In conversation | Dr Richard Ahlfeld 18 and 18 months. They’re wasting huge amounts of money, two to five times what they should be spending, from our experience,” Ahlfeld says. “But the longer you’ve been in the game, the better you’ve learned how to safety-rate a cell using a shorter and smarter test regimen. Peter Attia, who used to lead battery testing at Tesla, was a Stanford researcher specialising in AI for battery testing. He proved he could reduce the time and cost of battery testing programmes by something like 98%. Within the controlled limits of battery tests, AI can learn how to solve these problems much faster than humans.” How AI learns Monolith’s AI platform functions on different types of learning, depending on the application. For optimising a battery test plan, the algorithm goes through a sequence of steps analogous to learning the rules of a game. “Once it learns the rules, it learns to optimise its next move. It is essentially reinforcement learning, with specific Bayesian optimisations to make it work robustly in dimensions with high noiseto-signal ratios, like battery tests, which often output very jagged, erratic curves in cells’ performance parameters,” Ahlfeld explains. “Inference is the most important thing: we give the platform a data set, and it has to draw conclusions and assumptions from that set in order to optimise the next one. It is an entire area of machine learning, and we’ve gone through all types available over the last five years, applied them to real test lab environments and picked the ones that worked, adding on user interfaces that make them easier to interact with. “For our test lab optimisation toolbox, users upload the results they have so far, and the algorithm goes through a search to figure out the rules of the game and the final goal in order to output recommendations.” He likens the fundamental behaviours of the toolbox to how some social media apps recommend content. From an algorithmic perspective, both types of system must work in noise-heavy environments and infer what should be chosen next, based on past choices (be they clicks, or the number and types of cell tests). While there are variations, fundamentally, each solution in the Monolith platform has been engineered in the same way: the team has identified an ideal machine-learning tool for solving the specific problem of how to test a battery faster, and then made it easier for engineers to work with it. AI in battery testing Three AI solutions within Monolith stand out with regard to battery testing. The first is its Anomaly Detector software, which monitors users’ test stations 24/7, and upon detecting something indicative of a fault condition, alerts them to stop the test. “In predictive maintenance, you have a finished product with well-established behaviours. With new battery cells, you don’t know how they’re supposed to work, but with this solution’s self-learning, deep learning-based algorithm, just 20 seconds of training it with a ‘golden run’ of ideal test results is enough for it to start learning what is normal for safe cell behaviour, thereafter telling you – almost in a paranoid way – whenever a cell does something unusual,” Ahlfeld says. The system can work across hundreds of channels with different sensor types, tracking and judging not only voltage or current sensors but also accelerometers, gyroscopes, torque sensors and other automotive-testing equipment. “Lots of things can trigger an alert to the Anomaly Detector in battery tests. We have seen it flag when a cell swelled up slightly or overheated, when a sensor drifted over time, technicians running the wrong script, unusual vibrations caused by a researcher doing a dance nearby, and it can be configured to prioritise certain signals over others,” Ahlfeld notes. “Users don’t have to fine-tune an advanced deep neural network. They just label acceptable data signals as okay, so the system learns what is normal behaviour for a cell, and hence what isn’t normal or acceptable, be it overheating, overcurrents and so on.” The second solution is a Next Test Recommender, which formulates a May/June 2024 | E-Mobility Engineering Monolith can be applied to anything with a prototype to be validated, sensors to pull data from and a moderately complex range of different test conditions
RkJQdWJsaXNoZXIy MjI2Mzk4