Skip to main content
Home  › ... Case Studies › Novosound

Case Study

Novosound + University of Strathclyde

To develop advanced machine learning techniques for wide area corrosion monitoring using thin film ultrasound sensors

Tomography image prediction on steel flat plate

Novosound

Novosound is a centre of expertise in Ultrasound. The core IP is a novel, flexible thin film piezoelectric material. 

Novosound Logo

The Challenge

What was the need?

The Challenge

Novosound’s Ceilidh platform continuously streams high-sample-rate ultrasonic data from four identical sparse arrays. Machine-learning models can exploit this large-scale of observations to learn non-linear wave dynamics directly from raw signals, extracting high-value insights and automating structural-health-monitoring decisions.

So far, the company’s deep-learning implementations achieve excellent accuracy but only for binary “defect” or “no-defect” classifications. That coarse flag may trigger manual follow-ups, yet it still leaves asset-integrity engineers uncertain about the location of corrosion, its growth rate and remaining wall thickness.

Our project therefore faced a dual challenge. First, we needed to invert sparse, highly multiplexed guided-wave data into a CT-style tomographic map that shows the footprint, location and remaining wall thickness of corrosion or erosion, richer information than any binary alarm so far. Second, we had to deliver that capability using just one of Ceilidh’s four arrays (a single transmitter and eight receivers) squeezing full imaging potential from limited coverage while mitigating mode conversions and noise that usually demand a full ring of transducers.

What did we do?

The Solution

Based on simulated data, we developed signal and image processing pipelines and trained neural network architectures capable of converting raw A-scans into a tomographic map predicting location, shape and depth of the material thickness profile. In the first stage, our baseline model employed a lightweight convolutional-based neural network architecture that delivers a coarse travel-time tomogram using only time of arrivals of certain ultrasonic guided wave modes. In the second stage, we employed transformer-based encoders with physics-constrained losses, which were trained using a large synthetic dataset generated from a digital twin model in OnScale, whose estimates were tuned and validated against laboratory experiments. This resulting full-waveform-inversion model refines the coarse map into a CT-like slice that pinpoints footprint, location and remaining wall thickness, meeting Ceilidh’s business need for actionable, visually intuitive corrosion monitoring capability. Close collaboration with Novosound’s engineers gave us privileged access to test rigs, allowing us to capture a carefully curated library of experimental A-scans that allowed extensive testing of the developed models.

What changed?

The Impacts and Benefits

Impacts for the Company

The project accelerates Novosound’s evolution from hardware-only sensor maker to a data-driven, service-centric business. By pairing Ceilidh’s sensors with a physics-informed tomography engine and a rich waveform-defect database, the company can now offer a competitive non-destructive-testing solution that includes real-time analytics dashboards for asset-condition monitoring. The new Python–OnScale toolchain lets R&D teams rapidly prototype bespoke arrays or validate customer use-cases, sharpening Novosound’s competitive edge. Finally, hands-on knowledge transfer in guided waves, 2-D wave propagation and latest deep learning technologies embeds cutting-edge expertise within the engineering team, strengthening Novosound’s reputation as a leader in smart ultrasonic monitoring.

Impacts for the Academic Team

Two journal manuscripts, one on physics-informed tomography, the other on sparse-array full-waveform inversion are now ready for submission to leading NDT journals, and two peer-reviewed conference papers have already been accepted, broadening the group’s international profile. Sharing these results with the ultrasonics community is expected to trigger collaborations on guided-wave inversion, signal processing and AI-driven asset monitoring, particularly with specialists across Europe and North America. Intensive laboratory work with Ceilidh hardware, bonding trials, bonding practises and calibration, gave the research team first-hand insight into deployment challenges and how sensor placement can be optimised for tomography. Finally, working at the intersection of guided waves, deep learning and two-dimensional wave-field imaging surfaced fresh research questions on mode-conversion physics, domain adaptation and uncertainty quantification, laying fruitful ground for the next round of grant proposals.

Impacts for the KTP Associate

The KTP has transformed my professional profile. Immersed in Novosound’s labs, I moved far beyond my original specialism, mastering ultrasonic guided-wave physics and developing successful experimental protocols. At the same time, I sharpened my algorithmic expertise, designing convolutional and transformer models and tuning signal and image processing pipelines, skills that now underpin my current projects. Continuous professional development was equally rich: KTP seminar-series courses deepened my understanding of commercialisation, while my development budget is allocated on two upcoming international ultrasonics conferences that should broaden my network across academia, start-ups and leading asset-integrity firms. Exposure to the agile culture of a high-tech SME taught me to iterate quickly, manage risk and communicate with diverse stakeholders. These experiences are already paying dividends: I have secured a new research-fellow post at the university.

The Impacts and Benefits

The People

Meet the Team

Dr Ioannis Matthaiou

KTP Associate

LinkedIn

Dr Katherine Tant

Knowledge Base Supervisor

LinkedIn

Matt McInnes

Company Supervisor

LinkedIn