Ultrasonic nondestructive testing (NDT) provides a valuable insight into the integrity of stainless steel structures, but the noise caused by the scattering of stainless steel microstructure often limits the effectiveness of inspection. This work presents a novel adaptive filtering approach to enhance the signal-to-noise ratio (SNR) of a measured ultrasonic signal from the inspection of a stainless steel component, enabling the detection of hidden flaws under strong noise. After the spectral modeling of the noisy ultrasonic NDT signal, the difference between the spectral characteristics of a flaw echo and that of grain noise is highlighted, and a reference spectrum model to estimate the frequency spectrum of the echo reflected by any possible flaw is developed. Then, the signal is segmented and the similarity between the spectra of data segments and the reference spectra is evaluated quantitatively by the spectral similarity index (SSI). Based on this index, an adaptive time-frequency filtering scheme is proposed. Each data segment is processed by the filtering to suppress the energy of noise. The processed data segments are recombined to generate the de-noised signal after multiplying weighting coefficients, which again is determined by the SSI. The performance of the proposed method for SNR enhancement is evaluated by both the simulated and experimental signal and the effectiveness has been successfully demonstrated.

Urban railway track infrastructures often suffer from damage that affects their service performance due to a variety of factors. In this study, an unsupervised feature selection and damage identification method based on globally sparse probabilistic principal component analysis (PCA) is proposed for urban railway tracks using the monitoring data of train-induced dynamic responses. A Bayesian framework is proposed for generating principal components on a basis of vectors (original variables) with a global sparseness pattern instead of separate patterns in a traditional sparse PCA. In this framework, a variational expectation-maximization algorithm is employed to obtain the tractable calculation of the marginal likelihood function for learning all uncertain parameters of the Bayesian model. The obtained principal components are linear combinations of the very same set of important variables, making our method better interpretable than the traditional sparse PCA. We can clearly understand which original variables are most relevant for describing the data. The track damage is identified simply by discriminating the corresponding measured dynamic responses using the binary elements of the latent vector inferred from the Bayesian globally sparse PCA algorithm. The usefulness is demonstrated by successfully identifying the track bed plate crack damage through the actual train-induced dynamic responses collected from the structural health monitoring system of an urban railway track infrastructure, where the method is able to achieve $F_1$ scores of 90% or higher for various scenarios.