22 Rejuvenation through AI Inevitably, there is more to it than that, and Dowdeswell is also overseeing the rejuvenation of the computing elements of sonar systems, partly through the application of AI. He explains: “AI and machine learning have brought big improvements to the processing and interpretation of sonar data, primarily for object detection and recognition related to side-scan sonar and also to bathymetry.” He says much of the pioneering work with AI applied to sonar had military applications, such as in mine countermeasures, for example, and it is still a very active area. “So, you may be looking for a cylindrical object on the seabed and to differentiate an oil drum from a mine. That’s where a lot of the industry’s r&d has been.” More recently, offshore construction of wind farms and their associated cable routes has come to the fore, and there is a lot of AI development going into the recognition and counting of boulders, which can be major obstacles to the installation of such infrastructure, and the set of applications is broadening. “Within the field of AI there are various branches, but the main division is between supervised or unsupervised approaches. In supervised approaches, you have raw data, you have the finished result, and you train some form of AI – generally a neural network – to map input to output. And that’s what I was doing 30 years ago with optical spectra,” Dowdeswell recalls. This was his PhD project, and it involved examining liquid effluent from a brewery in Bedford using a spectrometer to measure light wavelengths across a spectrum, revealing gaps in the normal spectrum caused by something absorbing particular wavelengths. “What it was looking for was abnormal instances where someone had left a valve open and the effluent was being contaminated – pollution breakthrough. You trained it on lots of examples of normal effluent and then told it to flag up something it had never seen before, an exception,” he says. “One analogy I sometimes use is: if I wanted an AI to tell the difference between dogs and rabbits, I would give it lots of pictures of dogs and rabbits, all labelled. The network would then build up a way of analysing a picture and giving you a determination. “If I then gave it a picture of a little pug with rabbit ears stuck on, it may struggle if the network says the dominant feature of rabbits is two big ears poking up, as it would go for that, whereas it might say there are other features of a dog that are dominant in determining whether it is a dog. “Such supervised AI is fine so long as what you’re asking it to do is within its knowledge base, whether it has been trained on it. If you give it something that it’s never seen before, it will give you an answer, but your confidence in its output will not be as high as it would be if the example were in the training set. “With bathymetry, we took the unsupervised approach, because in order to cover every example that you might want to survey – sandy and rocky seabeds, coral reefs, wrecks, harbour walls, etc – you would have to give it labelled examples, and in different water conditions at different depths that’s an awful lot of data,” says Dowdeswell. In unsupervised learning, the algorithm is given an unlabelled dataset and tasked with finding patterns or structure within the data on its own, without explicit guidance. Presented with new data, unsupervised learning algorithms typically focus on uncovering hidden patterns, similarities or relationships within the data. When Dowdeswell joined the company in 2017, he found that processing raw sonar data was a pattern recognition problem, so the company started an r&d programme with a team from the Computer Science Department at the University of East Anglia. They included two master’s students, one of whom focused on supervised AI while the other took the unsupervised approach. “Both worked reasonably well, but there’s a lot of effort that has to go into providing supervised AI with labelled data, which is why we went down the unsupervised route.” AI for interferometric sonar The first of GeoAcoustics’ sensor systems to benefit from AI is the GeoSwath interferometric sonar, which Dowdeswell contrasts with multi-beam echo sounder (MBES) echo sounder, although both June/July 2024 | Uncrewed Systems Technology In conversation | Dr Richard Dowdeswell Output from a GeoPulse Compact sub-bottom profiler. Layer boundaries in rock and sediment reflect different wavelengths, revealing important features below the seabed (Image courtesy of GeoAcoustics)
RkJQdWJsaXNoZXIy MjI2Mzk4