Impact-force identification using deep learning and Bayesian inference with application on pipeline structures

Structures like bridges and pipelines are vulnerable to impacts from various sources, such as falling debris, over-height vehicles, or floating objects, which can compromise their integrity. Traditional inspection methods are costly and complex and can lead to economic losses due to shutdowns. This research introduces a probabilistic and more efficient approach by combining deep learning and Bayesian inference techniques to accurately measure and analyze these impacts, aiming to enhance the longevity and safety of these structures while mitigating potential critical failures. In this methodology, a novel two-stage approach is implemented to resolve the inverse problem of identifying impact forces on structures. Initially, a convolutional neural network (CNN) classifier for each of the four sensors is employed to determine the impacting material (aluminum, rubber, plastic). This classification stage utilizes ground truth data to identify the nature of the impact accurately. Following this, the preclassified data, categorized by the actual impacting materials, is directed into one of three 5-layer artificial neural networks (ANNs), each designated for a different impacting material. The ANNs act as surrogate models for Bayesian inference to determine impact force and position, resulting in 12 specialized models, each corresponding to a specific combination of sensor and tip type. The ANNs were evaluated using mean squared error, showing precise predictions of the pipeline’s acceleration frequency signals, while the CNN classifier achieved over 99% F1 scores. Using ANN and CNN results, the approximate Bayesian computation with subset simulation technique showed over 90% precision with 7% uncertainty in inferring impact force and 92% precision with 2% uncertainty in determining the impact position along the pipe. Data fusion, combining sensor responses, further improved precision and reduced uncertainty, achieving over 92% precision with 8% uncertainty for impact force and over 98% precision with 1.8% uncertainty for depth. These results demonstrate the method’s reliability and effectiveness in accurately identifying impact forces and locations, even with a single distant sensor.