Machine learning applied to fiber optic SHM of inflatable structures for anomaly detection and classification

Machine learning (ML) algorithms were applied to data from fiber optic sensors embedded in inflatable softgoods space habitat structures to enable structural health monitoring (SHM). Fiber optic distributed strain sensing (DSS) with data spacing of 2.6 mm along each sensor was foundational to the presented approach of detecting changes in the structure over time. The ML models were applied to the distributed fiber optic signals to automatically detect structural failures as well as indicators of future failure of the inflatable’s restraint layer materials. In addition to the detection of the event, each event was classified, and the physical location of the structural event was determined. The ML models were trained using a dataset from a one-third-scale inflatable creep test at NASA Johnson Space Center that employed space-grade materials and fabrication techniques. Convolutional neural networks (CNNs) were designed for processing 1D and 2D formats of strain history data to detect features such as yarn and strap breakage events, persistent damage, pressure drops, and creep trend anomalies that indicate impending burst. Parallel convolutional layers within the neural network were used to detect features of vastly different timescales. The CNN models achieved an accuracy of 99.6% or better. There are no other techniques known to the authors that are space-deployable and can yield this level of SHM that gives insight into the inflatable’s creep evolution, dynamic damage events, and overall structural integrity.