Industrial CT scan optimization for EV's
Future Software Technologies
Semester programme:Complex Software Systems
Client company:INNER
Project group members:William Terterian
Weiming Wu
Tom Peereboom
Dick van de Meulenhof
Rodi Janssen
Duy Nguyen
Project description
This project by INNER is all about making CT scan reconstruction faster and smarter. Using ASTRA Toolbox on NVIDIA GPUs, we reconstruct industrial cone-beam CT sinograms into full 3D volumes. Benchmarks are then performed on various machines with different computer components and compared to see how the reconstructions performs on those systems, specifically the GPU.
Context
INNER is a Dutch tech company that specializes in diagnostic solutions for EV battery packs. They use CT scanning technology combined with machine learning to detect internal defects like torn wires, swollen cells, cracks, and leakage in fully assembled batteries, without opening them or running lengthy electrical tests. They can spot issues down to 100 microns of precision.
Their main products are an integrated X-Ray/CT scanner for manufacturers and a battery diagnostics service currently in beta with Fraunhofer EZRT. They work with OEMs and battery manufacturers to help them catch defects early during production, reducing recalls, warranty costs, and safety risks.
Results
It started with a simple question: can we make CT reconstruction faster and more reliable without throwing expensive hardware at the problem?
The main thing that came out of it is a full cone-beam CT pipeline that takes raw sinogram data, auto-crops it, reconstructs it on the GPU using SIRT3D, and spits out clean uint16 TIFFs with a JSON log of everything that happened. It's been run on different systems and holds up consistently, putting it solidly at TRL 5-6, meaning it works in a real environment.
The most surprising and useful finding came from the downsampling tests. A 4x spatial downsample gives you a 4.5 to 4.9x speedup, dropping reconstruction from about 5.8 seconds down to under 2 seconds. That's a meaningful result because it means older or cheaper GPUs can stay in the game if you're willing to trade a bit of resolution.
The benchmarking framework is probably the most reusable piece of the whole project. It quiets down the OS, runs 50 iterations, and tracks CPU, RAM, VRAM, disk I/O and PCIe throughput in a way that's actually reproducible. That puts it at TRL 6-7.
We did not get non-NVIDIA GPU's to work with any of our toolkits. This has to do with the usage of CUDA in the algorithms. The available work arounds like ZLUDA did not function correctly. That failure is a TRL 3 result but it saves anyone downstream from wasting time on it.
About the project group
We are primarily software engineering students with diverse experience spanning artificial intelligence, game design, cybersecurity, smart industry, and IoT technologies.
We worked with an agile methodology, working in sprints of three weeks. The project duration has been the whole of semester 6. During this period, we dedicated 4 and a half days a week to it. After roughly a month, we split the group up into 2 mini-groups, consisting of three people. One was dedicated to benchmarking, and the other was focussed on researching and optimizing the reconstruction algorithms.