Machine and component residual life estimation through the application of neural networks

Abstract

Analysis of reliability data plays an important role in the maintenance decision making process. The accurate estimation of residual life in components and systems can be a great asset when planning the preventive replacement of components on machines. Artificial intelligence is a field that has rapidly developed over the last twenty years and practical applications have been found in many diverse areas. The use of such methods in the maintenance field have however not yet been fully explored. With the common availability of condition monitoring data, another dimension has been added to the analysis of reliability data. Neural networks allow for explanatory variables to be incorporated into the analysis process. This is expected to improve the quality of predictions when compared to the results achieved through the use of methods that rely solely on failure time data. Neural networks can therefore be seen as an alternative to the various regression models, such as the proportional hazards model, which also incorporate such covariates into the analysis. For the purpose of investigating their applicability to the problem of predicting the residual life of machines and components, neural networks were trained and tested with the data of two different reliability related datasets. The first dataset represents the renewal case where repair leads to complete restoration of the system. A typical maintenance situation was simulated in the laboratory by subjecting a series of similar test pieces to different loading conditions. Measurements were taken at regular intervals during testing with a number of sensors which provided an indication of the test piece’s condition at the time of measurement. The dataset was split into a training set and a test set and a number of neural network variations were trained using the first set. The networks’ ability to generalize was then tested by presenting the data from the test set to each of these networks. The second dataset contained data collected from a group of pumps working in a coal mining environment. This dataset therefore represented an example of the situation encountered with a repaired system. The performance of different neural network variations was subsequently compared through the use of cross-validation. It was proved that in most cases the use of condition monitoring data as network inputs improved the accuracy of the neural networks’ predictions. The average prediction error of the various neural networks under comparison varied between 431 and 841 seconds on the renewal dataset, where test pieces had a characteristic life of 8971 seconds. When optimized the multi-layer perceptron neural networks trained with the Levenberg-Marquardt algorithm and the general regression neural network produced a sum of squares error within 11.1% of each other for the data of the repaired system. This result emphasizes the importance of adjusting parameters, network architecture and training targets for optimal performance The advantage of using neural networks for predicting residual life was clearly illustrated when comparing their performance to the results achieved through the use of the traditional statistical methods. The potential of using neural networks for residual life prediction was therefore illustrated in both cases.