On the application of heterogeneous transfer learning to population-based structural health monitoring

Population-based structural health monitoring (PBSHM) is a branch of structural health monitoring (SHM) which seeks to leverage information from across a population of structures, with the aim of making robust data-based models that generalise across the population, allowing information to be exchanged and harnessed in constructing better inferences than considering an individual structure alone. PBSHM approaches overcome many of the challenges associated with conventional data-based SHM, such as limited labelled observations in training, classifiers failing to generalise when structural modifications or environmental variations occur, etc. Transfer learning provides an important set of tools in performing PBSHM, with the technologies offering mechanisms for transferring label information between structures, and the ability to harness all the available information from all structures in the population, creating a single classification model that generalises across the complete population. This paper explores heterogeneous transfer learning, a branch of transfer learning where datasets have inconsistent feature spaces, i.e. the dimensions of datasets from one structure are different to those from another. In PBSHM, this scenario arises for several reasons; for example, the data acquisition processes on each structure may be different: e.g. the sample rates and durations were different for each structure, leading to transmissibilities with a different number of spectral lines. The paper compares two heterogeneous transfer learning approaches that are formed in a discriminative manner, namely kernelised Bayesian transfer learning and heterogeneous feature augmentation. The techniques are benchmarked against conventional approaches to data-based SHM, with the benefits of a heterogeneous transfer learning approach highlighted by a case study on a Gnat aircraft wing.