Engineering design and operational decisions depend largely on engineers’ understanding of applications. This includes assumptions made to simplify problems in order to solve them. However, these assumptions often times introduce errors compared to the actual behavior of an application. Increasing access to sensor or virtual data and computational resources, combined with democratization of advanced Machine Learning (ML) algorithms leads to greater use of ML. This brings field data and engineering knowledge together allowing for an increased level of overall accuracy in decision-making and design performance improvement.
ML was first defined by computer scientist Arthur Samuel as “a field of study that gives computers the ability to learn without being explicitly programmed”. It differs from traditional rule-based programming, as its performance is a function of the data it learns from, as opposed to traditional programming where a set of rules governs the logic of the application. As a result, predictive models from ML continuously improve with data, as opposed to traditional rule-based programming that requires explicit updates by users.
ML can use field or simulation data and it can be of two types; supervised or unsupervised. Let’s look at two applications of ML for automotive industry. In the first application, Altair Multi-Disciplinary Design Optimization Director (MDOD), uses simulation data for supervised learning. In the second application, anomaly detection in bearings uses sensor data for unsupervised learning.
Multi-disciplinary Design Optimization (MDO) has become increasingly important to meet demanding requirements for product performance and time-to-market. Those that have worked on MDO applications know that setting up and solving MDO problems can be labor intensive and computationally expensive, especially if it is a large-scale application such as a full automotive Body-in-White (BIW). MDOD is an environment developed to break these barriers for utilizing MDO in engineering applications. In addition to its intuitive user interface, it employs ML to give rapid feedback to the design teams. For this, it uses sub-response surfaces to decouple the problem and reduce the amount of data required to accurately predict the performance. Data is gathered using efficient and extensible sampling methods that are conducive to ML. ML is also employed during the optimization step as part of its global search method to reduce the required data size.
The use of sensors for diagnostics and prognostics is increasing as IoT use grows. Bearings are critical components in automotive industry. In this example, we have sensor data for 4 bearings sampled at a rate of 20kHz resulting in sampling every 10 minutes for 1 second for 9 days. This adds to a total of 20 million records. The first sampling corresponds to the new bearing and is used as a reference for anomaly detection as it ages. In this project, the objective is to recognize these anomalies as soon as they occur before they lead to irreversible issues, such as parts failures. In the ML process, first Principal Component Analysis (PCA) is used for feature recognition. The samples are then correlated to the healthy sample. Finally, the anomaly is detected using a threshold for correlation drop and value.
There are many more automotive applications of ML such as nonlinear concept design. More information on Altair’s market-leading Multiphysics simulation and optimization tools along with its IoT analytics offering can be found on http://iot.altair.com/
This article was originally written for and published in the July 2018 issue of Ingenieurs de l’Auto by SIA.