Blog post

How to Boost Quality Control and Transform the Automotive Domain Using the Power of ML and AI

Renault Senior Data Scientist Mohamed-Achref Maiza shed the light at how Renault takes the digital transformation journey adopting artificial intelligence in automotive

December 17, 2019

5 mins read

Artificial intelligence (AI) and machine learning (ML) are taking automotive businesses to the next level of efficiency. At IT Arena 2019, we had a chance to meet Mohamed-Achref Maiza, a senior data scientist at Renault. Responsible for developing advanced analytics solutions for a world-class automaker, Maiza knows the ins and outs of integrating AI and ML in the automotive industry as well as how to combine quality control and AI.

How can AI and ML technologies assist digital transformation in the automotive industry?

Mohamed-Achref Maiza: The first thing that AI- and ML-powered digital transformation in the automotive industry should aim at is to generate savings through process optimization. At the same time, artificial intelligence can be really helpful when it comes to testing new features of a vehicle. For example, using ML, automakers can analyze automotive big data and see what features are being used correctly and where the bottlenecks are. This approach will help them understand better what they should focus on to increase their competitiveness and to provide a better product.

In addition, the usage of artificial intelligence in automotive can help automakers perform more efficiently with a cost-effective fleet management platform by monitoring fleet data and allocating vehicles accordingly.

How can automakers combine quality control and AI in the automotive industry?

Traditional machine vision systems are robust and fast when it comes to analyzing structure schemes. However, problems begin with abnormality detection during the production process. You can’t predict how this or that defect will look. So in this respect, automakers still rely on human intelligence. It is not efficient, since a human being can’t analyze all the variations in the car data. On the contrary, deep learning technology mimics human intelligence while analyzing all the variations of the data automatically. Thus, it is like a good combination of human-like flexibility and the speed and robustness of computerized systems.

But since defects are pretty rare, the question is how we can we combine quality control and AI and train these ML models on small data sets. Here’s where few-shot learning comes into play. Few-shot learning algorithms have proved to increase the efficiency of machine learning on small data sets. Furthermore, they have great potential for solving manufacturing problems in the future.

What are the challenges of mixing ML and automotive safety?

I think automakers should focus on safety in every phase of the production cycle. For instance, when it comes to validation testing, testing plans need to be optimized in terms of time and volume. We need to be able to run a larger and larger number of scenarios and generate more validation results in a timely manner.

Furthermore, you can use DL-driven [deep learning-driven] technologies to enhance visual inspection and thus be able to check 100% of your production, discovering hidden patterns in data.

Neural networks have certain shortcomings. For example, if a model is trained on small objects, it won’t work on larger ones. Do you see any solutions to this?

When talking about ML and automotive safety, the question is how we can shrink ML models in order to optimize their performance. In this respect, there are hardware and software solutions. On the hardware side, there is Google Coral Edge TPU, which is designed by Google to accelerate metrics, multiplications, and beyond.

There is also another solution called the Dev Board, which has been designed by Google as well. Another product that automakers should adopt is Movidius. Created by Intel, Movidius is a new-gen microprocessor designed for vision applications. The good thing about Movidius is that it is powered by TensorFlow Lite, which is a deep learning framework that makes it possible to use existing TensorFlow models. So instead of building a new TensorFlow model from scratch, you can convert it to a more compressed model and deploy it on mobile, embedded, and IoT devices.

As for the “purely software” techniques, I can’t but mention the 1×1 convolution method. It can shrink the number of depth channels in the neural network, making the volume much smaller without losing a lot of efficiency. It has proved successful in many DL algorithms, like Inception and so-called Google Nets.

Can you suggest new approaches to the development and use of new evaluation metrics, including automation aspects?

There are a bunch of evaluation metrics that can be used to measure the robustness of machine learning algorithms. The reason why I deal with these evaluation metrics is always to create a connection with business impact.

Each machine learning problem is different, and you have to adapt to it. For example, if you are using a metric to measure the cost of the false positive or negative error rate of your model, you can try to convert it to another metric that reflects the financial gain or financial loss. I believe that such an approach is very powerful in terms of communication with teams or when you are trying to convince them to use ML models. I think every problem is different, and when it comes to the evaluation of machine learning models, we should think about the real value of AI.

Which of today’s machine learning problems and challenges the industry is not ready to solve?

Probably it is car styling. ML and car safety is more or less clear. But ML models — specifically DL ones — have been successful at generating art images, using so-called neural style transfer. However, we doubt if it will work correctly for generating a new style for a car that will trigger the emotion of the customer, comply with safety and aerodynamic requirements, and match the visual signature of a brand. So I think it will be hard to solve this kind of machine learning problems in automotive. We still need designers.

While traditional automotive businesses still depend a lot on human intelligence, Renault Digital reaps the benefits of combining quality control and AI as well as ML and automotive safety. As our interview confirms, the ability of machines to process information and react like human beings helps automotive companies handle a wide range of tasks. The experience of Renault demonstrates that the adoption of AI and ML allows for optimizing production processes, revolutionizing quality control, and increasing safety. And though ML-powered technologies still have much room for improvement, they’ve already turned the automotive market upside down.

Your subscription is confirmed.
Thank you for being with us.

No votes Thank you for your vote. 21330 4c0ba981fc

Tell us about your project

I give consent to the processing of my personal data given in the contact form above under the terms and conditions of Intellias Privacy Policy. I want to receive commercial communications and marketing information from Intellias by electronic means of communication (including telephone and e-mail).
* I give consent to the processing of my personal data given in the contact form above under the terms and conditions of Intellias Privacy Policy.

Awards and recognition

iso-27001 (1)
iso-2001-2015 (1)

Thank you for your message.
We will get back to you shortly.

Thank you for your message.
We will get back to you shortly.