Abstract
In modern industries, smart maintenance approaches have become crucial for optimized and uninterrupted production processes. The traditional maintenance techniques are less effective to fulfill the competitive requirements of the advanced manufacturing systems. Conventional corrective maintenance increases the cost of products or services due to the disruption in production process. Similarly, time-based preventive maintenance escalates the cost of maintenance due to unnecessary repairing actions. In the conducted research, (i) advanced Remaining Useful Lives (RULs) prediction methods subject to industrial slurry pumps, have been developed to facilitate Condition Based Maintenance (CBM). Similarly, (ii) an improved method for predicting Corrosion Under Insulation (CUI) in pipes has been established to enrich the predictive maintenance.In the literature, many studies are available where researchers have developed RUL prediction models with an ideal database. Those databases were equipped with a huge amount of “run to failure” and “run to prior failure” datasets. However, in the real world, run to failure data for an in-operation industrial machine is difficult to exist since machines are never allowed to work until their failures. In such a situation, the available option is to utilize the “run to prior failure” data of a particular machine for estimating the RUL.
In the first part of the conducted research, the online RULs of industrial slurry pumps have been predicted. The novel aspect of this prediction was its estimation with only available “run to prior failure data”. The RULs were predicted with two different deep learning-based prognostics methods. In the first proposed prognostic methodology, the vibration signals obtained from slurry pumps were cleaned manually. Afterward, the cleaned signals were utilized for generating performance degradation trends using Fast Fourier Transform (FFT) technique. Then a hybrid nonlinear autoregressive (NAR)-LSTM-BiLSTM model was developed for predicting the online RULs. The studies available in literature have utilized statistical methods for only predicting the overall RULs of slurry pumps. Furthermore, their degradation trends were progressing towards the threshold line in a very calm manner. The presented work is the first study that has engaged a combination of traditional and deep learning neural networks for predicting not only the overall RULs but also the continuous RULs. The developed NAR-LSTM model was delivering satisfactory results either the degradation trends were progressing in a smooth or complicated way. The proposed methodology was also applied to publicly available NASA’s C-MAPSS dataset for validating its applicability, and in return, acceptable results were achieved.
In the second proposed prognostic method, the available “run to prior failure” vibration datasets of a slurry pump were filtered automatically by using a novel criterion. Then the filtered datasets were utilized for constructing the Health Degradation Trends (HDTs) using the principal component analysis and a moving average method. Subsequently, a hybrid deep LSTM model embedded with a smart learning rate mechanism was developed to utilize the formulated HDTs for estimating the online RULs. As per literature, it was a research gap that the prediction process should start from which point of HDTs. This study suggested that the prediction process should start from a point with an increasing slopes trend. The available models in literature are predicting the online RULs of slurry pumps via utilizing curve fitting methods. The results obtained by the developed hybrid deep LSTM model were also compared with curve fitting and machine learning methods.
In the second part of the study, a predictive maintenance technique was designed to predict the CUI in pipes. This work was based on American Petroleum Institute (API 581) Risk-Based Inspection manual and CUI corrosion rate field data. The utilized field data were obtained from a local gas processing plant in Malaysia. Due to the varying nature and scarcity of CUI data, a fuzzy logic-based CUI corrosion rate prediction model was developed in the conducted research. The fuzzy logic model available in literature has predicted CUI corrosion rates for only two input parameters, i.e., “operating temperature” and “type of environment”. There was a need to improve the existing model. Therefore, an advanced fuzzy logic model with three additional input parameters, i.e., insulation type, pipe complexity, and insulation condition, was constructed in this study. Results generated by the developed fuzzy logic model were found to be significant when they were validated using the Kruskal-Wallis test. After the development of fuzzy logic model, sensitivity analysis (SA) of all the input parameters was performed to determine the most prominent input parameter for the cause of CUI. At the end of the study, CUI 3D surfaces were modeled, which were showing the relationship between any one/two input parameters with CUI corrosion rate at an instant. The developed fuzzy logic model performed CUI assessment on an impressive scale and produced comparable results to the industrial standards.
| Date of Award | 8 Dec 2022 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Wai Tat Peter TSE (Supervisor) & Inez ZWETSLOOT (Supervisor) |