Deep PHM: IoT-Based Deep Learning Approach on Prediction of Prognostics and Health Management of an Aircraft Engine
Every industry requires efficient management of assets, especially for systems where failure and downtime are costly. A low-cost combined health monitoring system has become increasingly important, especially in drive and capital-intensive industries, where the aforementioned economic losses due to lost time and failure are significant. A Prognostics and Health Management (PHM) framework that effectively detects faults and failures can help scheduling proactive maintenance actions that avoid catastrophic failures, moreover reducing overall maintenance cost. In aerospace industry, implementation of a PHM system is quite challenging. The system architecture should be secure, reliable, and simultaneously capable of collecting, organizing, and analyzing substantial data from numerous diverse assets. Further complication lies in capturing information and processing data from assets that are co-located and/or geographically distributed over wide area. Such complication is one among the reasons why PHM has not been used in the past on individual aerospace assets. This resulted in systems having limited or no PHM capabilities requiring a statistical PHM approach that is only fleet-based. However, an advanced PHM system capable of analyzing individual assets will be an attractive option in view of reducing large lifecycle costs, understanding the current state of every asset, and ultimately assessing the overall readiness of the fleet. This provokes to search into other comparative industries for technologies that could help overcome the difficulty in aerospace industry. Predictive maintenance is one that predicts the remaining lifecycle of an equipment and also could discern trends that indicate the upcoming failure. This enables repairs to be done well before the onset of major issues. The difficulty here is in analyzing the available data. The technicians must know and understand the previous issues that happened with the equipment. All views, currently, have started turning toward the Internet of Things (IoT) which could collect data and foretell the data patterns which indicate first-stage equipment failure. Following are the different ways IoT can be utilized for improving predictive maintenance for Maintenance, Repair and Overhaul (MRO) tasks. OEM parts companies, airline operators, and technicians can utilize the prognostic and diagnostic strategies to find potential symptoms of aircraft failure. Data for analyzing symptoms of aircraft failure can be effectively gathered by placing sensors on significant areas such as engines, landing gear, wing flaps, and bleed valves. The possible failures can be noted down and predictive maintenance schedules can be created to order parts and schedule relevant workers for timely fixing of the equipment. Commonly used machine learning models of predictive maintenance are based upon feature engineering. Feature engineering is using domain expertise to construct right features manually. However, versatility of the models through this method is questionable since feature engineering is problem specific and the data available varies for every business. In this scenario, applying deep learning into predictive maintenance becomes attractive, as deep learning networks can analyze the data extracts the right features automatically. Hence, the need for manual featuring engineering is eliminated. Deep learning is now demonstrating outstanding results in disciplines such as picture classification and object recognition. Moreover, it is gaining reputation in fields where time series data plays a significant role. Predictive maintenance monitors the state of an asset and predicts failures through susceptive data patterns, which in turn found analyzing data collected over a period of time. Hence, it is obvious that deep learning has its application in predictive maintenance. Because of its higher ability to learn from sequences, the Long Short-Term Memory (LSTM) network, one of the deep learning approaches, is more appealing for predictive maintenance. This enables it to look back into the time series data collected for long periods to detect patterns corresponding to failures. An example to prove the applicability of LSTM network is also shown. In this example, an LSTM network is built using a publicly available turbofan engine dataset (C-MAPSS dataset) for predicting the Remaining Useful Life (RUL) of aircraft engines. Simulated aircraft sensor values are used in this LSTM network to predict the time of engine failure, which could help advance planning of maintenance works. Basic neural network (NN), deep neural network (DNN), recurrent neural network (RNN), and deep convolution neural network(DCNN) are some of the other deep learning networks networks. . Prognostic analysis using these network architectures to estimate RUL were carried out on the equivalent C-MAPSS dataset of a turbofan engine unit. By having a high predicted accuracy on RUL estimation, LSTM network, on the other hand, established its superiority. This is established by associating the findings to the state of the art on the equivalent dataset. As a result, the LSTM network emerges as a fresh and promising tool for providing systematic and accurate prognostics. As in many fields, IoT and deep learning methods opened up chances for PHM to be implemented to all kinds of assets distributed in all sectors, hence introducing new opportunities in aviation business. In this chapter, the reader will be introduced to the basic concepts of PHM and the prospective role of IoT and deep learning networks in expanding new opportunities.
Read more