IMPACT-FORCE IDENTIFICATION USING DEEP LEARNING AND BAYESIAN INFERENCE WITH APPLICATION ON PIPELINE STRUCTURES

Abstract

Structures, ranging from bridges to pipelines, frequently encounter various impact events that can jeopardize their integrity. For example, a pipeline might experience impacts from falling debris during a construction project, or a bridge might suffer hits from over-height vehicles. In coastal areas, piers and offshore structures often face impacts from floating debris or boats. Each of these scenarios can induce stresses that, if undetected, might lead to critical failures. Therefore, it is essential to measure and analyze these impacts accurately to ensure the longevity and safety of such structures. Traditional methods such as hiring an inspector for this task are often costly, and complex, and can require pipeline shutdowns, leading to economic losses. To address these issues, this research employs a combination of deep learning and Bayesian inference techniques, offering a more efficient approach. 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) impacting the structure. This classification stage utilizes ground truth data to accurately identify the nature of the impact. Following this, the pre-classified 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. These ANNs serve as surrogate models for Bayesian inference, which is used to infer both the impact force and position. By employing this method, the approach effectively creates a total of 12 distinct and specialized surrogate models, corresponding to each combination of sensor and tip type. The testing phase of the ANNs demonstrates a low Mean Squared Error (MSE), indicating a precise prediction of the pipeline's acceleration frequency signals while the CNN tip classifier illustrated 99% and above F1 score in all sensing elements. Here, the Approximate Bayesian Computation with Subspace Simulation (ABC-SS) technique is utilized, with the ANN and CNN test results serving as input. This final step showed promising results with each sensor having over 90% precision with 7% uncertainty in inferring the impact force and 92% with 2% uncertainty in inferring its position along the length of the pipe (will be referred to as the depth). A more robust inference of the impact force and depth location could be done by data fusion where each response coming from each sensor is combined by introducing a new combined distance metric. This way a higher precision and less uncertainty for the depth could be obtained with over 92% precision with 8% uncertainty for the impact force and over 98% precision with 1.8% uncertainty for the depth (location). The obtained results demonstrate the method's reliability and effectiveness in pinpointing impact forces and the location of the impact using even a single far-away sensor.