Pinheiro Giovanni de Souza, Setúbal Fábio Antônio do Nascimento, Filho Sérgio de Souza Custódio, Mesquita Alexandre Luiz Amarante, Nunes Marcus Vinicius Alves
Institute of Technology, Federal University of Pará, Belém 66075-110, Brazil.
Marabá Industrial Campus, Federal Institute of Pará, Marabá 68740-970, Brazil.
Sensors (Basel). 2024 Oct 17;24(20):6675. doi: 10.3390/s24206675.
The knowledge of the forces acting on a structure enables, among many other factors, assessments of whether the component's useful life is compromised by the current machine condition. In many cases, a direct measurement of those forces becomes unfeasible, and an inverse problem must be solved. Among the solutions developed, machine learning techniques have stood out as powerful predictive tools increasingly applied to engineering problem-solving. This study evaluates the ability of different machine learning models to identify parameters of multi-force systems from accelerometer measurements. The models were assessed according to their prediction potential based on correlation coefficient (R), mean relative error (MRE), and processing time. A computational numerical model using the finite element method was generated and validated by vibration measurements performed using accelerometers in the laboratory. A robust database created by the response surface methodology in conjunction with Design of Experiment (DOE) was used for the evaluation of the ability of machine learning models to predict the position, frequency, magnitude, and number of forces acting on a structure. Among the six machine learning models evaluated, k-NN was able to predict with a 0.013% error, and Random Forests showed a maximum error of 0.2%. The innovation of this study lies in the application of the proposed method for identifying parameters of multi-force systems.
了解作用在结构上的力，在诸多因素中，有助于评估部件的使用寿命是否会因当前机器状态而受到影响。在许多情况下，直接测量这些力变得不可行，必须解决一个反问题。在已开发的解决方案中，机器学习技术已成为越来越多地应用于工程问题解决的强大预测工具。本研究评估了不同机器学习模型从加速度计测量中识别多力系统参数的能力。根据模型基于相关系数（R）、平均相对误差（MRE）和处理时间的预测潜力对其进行评估。使用有限元方法生成了一个计算数值模型，并通过在实验室中使用加速度计进行的振动测量进行了验证。由响应面方法结合实验设计（DOE）创建的一个强大数据库被用于评估机器学习模型预测作用在结构上的力的位置、频率、大小和数量的能力。在所评估的六个机器学习模型中，k近邻算法能够以0.013%的误差进行预测，随机森林算法的最大误差为0.2%。本研究的创新之处在于将所提出的方法应用于识别多力系统的参数。
