Estimation of remaining useful life with model based predictive maintenance

Asynchronous motors are widely used in factories due to their several advantages. The healthy continuation of production in enterprises depends on the proper functioning of the devices. Like all equipment, asynchronous motors tend to age over time and various malfunctions may occur. It is critical for businesses to predict these failures in order to prevent production downtime and loss of time and cost. In order to decrease operational expenses, increase operational reliability and end user satisfaction, continuous monitoring of engine operation and health status and detection of developing malfunctions before they become more costly and destructive has become an increasingly important and developing technology. Motor faults can be basically divided into two categories, electrical and mechanical faults. In different studies, it has been shown that bearing failures take the first place among motor failures. Therefore, bearings are among the most critical mechanical components and have wide applications in modern industrial rotating machinery. Bearings function to reduce friction between mechanical components or limit movement in the preferred direction. They generally function in harsh operating environments such as overload, high speed and insufficient lubrication. After a certain period under these circumstances, clearance, frictional force, overheating, etc. performance decrease and failures may develop due to reasons. Failure to detect these faults in a timely manner and to perform proper maintenance may cause to failure. So, it is necessary to constantly monitor some technical data (mechanical vibration, temperature, electrical current, oil analysis, etc.) on the equipment and to apply condition-based maintenance scenarios. Condition-based maintenance scenarios basically consist of two parts, prognostic and diagnostic. Prognostics is concerned with estimating when the fault will occur from the moment the deterioration first started, while diagnostics is concerned with the type and cause of the fault after it has occurred. In this sense, prognostic applications are much more effective in terms of no loss and efficiency. The methods used for the analysis of condition monitoring data are basically classified into three categories. These are data based, model based and hybrid approaches that can be considered a combination of these two. The purpose of predictive maintenance is to predict the failure that will occur, with a high accuracy and sufficiently in advance, when the failure may occur after processing the physical data received from the machinery or equipment via sensors. Predictive maintenance describes a number of techniques for accurately monitoring the health status of any equipment. These techniques, which can work locally or cloud-based, include various data analysis and machine learning algorithms, provide a cost advantage over other maintenance methods such as reactive or periodic maintenance. In this study, an algorithm has been developed to predict when a failure will occur in the bearings by analyzing the vibration signal measured from the identical bearings placed on a shaft connected to an induction motor. The data set originally belongs to Center for Intelligent Maintenance Systems (IMS), University of Cincinnati, and shared openly for the use of researchers on the NASA-Prognostics Center of Excellence website. During the tests, the natural aging and failure processes of the bearings were examined by rotating the shaft under constant speed and load without applying any additional processes that would cause failure or accelerate the occurrence of failure. Four bearings with the same technical specifications on the shaft started the test together and the test was terminated when any of them failed. Three rounds of testing were carried out in this way with the renewed bearings, and three datasets were shared, each one eventually resulting in failure. The information on which bearing and what failure occurred was also shared by the data owners. Since vibration data naturally contains noise, it was first subjected to noise reduction. Then, discrete wavelet transform has been applied to obtain some features which are characterizing the original vibration signal which is non-stationary. After this process, various frequency and time domain features are extracted and a feature table was created. A certain part of the data in the table is accepted as training data, and a certain number of features were selected according to the monotonicity levels over these data. As the next data processing step, principal components analysis (PCA) was applied to the selected features and a single health indicator was obtained that indicates the aging-related health status of the bearing. Some data processing steps such as smoothing and finding the TSP (time to start prediction) point were applied on the health indicator data, and the final health indicator data was applied to exponential degradation model. The model run with the default initial parameters gives the estimated remaining useful life information obtained when each test data point arrives. The developed prognostic model was applied to the first and second data sets, and α-λ graph and CRA (cumulative relative accuracy) metrics were used for performance evaluation. The CRA value was calculated as 0.74 for the first data set and 0.85 for the second data set, and it was seen that these performance values were sufficient for this application and the remaining useful life of the bearings could be estimated successfully. In addition, the presence of faults were shown by performing a diagnostic study and using the envelope analysis technique. The algorithm applied in the study and the model obtained can be used for other equipment with similar data structures. The accuracy of the algorithm may be tested and increased by iteratively changing the methods applied in the data processing steps, selected parameters and some assumptions.