognostic method for photovoltaic module health degradation was developed based on two approaches to achieve more accurate predictions: online diagnosis and data-driven prognosis. This framework of forecasting integrates the strengths of real-time monitoring in the first approach and relevant vector machine in the second. The results show that the proposed method is plausible due to its good prediction of RUL and can be effectively applied to many systems for monitoring and prognostics.

Bearing life prognosis based on monotonic feature selection and similarity modeling

In data-driven prognosis approach, indicator information plays an important role for reliable prediction. Although lots of researches have been carried out on prognosis algorithms, only few have paid attention on developing an effective method to select ‘good’ degradation indicators. This paper presents a novel strategy to address the problem, which mainly proposes methods of monotonic feature selection using rank mutual information, and similarity-based modeling for remaining life estimation. The proposed system is demonstrated based on open source data of bearing life cycle. The experiment results show that satisfactory prognostic performance can be obtained with advantages of simplicity, accuracy and generality.

Prognosis of Underground Cable via Online Data-Driven Method With Field Data

In order to prevent unexpected electrical outage and to save on repair expenses, an online data-driven prognosis method for monitored underground cable is proposed. The method is nondestructive and nonintrusive. The field data of voltage and current were collected from an underground distribution cable lateral installed in a residential area. Potential useful features were selected based on the data-driven method using the training data. The rationality of selected features was verified in view of cable aging mechanisms; after that, the useful features were finally determined. The remaining life of cable was forecasted by predicting the time for the cumulative effect of the selected features to reach the threshold. The performance of the developed prognosis method was tested and evaluated using the field testing data. The standard deviation of the prognosis results at different arrival times is used as a source of uncertainty estimation (reverse of confidence estimation). The results demonstrate that the method can continually offer valuable remaining life prediction of the monitored underground cable as time goes on. When the predicted fault times approach the actual fault time, the standard deviation value is small. When the standard deviation value is small, there are steady forecast results and usually small residuals.

Data-driven prognosis: a multi-physics approach verified via balloon burst experiment.

A multi-physics formulation for data-driven prognosis (DDP) is developed. Unlike traditional predictive strategies that require controlled offline measurements or 'training' for determination of constitutive parameters to derive the transitional statistics, the proposed DDP algorithm relies solely on in situ measurements. It uses a deterministic mechanics framework, but the stochastic nature of the solution arises naturally from the underlying assumptions regarding the order of the conservation potential as well as the number of dimensions involved. The proposed DDP scheme is capable of predicting onset of instabilities. Because the need for offline testing (or training) is obviated, it can be easily implemented for systems where such a priori testing is difficult or even impossible to conduct. The prognosis capability is demonstrated here via a balloon burst experiment where the instability is predicted using only online visual observations. The DDP scheme never failed to predict the incipient failure, and no false-positives were issued. The DDP algorithm is applicable to other types of datasets. Time horizons of DDP predictions can be adjusted by using memory over different time windows. Thus, a big dataset can be parsed in time to make a range of predictions over varying time horizons.

Prognosis of anterior cruciate ligament reconstruction: a data-driven approach.

Individuals who suffer anterior cruciate ligament (ACL) injury are at higher risk of developing knee osteoarthritis (OA) and almost 50% display symptoms 10-20 years post injury. Anterior cruciate ligament reconstruction (ACLR) often does not protect against knee OA development. Accordingly, a multi-scale formulation for data-driven prognosis (DDP) of post-ACLR is developed. Unlike traditional predictive strategies that require controlled off-line measurements or 'training' for determination of constitutive parameters to derive the transitional statistics, the proposed DDP algorithm relies solely on in situ measurements. The proposed DDP scheme is capable of predicting onset of instabilities. As the need for off-line testing (or training) is obviated, it can be easily implemented for ACLR, where such controlled a priori testing is almost impossible to conduct. The DDP algorithm facilitates hierarchical handling of the large dataset and can assess the state of recovery in post-ACLR conditions based on data collected from stair ascent and descent exercises of subjects. The DDP algorithm identifies inefficient knee varus motion and knee rotation as primary difficulties experienced by some of the post-ACLR population. In such cases, levels of energy dissipation rate at the knee, and its fluctuation may be used as measures for assessing progress after ACL reconstruction.

Comparison of binary classifiers for data-driven prognosis of jet engines health

A reliable prognosis is crucial to manage asset health and predict maintenance needs of large civil jet engines, which in turn contribute to enhanced aircraft airworthiness, longer time on wing and optimized lifecycle costs. With the accumulation of large amount of data over the last decade, one can relate the number of components serviced during a maintenance visit to the history of parameters inside and outside the engine (temperatures, pressure, shaft rotation speeds, vibration levels, etc.). While established statistical models had been developed for small samples, more recent computer-intensive statistical techniques from the field of Machine Learning (ML) can handle more complex datasets. In particular, binary classifiers constitute an attractive option to predict the probability of servicing the components of a given jet engine at the next maintenance visit. This paper demonstrates the validity of such data-driven methods on an industrial case study involving failures of thousands of compressor blades in aeronautical turbomachines. The prediction accuracy obtained with the ML techniques presents a significant improvement over the state-of-the-art. Moreover, the performance of six binary classifiers with different characteristics - logistic regression, support vector machines, classification trees, random forests, gradient boosted trees and neural networks - was compared according to four qualitative and quantitative criteria. Results show that there is no clear winner, although ensemble models based on trees (random forests and boosted trees) offer a good overall compromise while neural networks offer the best absolute performance. In the industrial world, the business objectives, the environment in which the models are deployed and the users’ skills should dictate the choice of the most adequate statistical technique.

A Survey of Health Indicators and Data-Driven Prognosis in Semiconductor Manufacturing Process

Semiconductor device fabrication is considered today as one of the most complicated manufacturing process, characterized by an important complexity of production context, in an uncertain environment. To improve process efficiency and productivity, it is of prime importance for engineers to dispose of reliable indicators to drive decision-making on maintenance operations. To this end, terabytes of different data are collected during the manufacturing process, to feed statistical analysis tools and production management systems. Prognostics and health management (PHM) is defined as the discipline that links studies of failure mechanisms to system lifecycle management. Among the different approaches existing for prognosis, data-driven techniques learn models directly from monitored operational data related to system health. There is therefore a great interest in applying data-driven PHM methodologies to address semiconductor manufacturing issues. This paper surveys works on data-driven approaches for two issues of PHM methodologies with applications focused on semi-conductor manufacturing process: the development of indicators for health assessment, and prognostic methods.