A data-driven normal contact force model based on artificial neural network for complex contacting surfaces. (July 2021)

• Record Type: Journal Article
• Authors: Ma, Jia; Dong, Shuai; Chen, Guangsong; Peng, Peng; Qian, Linfang
• Is Part Of: Mechanical systems and signal processing. Volume 156(2021)
• Publication Date: 2021-07
• Subjects: Complex contacting surfaces -- Data-driven modelling framework -- Artificial neural network -- Normal contact force model; Structural dynamics -- Periodicals; Vibration -- Periodicals; Constructions -- Dynamique -- Périodiques; Vibration -- Périodiques; Structural dynamics; Vibration; Periodicals; 621
• DOI: 10.1016/j.ymssp.2021.107612
• Languages: English
• ISSNs: 0888-3270

Abstract

Highlights: This work provides a data-driven modelling framework for the research of contact/impact process between complex contacting surfaces. Indoor experiment rig between complex contacting surfaces is manufactured and displayed. A neural-network-based contact force model between barrel and bourrelet is established. Results obtained confirm that the proposed model can achieve high accuracy and also present excellent generalization ability. Abstract: Proper modelling of contact/impact phenomenon is critical to ensure reliable description of the overall dynamic behaviors of mechanical systems. The past few decades witnessed substantial developments on contact/impact dynamics modelling, especially for the smooth contacting surfaces, like spheres or cylinders. Contrastingly, less attention has been paid to the urgent modelling demand for complex contacting bodies. By utilizing the data-driven modelling framework based on artificial neural network, this paper aims to provide a new and feasible scheme for the research of contact/impact process between complex contacting surfaces. Taking the contact/impact process between barrel and bourrelet as our research object, the indoor experiment rig is manufactured and displayed for the first time. Measurement results collected under different initial indentation velocities serve as the training datasets of the learning process for the data-driven normal contact force model. After that, the optimum hyper-parameters of the neural network, mainly including the performance index, activation function, structure of network, and learning algorithms, are tuned for the contact/impact process between barrel and bourrelet through trail-and-error method. Eventually, the neural-network-based normal contact force model can be established, of which the prediction performance for interaction modelling is further analyzed and verified. Simulation results confirm that the proposed data-driven normal contact force model can achieve high accuracy and also present excellent generalization ability. Great agreements with the experimental results under the chosen network structure demonstrate the effectiveness of data-driven interaction modelling methodology presented for complex contacting geometries.