Although guided wave structural health monitoring (SHM) has been proven useful for ensuring safety, estimating performance deterioration, and detecting damages in structures, it experiences setbacks in accuracy due to varying environmental, sensor, and material factors. To combat these challenges, environmentally variable sensor data can be transformed to match a baseline signal and enable accurate damage detection. Methods currently exist to perform temperature compensation on guided waves, including the scale transform and optimal signal stretch methods. However, these methods have limitations because as stretch factors reach extreme values, performance declines. This leads to less accurate signal reconstructions and causes less accurate damage detection results.This paper addresses this challenge by demonstrating a machine learning (ML) method that can predict stretch factors in guided waves more effectively than the scale transform and other related stretch-based temperature compensation techniques. This effort uses feedforward neural networks (FFNN) due to their ability to accurately approximate complex, continuous functions. The FFNN was trained with 50,000 simulated ultrasonic guided Lamb wave pairs, each containing a baseline and a “stretched” baseline with a different velocity. Each guided wave signal was 2,500 data points, representing 10ms of data. The velocity variation ranged from 0.97 to 1.03 times the baseline velocity. These velocity stretch factors were saved as labels (the output of the FFNN). The first 40,000 measurements were used for training, and the last 10,000 were used for testing.&nbsp; The machine learning framework PyTorch was used to train and create two FFNN models. The first model was a 3-layer model with tanh activation functions. The second model omits the activation functions and instead applies a natural log to the input data and an exponential operation after the hidden layer. Both approaches were constructed the same, besides the activation function, for direct comparison. Their input was of size 5000, the single hidden layer was of size 2500, and the output was of size 1. Each model used 500 epochs, a batch size of 100, a mean squared error loss function, and a learning rate of 1.0 x 10^-5.The testing data was applied to both FFNN models as well as with the scale transform to estimate the stretch factors and compare results. Mean squared error (MSE) was used as a measure of performance. All models successfully predicted stretch factors with small errors. The scale transform MSE was 1.1 x 10^-4, the standard FFNN MSE was 2.8 x 10^-4, and the logarithm FFNN MSE was 5.0 x 10^-5. The standard FFNN performed slightly worse than the scale transform and the log FFNN overcame the scale transform by a factor of 2.2. It is believed that the natural logarithm performs better because the logarithm transformation changes multiplication to addition and enables inner products to be computed with a neural network. This work is the first to demonstrate that we can potentially improve on existing environmental compensation methods with neural networks, enabling higher accuracy in damage detection.&nbsp;Acknowledgements: This work was supported by the National Science Foundation under award number ECCS-1839704.

Estimating Guided Wave Velocity Variation With Neural Networks

Paper Proceedings: https://asmedigitalcollection.asme.org/qnde/QNDE2021/volume/85529

ASME 2021 48th Annual Review of Progress in Quantitative Nondestructive Evaluation (QNDE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.