News facts:
- ‘Airbus AI Gym’ challenge to find most accurate unsupervised predictive AI capability for helicopters awards first prize to Fujitsu
- Winning solution achieved 93% precision; identifies when sensors are functioning unusually and shows early warnings for vehicle faults
Fujitsu has been awarded first prize by leading global aerospace manufacturer Airbus SE in a worldwide competition to find the most accurate use of an unsupervised artificial intelligence (AI) system.
Top ranking in the Airbus AI Gym1 challenge for accurate sensor monitoring went to Fujitsu for developing a way of using unsupervised AI to detect anomalies in accelerometer data from Airbus pre-certification helicopters, ahead of 140 other teams participating in this helicopter challenge.
Flight engineers attach large numbers of sensors onto test helicopters to capture every nuance of behavior. To enhance detection of early-warning signals in this vast amount of data Airbus established its AI Gym challenge, fostering research into a new way of accurately locating potential issues, especially data outliers. A multi-disciplinary team of specialist engineers supports each flight to study this mass of observations – a major investment in every flight made. Because almost all sensor data is considered ‘normal’ this mechanism should be able work without prior guidance from engineers.
Fujitsu’s winning solution achieved 93% precision, leveraging its “DeepTAN” Unsupervised AI Model created by the company’s sub-division, Fujitsu Systems Europe2 (FSE). The solution took data sequences from multiple sensors and analyzed them across a fixed time period, detecting abnormal sensor behavior using a deep learning algorithm based on Multivariate Anomaly Detection with Generative Adversarial Networks3 [MAD-GAN]. FSE trained and validated the algorithm in its own data center using 1,677 one-minute-sequences of accelerometer data from test helicopters flying at various locations, angles and flights.
Fujitsu plans to industrialize its solution for unsupervised time-series analysis solution, complementing DeepTAN with end-to-end functionality, integrated data pipelines and further evolved algorithms.
New functions include a semi-supervised mechanism to classify the type of sensor anomaly, addressing the imperative for engineers and maintenance services to find the root cause of anomalies, and to interpret multi-variate data and correlations between all test flights in a program. The company will then be able to bring value to customers across the aircraft lifecycle, from test flight and pre-delivery, to airlines and Aviation MRO4 organisations.
Ian Godfrey, Director Solutions Business at FSE, says: “Winning first prize in this data challenge not only underlines Fujitsu’s world-leading AI expertise and technologies – it also provides concrete evidence of our ability to apply them to real-world business scenarios. The concepts we applied to this specific problem have shown us how these new deep learning techniques not only help manufacturers but the firms working to sustain aircraft in service as well.”
