Zhongmin Deng

This study introduces a machine learning (ML) framework for efficient aero-structural characterization of wing planform shapes, addressing the computational challenges posed by traditional CFD and FEA methods. The goal is to develop an ML-based aero-structural optimization framework that replaces time intensive computational tasks with faster yet reliable approach. Leveraging advanced parameterization, data normalization, and reduced-order modeling (ROM), two regressor chain-based surrogate models are developed to predict aerodynamic and structural responses from identical shape parameters. For CFD buffet predictions, a Proper Orthogonal Decomposition-based surrogate model (POD-ML) is employed, while for stress predictions, the data is first clustered, and then POD is applied within each cluster, forming the K-means ML model. Using the ONERA M6 wing as a baseline, the surrogate models' performance is evaluated across various shape parameters like sweep, dihedral, and twists. The models effectively predict aerodynamic responses, such as pressure coefficients ( C _p ), with a maximum Mean Squared Error (MSE) of 0.05, requiring only 1,000 training samples — significantly fewer than conventional neural network models. The models also handle complex structural predictions with good accuracy, particularly for stress fields, achieving a maximum MSE of 0.10. Comparisons reveal that K-means ML outperforms POD-ML in stress field prediction, though it performs slightly worse for aerodynamic responses. Together, these methods provide substantial computational savings, facilitating rapid design iterations and advancing the optimization of wing architectures.

Because of the operating environment and load, the main fault form of flywheel bearing is the friction fault between the cage and the rolling elements, which often lead to an increase in the friction torque of the bearing and even to the failure of the flywheel. However, due to the complex mechanism of the friction fault, the characteristic frequencies often used to indicate cage failure are not obvious, which makes it difficult to monitor and quantitatively judge such faults. Therefore, this paper studies the mechanism of the friction fault of the flywheel bearing cage and establishes its fault feature identification method. Firstly, the basic dynamic model of the bearing is established in this paper, and the friction between the cage and the rolling elements is simulated by the variable stiffness. The influence law of the bearing vibration response reveals the relationship between the periodic fluctuation of cage-rolling element friction failure and the bearing load. After analyzing the envelope spectrum of the vibration data, it was found that when a friction fault occurred between the cage and the rolling element, the rotation frequency component of the cage modulated the rotational frequency component of the rolling element, that is, the side frequency components appeared on both sides of the characteristic frequency of the rolling element (with the characteristic frequency of the cage as the interval). In addition, the modulation frequency components of the cage and rolling element changed with the severity of the fault. Then, a modulation sideband ratio method based on envelope spectrum was proposed to qualitatively diagnose the severity of the cage-rolling element friction faults. Finally, the effectiveness of the presented method was verified by experiments.

Finite element model updating precision depends heavily on sufficient vibration feature extraction. However, adequate amount of sample collection is generally time-consuming in frequency response (FR) model updating. Accurate vibration feature extraction with insufficient data has become a significant challenge in FR model updating. To update the finite element model with a small dataset, a novel approach based on transfer learning is firstly proposed in this paper. A readily available fault diagnosis dataset is selected as ancillary knowledge to train a high-precision mapping from FR data to updating parameters. The proposed transfer learning network is constructed with two branches: source and target domain feature extractor. Considering about the cross-domain feature discrepancy, a domain adaptation method is designed by embedding the extracted features into a shared feature space to train a reliable model updating framework. The proposed method is verified by a simulated satellite example. The comparison results manifest that sample amount dependency has prominently lessened this method and the updated model outperforms the method without transfer learning in accuracy with the small dataset. Furthermore, the updated model is validated through dynamic response out of the training set.

This paper proposes an improved reliability analysis method for composite structures. A new uncertainty model is introduced to quantify the uncertainty in the elastic properties and strength characteristics of the constituents. A multi-scale analysis procedure is developed to link scales in elastic properties, and solve the mechanical responses of all scales. A set of limit state functions are defined, corresponding to the failure modes of intralaminar constituents’ failure and interlaminar delamination. First-order reliability method presents good convergence and accuracy of reliability estimation. Numerical cases are studied to demonstrate the validity and improvements of the proposed method.

A unidirectional convolutional neural network (UCNN)-based approach is proposed in this paper to take advantage of frequency response (FR) data for finite element model updating. UCNN is designed to acquire the high-precision inverse mapping from the FR data to the updating parameters without any artificial feature extraction. Unidirectional convolution is applied at the frequency and location dimensions of the FR data separately to avoid data coupling. UCNN outperforms the residual-based model updating method and two-dimensional convolutional neural networks on a satellite model updating experiment. It achieves highly- precision results both inside and outside of the training set.

In this paper, the propagation of hybrid uncertainties is studied in transient heat conduction problems. Based on the layer-by-layer analysis strategy, a novel mixed method using stochastic theory and convex model is presented. First, uncertainties are modelled as random parameters and uncertain-but-bounded parameters respectively. Then matrix perturbation theory is utilized to deal with random parameters, obtaining the temperature response expectation and variance. Finally, using the Taylor series expansion and the Lagrange multiplier method to analyze the convex model, derive the intervals of the temperature response probabilistic characters. Four numerical examples are given to verify the feasibility and practicality of the proposed method to deal with transient heat conduction problems with random and uncertain-but-bounded parameters or pure uncertainties, the results compared with Monte Carlo method. In addition, the proposed method is also applicable to the steady-state problems.