A Data-Driven Framework Based on Sparse Representation of Ultrasonic Guided-Waves for Online Damage Detection of Pipelines
This work addresses some of the open challenges in guided-wave based structural health monitoring (SHM) of pipelines. In this dissertation, we review these challenges under three headings: (a) Multiple modes, (b) Multi-path reflections, and (c) Sensitivity to environmental and operational conditions (EOCs). The objective is to develop damage detection methods that (1) simplify guided-wave signals, and (2) have low sensitivity to EOC variations. First, I propose a supervised method for online damage detection. The detection performance is maximized under variety of EOCs, by imposing a sparsity constraint on the signals. In the training stage, data is recorded from an intact pipe, as well as the pipe with structural abnormalities. During the monitoring stage, test signals are projected on the extracted sparse discriminant vector, and these projections are used as damage-sensitive features. I conduct a diverse set of laboratory and field experiments to investigate and to validate the extent to which EOC variations, as well as the differences in characteristics of the structural abnormality in training and test data, can influence the discriminatory power of the damage-sensitive features. The validation results suggest that a simple binary-labeled training data (i.e., undamaged/damaged), obtained under a limited range of EOCs, is sufficient for the proposed method. In other words, the detection method does not require prior knowledge about the characteristics of the damage (e.g., size, type, location), and/or a training dataset that is obtained from a wide range of EOCs. ix Next, I propose an unsupervised method to address some of the limitations of the aforementioned supervised approach. The unsupervised approach eliminates the need for training data captured from a pipe with structural abnormality, which can be a challenge for some applications such as pipes with restricted accessibility. Therefore, the damage-related training parameters that may affect the detection performance of the supervised method are not an issue for the unsupervised approach. The proposed unsupervised method takes advantage of two facts that are further verified throughout this work: (1) high-energy arrivals are less sensitive to EOC variations compared to the rest of the signal, and (2) damage changes the energy-content and/or time-location of high-energy arrivals in the signal. For this method, the training dataset is not labeled, but is assumed to mostly include intact signals. In the training stage, a sparse subset of high-energy arrivals from intact pipes is extracted so that energy estimation error is minimized. My experimental analysis proved that high-energy arrivals in intact signals are located at different time-points than those in damaged signals. Therefore, using the extracted sparse subset (which is mainly composed of high-energy arrivals), the energy estimation error will increase as damage occurs. The proposed method is proved successful for online detection of damage under varying EOCs. It is also shown that the wider the range of EOCs in the training dataset, the better the detection performance. This range, however, is not required to be comprehensive of all possible testing scenarios. For example, for a test dataset in which temperature varies between 24℃ and 39℃, a training dataset with temperatures ranging between 24℃ and 30℃ results in separation accuracy of 99%, and detection delay of five observations, captured in one-minute intervals.