Recurrent and Deep Learning for Machine Prognostics

Title Recurrent and Deep Learning for Machine Prognostics
Summary Construct and optimise Recurrent Neural Networks for industrial applications on machine prognostics; Augmenting industrial data for supervised learning
Keywords Recurrent Neural Networks, time series forecasting, supervised learning, Prognostics and Health Management, fault detection
TimeFrame Fall 2017
References [[References::[1] Liu, Jie, et al. An adaptive recurrent neural network for remaining useful life prediction of lithium-ion batteries. NATIONAL AERONAUTICS AND SPACE ADMINISTRATION MOFFETT FIELD CA AMES RESEARCH CENTER, 2010.

[2] Malhi, Arnaz, Ruqiang Yan, and Robert X. Gao. "Prognosis of defect propagation based on recurrent neural networks." IEEE Transactions on Instrumentation and Measurement 60.3 (2011): 703-711.

[3] Heimes, Felix O. "Recurrent neural networks for remaining useful life estimation." Prognostics and Health Management, 2008. PHM 2008. International Conference on. IEEE, 2008.

[4] Marco Rigamonti et al., "Echo State Network for the Remaining Useful Life Prediction of a Turbofan Engine." Third European Conference of the Prognostics and Health Management Society 2016, Bilbao, Spain, 5-8 July, 2016. PHM Society, 2016.

[5] Mandic, Danilo P., and Jonathon A. Chambers. Recurrent neural networks for prediction: learning algorithms, architectures and stability. New York: John Wiley, 2001.

[6] Jaeger, Herbert. Tutorial on training recurrent neural networks, covering BPPT, RTRL, EKF and the" echo state network" approach. Vol. 5. GMD-Forschungszentrum Informationstechnik, 2002.

[7] Sun, Jianzhong, et al. "Application of a state space modeling technique to system prognostics based on a health index for condition-based maintenance." Mechanical Systems and Signal Processing 28 (2012): 585-596.

[8] Wang, Wilson Q., M. Farid Golnaraghi, and Fathy Ismail. "Prognosis of machine health condition using neuro-fuzzy systems." Mechanical Systems and Signal Processing 18.4 (2004): 813-831.]]

Prerequisites Artificial Intelligence and Learning Systems courses; good knowledge in machine learning and neural networks; programming skills for implementing machine learning algorithms
Author Kunru Chen
Supervisor Sławomir Nowaczyk, Sepideh Pashami, Yuantao Fan
Level Master
Status Finished


The current paradigm for maintaining industrial equipments is a combination of reactive and preventive actions. Take commercial transportation vehicles as example, they are typically maintained after an equipment failure occurs or according to preplanned visits to the workshops based on mileage or calendar time. This mixture of maintenance strategy is not ideal: i ) it does not perform maintenance pro-actively well before the failure happens, i.e. severe component failures usually result in extra damage to the system and could be prevented; ii ) planned maintenance with fixed time intervals does not guarantee all routinely changed parts have used all their potentials. Therefore, a shift of current maintenance strategy to one with more predictive maintenance is required: to inspect and repair components (well) before they causes a breakdown or severe damage to the system.

Nowadays, with the development of electronic devices and the emergence of Internet of Things, huge amount of sensor data collected and transmitted remotely can be utilized for equipment monitoring, fault detection and prognostics. By processing sensor data during operations, condition of the equipment will be accessed and maintenance decision will be made. A common approach for machine prognostics is to estimate Remaining Useful Life (RUL) of the system. Many researchers applied Recurrent Neural Networks (RNNs) for estimating RUL based on sensor measurements, e.g. see reference [1,2,3,4]. Another common approach for prognostics is to generate index that reflects health status of equipment, e.g. reference [7,8], based on it's condition.

A recurrent neural network (RNN) is a type of artificial neural network where connections between units form a directed cycle. It can use their internal memory to process arbitrary sequences of inputs and capture dynamic temporal behaviour, tutorials can be found in [5,6].

In this project, the student will construct RNNs for predicting RUL of machines in different domains, including simulated data and real data from industrial application. The industrial data collected from large amount of vehicles performing transportation tasks in the field. Vehicle’s configuration, aggregated sensor values collected at different time are available. Service record that contains assessments and repair actions is provided. Based on sensor data and service records, a machine learning method for predicting RUL of equipment is expected to be proposed and evaluated.

This project is programming oriented, the student will be working with some of the libraries that includes RNNs implementations, e.g. Theano and Keras.


1. Construct Recurrent Neural Networks, optimise it's architecture and training algorithms for predicting Remaining Useful Life of equipment.

2. Investigate and propose different scenarios augmenting industrial data for machine prognostics, e.g. generating targets/teaching signals for RNNs to learn?

3. Investigate what type of data representation technique can be employed for this application, e.g. histograms of aggregated values as input and use CNN-RNN model for the networks.

Research Questions:

1. How to construct and train RNNs for predicting RUL?

2. How to augment industrial data for this study? - What type of representations can be used? - How to generate targets/teaching sequences based on service records? Does imposing arbitary conditions based on application improves prediction performance?

3. How to augment the network architecture for different types of data representation (e.g. histograms with aggregated values as input)?