Introduction
The recent advancements in connected vehicle technologies, including streaming massive amounts of vehicular data, and cloud computing capabilities are driving a paradigm shift in vehicle health management for the automotive industry. Insights from vehicle data can enable Automotive OEMs and their partners to understand vehicle usage pattern, improve product quality, and provide better customer service etc.
Real-time estimation of remaining useful life of vital automotive components is one of the areas drawing interest from both customers and OEMs alike. Traditionally these algorithms are deployed in ECUs. However, ECUs are bound by limited computational power and are not best placed to execute these algorithms. In addition, processing of diverse data sets pose certain other challenges, viz:
- Handling noisy sensor data
- Unavailability of labelled data
- Scalability and reusability of algorithms across components
Deep neural networks have proven to be effective for predictive analytics due to their ability to understand nonlinear patterns in large & unlabeled data sets. These techniques can leverage a wide range of parameters such as sensor data, CAN data, driver behavior, route profile etc., to derive vehicle ageing. These algorithms can be deployed in cloud and scaled easily to process data from few vehicles to large fleets.
This article focuses on the implementation of deep learning methods for Remaining Useful Life (RUL) prediction for vital components.
Need for RUL algorithms
Remaining Useful Life (RUL) prediction is an essential component of Integrated Vehicle Health Management (IVHM). RUL defines the fault progression in a component with the help of relevant degradation parameters. This prediction enables OEMs to either take preventive actions, such as arresting failure progression or corrective actions, such as replacing the component. This results in dual benefits, ensuring vehicle uptime, and simultaneously enabling a better understanding of the real time component performance, which can act as input for future product development.
Traditional Machine learning methods were utilized initially for RUL prediction, however, they failed to process the complex data acquired from sensors in a reliable way. The failure was primarily attributed to the need for domain specific feature engineering and large quantities of labelled data.
On the contrary, Deep learning has shown promising results in processing data with high variations and non-linearity with minimal human input [1]. It has become an active field for applications such as Integrated Vehicle Health Management and prognostics. The increase in GPU power and efficient data collection mechanisms have enabled deep learning to improve prediction results at affordable costs.
In this article, we explored the application of Feed Forward Neural networks for RUL prediction of component degradation. Using this method, we were able to predict RUL without domain know-how pertinent to component degradation characteristics. This implies that the algorithm is trainable without domain specific feature engineering requirements, thus saving efforts in understanding the correlation and causation between parameters (referred to as unsupervised learning).
Typical challenges in data collection
We need to collect and process vehicle performance data, route information and usage patterns for predicting RUL. These data sets are highly jittery in nature and create unnecessary complexities if left untreated.
There are some standard methods such as Fast Fourier Transform (FFT) or filters to smoothen the noise. However, these methods have the following shortcomings.
- Thresholding: Requires domain or statistical know-how for appropriate threshold definition and noise filtration
- Averaging: Often produces averaged values as output, which aren’t real values
The above challenges can be tackled by using de-noising autoencoders. An autoencoder is an unsupervised Neural Network which tries to learn an approximation function by replicating the original signal [2].
De-noising autoencoder architecture
Other major challenges around data collection include:
- Difficulty in obtaining high quality labeled training data for failure scenarios
- Diverse failure modes and operating conditions that increase the degradation complexity
Countering these challenges | KPIT RUL prediction methodology
KPIT has used a novel prediction methodology, which combines elements of unsupervised and supervised learning models. Using this method, we are able to automatically extract complex degradation features with high accuracies, from relatively smaller amounts of data.
Pre-training with unsupervised learning enables noise elimination. It further codifies complex non-linear relationships between the input and output parameters, to enable faster and standardized data labeling.
This pre-trained data is then fed to a supervised learning algorithm for estimation and prediction of RUL values.
Significant reduction in model development time, is one of the biggest advantages of this combined approach.
Methodology | Salient features
Domain encapsulation
Pre-training deep learning and Bayesian network models with engineering data and associated artefacts, enables domain encapsulation, thereby eliminating the need for explicit feature engineering. The encapsulation identifies relationships between the parameters from the pre-trained data and engineering design formulae, thereby drastically decreasing the computing time for correlations.
Custom Neural Network
A proprietary Neural Network architecture with an 18-layer Feed Forward Neural network has been used for processing the data. Each feed comprises 100 Neurons per layer for auto-encoding. The activation function used is hyperbolic tangent (Tanh). This multi-layer neural network enables representation of complex non-linear relationships between the input and output parameters. In addition, it automatically populates the missing data points in the degradation characteristics of the automotive component in question.
Auto encoders
Since the sensor data is noisy and not smooth, there is a need to flatten the noise. Using deep learning based auto-encoders, noise in the data can be eliminated [4]. Auto-encoders have the ability to represent the superset, with a subset, having reduced data. This enables elimination of data outliers that do not contribute to the variance of the model. An additional benefit relates to the reduced amount of data required to train the model. It is found that the output of the auto-encoder i.e. the reconstruction data, has an error rate of 0.0015, which is extremely negligible.
Scalable System Architecture
KPIT has developed a scalable architecture for running machine learning algorithms on large datasets. The system can scan the data in real time and provide actionable insights in the form of alerts and system monitoring dashboards. New RUL algorithms for different auto components can be deployed easily with this architecture.
Predictive analytics system architecture
The architecture can also be integrated with OEM systems through Rest APIs. For example, remaining useful life data can be delivered to After-sales service tools for use by technicians in service dealerships.
What are its real-world benefits?
Real time assessment of vehicle component performance and remaining useful life modeling provides an integrated system for decision making in condition-based maintenance [3]. The evolution of Shared mobility and transportation marketplaces have resulted in a significant increase in fleet-based operations, both in the passenger and commercial vehicle segments.
Predicting Remaining useful life is one of the fast-growing areas in analytics. It enables OEMs to continuously monitor the overall health of the asset as well as the associated components.
The RUL prediction methodology from KPIT improves operational efficiency and reduces failure rates of components and vehicle subsystems. OEMs can generate value for customers by increasing the reliability and durability of their products.
Additionally, they can help fleet owners in optimizing their fleet operating costs through real time monitoring of vehicle health and performance. Fleet operators can also pre-plan their maintenance schedules and avoid unexpected breakdowns and unplanned downtimes, thus improving the Total Lifetime Value (TLV) of their assets.
Conclusion
The approach defined in this paper promises best-in-class RUL prediction efficiency with reduced amounts of labeled training data.
The stated methodology, using a combination of unsupervised and supervised learning techniques can deal with multiple failure modes and operating conditions, and helps improve understanding of component degradation phenomena.
Given the challenges faced in collecting and processing large amounts of data, the method using unsupervised learning techniques combined with domain encapsulation offers a highly relevant and an accurate solution for real world prognostics applications.
References
- Y. Hu, S. Liu, H. Lu and H. Zhang, “Remaining Useful Life Assessment and its Application in the Decision for Remanufacturing”, 21st CIRP Conference on Life Cycle Engineering, vol. 15, pp. 212-217, 2014. Available: 10.1016/j.procir.2014.06.052 [Accessed 9 December 2019].
- R. Kewalramani and R. Kumar, “Estimation of Remaining Useful Life of Electric motor using supervised deep learning methods”, International Transportation Electrification conference, vol. 2, pp. 110-121, 2019. [Accessed 9 December 2019].
- X. Si, Z. Zhang and C. Hu, Data-Driven Remaining Useful Life Prognosis Techniques, 1st ed. Beijing, China: Springer, 2017, pp. 421-445.
- C. Hu, X. Si, W. Wang and D. Zhou, “Remaining useful life estimation – A review on the statistical data driven approaches”, European Journal of Operational Research, vol. 213, no. 1, pp. 1-14, 2011. Available: https://www.sciencedirect.com/science/article/abs/pii/S0377221710007903 [Accessed 9
December 2019].