Abstract
Predictive maintenance has attracted many researchers with the increased growth in the digitization of industrial, locomotive, and aviation fields. Simultaneously, extensive research in deep learning model development to its deployment has made its way to industrial applications with unprecedented accuracy. The most crucial task in predictive maintenance is to predict the machine's remaining useful life, yet the most beneficial one. In this chapter, the authors address the problem of predicting the remaining lifecycle of an engine using its sensor data. The authors provide practical implementation of predicting the RUL of an engine by proposing a deep learning-based framework on the open-source benchmark NASA's Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) engine dataset, which contains sensor information of around 100 engines with 22 sensors. The proposed framework uses the bi-directional long short term memory algorithm. The authors optimize hyperparameters using advanced deep learning frameworks.
TopIntroduction
A forecast of equipment failure requirements aids the productions team in planning equipment maintenance. With such a maintenance facility, one can avoid performing unnecessary maintenance checks, called periodical maintenance, and save a lot of energy and time. In modern factories, a huge amount of big data is captured from the sensors attached to equipment and private cloud servers (Yasumoto et al., 2016). Such sensor data contains valuable information about the behavior of equipment during its normal and failed operation. A type of sensor data analysis is termed predictive analytics and is one of the most interesting data analytics problems. A propulsion system generates a thrust for an airplane to move through the air. Turbofan engines are used by modern airliners for such purposes. For aircraft maintenance, it is crucial to predict the turbofan engine’s remaining useful lifecycle (RUL). This chapter analyzes the available sensor data to train a deep learning-based algorithm for predicting rul. The dataset is from NASA turbofan jet engine sensor measurements. Predictive maintenance can be performed in a classification or regression way. It predicts the likelihood of failure in subsequent n-steps in its classification approach. Predicting the remaining time left before the subsequent failure is the regression approach. Some of the real-world issues solved using data-driven approaches for RUL prediction are 1) multifaceted chronological dependencies amongst sensors: numerous modules act together with each other in myriad ways leading to complex dependencies amongst sensor readings. For example, an alteration in one sensor may lead to a variation in another sensor after an interval of a few instants. 2) fractional absence of sensor data: some data may be moderately unattainable due to reasons such as loss of communication in the network and broken or defective sensors. 3) health degradation: it is challenging to build physics-based models for complex machines with numerous constituents. 4) noisy sensor readings: varying environmental noise levels affect the sensor readings. The quantity of noise varies across sensors. Learning-based algorithms are one of the growing areas that have a high impact in the predictive maintenance domain, and that is the focus of this chapter.
Key Terms in this Chapter
MinMaxScalar: It scales the feature value between 0 and 1.
Open Source Benchmark Dataset: Dataset available free to use for research purposes for benchmarking and validating the new proposed implementation and comparative analysis purposes.
API: Application Programming Interface is a list of rules and definitions for developing and integrating the application.
Encryption: It is the process of transforming the data into a secure format.
UNIX: An operating system widely used for internet servers and workstations.
Epoch: Number of times the deep learning model goes through the dataset.
Polymorphism: It is the concept that describes situations when there are multiple forms of the same object.