Machine Learning for In-Water Inspection of Submarine Hull Coatings

M. An, J. Cipolla, A. Shakalis, B. Hiriyur, R. Tolimieri
Prometheus Inc.,
United States

Keywords: MIRK, Submarine Hull, Submarine Availability, data-driven learning, Autoencoder


Employing machine learning, we have adapted the methods from previous projects to detect elastic material properties in sonar returns in real time, to enable the capability to perform in-water submarine hull coating delamination detection. This will reduce the cost and duration of submarine hull coating inspection, a significant factor in overall submarine availability to the fleet. In turn, this will offer important cost savings and risk mitigation for the Navy. Previously, our project for NASA Langley characterized aircraft composites using a data-driven method based on wavenumber images (spatial FFT), and the MIRK (Material Identification Reflectivity Kernel) technology. In addition, research at NUWC by Dr. A. Hull revealed that the physics of delaminated coatings underwater yielded a stronger differentiated scattering pattern than fully-bonded coatings, motivating our research into acoustic methods for detection of delamination. Our testing used a set of eight panels in a Navy tank environment to demonstrate delamination detection on bio-fouled and non-fouled panels. On each panel, six sites with varying sizes (zero to large) of delamination were created. In the test tank, two separate waveforms and various angles of incident ensonification were used to generate the test data set of roughly 1500 independent measurements. Our analysis of the data set showed that the delaminated areas affected the scattered acoustic signal, and that this effect could be detected and classified, even in the presence of bio-fouling. Even without optimizing the measurement waveform, frequency, angle of incidence or other parameters, and with a small dataset, we achieved >80% accuracy in classifying local delamination targets. Challenges encountered in employing machine learning to analyze the test results included: I. The data set was small for robust data-driven learning (eight panels, only six sites per panel were ensonified); II. Feature inputs are not scalar features, but discrete time-signals (dimensionality); III. Multiple signal channels are all pointing to a single panel (no variation in the labels). Our approach to mitigate these challenges and obtain the success that we achieved included: A. To maximize the size of the dataset, we treated each signal as a data point; B. Dimensionality reduction methods were used to lower the dimension of the feature inputs; C. Autoencoder, a learning-based nonlinear dimensionality reduction method mathematically comparable to the Gabor / Wavelet transform, was used; D. After reduction, given the reduced-dimension features along with characteristic features (angle, wavetype, etc.), multiple data-driven classifiers were trained. The Navy is in the process of arranging for data to be gathered just before a submarine is taken into drydock, so that we may employ our new delamination detection technology on this data and compare the results to what is subsequently found using traditional methods, thus providing a true blind test. When proven successful, this technology will potentially transition to various classes of U.S. submarines.