Research paper accepted by IEEE Transactions on Reliability
In this paper, we develop a data-efficient learning-based approach for accurately estimating the time-varying kinematic reliability of mechanisms. To improve data efficiency, we develop an adaptive sampling scheme that is guided by the prediction uncertainty of the resultant deep learning model. Overall, the proposed methodology follows a four-step procedure. First, a spectral-normalized neural Gaussian process (SNGP) is employed to quantify the prediction uncertainty of the deep learning model used to evaluate the time-varying kinematic reliability of mechanisms. Unlike other uncertainty quantification methods, SNGP has two unique characteristics: SNGP normalizes the weights of hidden layers of neural network to preserve the relative distances among samples during data transformation; SNGP replaces the conventional dense output layer of neural network with a random Fourier features-based Laplace approximation of Gaussian process for distance-aware uncertainty quantification. Next, based upon the uncertainty estimated by SNGP, we devise an active learning function to adaptively identify and select the most valuable next data sample to add to the training set. Thirdly, based on a predefined data acquisition metric, we iteratively add data points that maximize the active learning function to the training set and continue updating the deep learning model until a termination condition is met. Finally, we benchmark the developed approach for time-varying reliability analysis of mechanisms against Monte Carlo simulation, PHI2, and the Monte Carlo dropout-based uncertainty quantification method in terms of estimation accuracy and computational efficiency using three numerical examples. Computational results suggest that the developed method substantially reduces the number of performance function evaluations while achieving reliability estimates that are more accurate than those from MC dropout and comparable to those from MCS and PHI2.
Research paper accepted by Decision Support Systems
Credit card fraud detection remains a significant challenge due to extreme class imbalance, evolving fraud patterns, and asymmetric misclassification costs. While existing machine learning methods often prioritize fraud detection performance, they frequently overlook model risk management, which is crucial for reliable decision making in high-stakes environments. To address this issue, we introduce DISCO — a framework that combines discriminative and robust representation learning with a conformal prediction-based mechanism for provable risk control. In essence, DISCO is established upon the principle of decoupling representation learning from risk-controlled decision-making. To this end, we first leverage deep metric learning (DML) to construct an embedding space that is inherently robust to class imbalance and resilient to concept drift. Subsequently, we employ a conformal risk control (CRC) mechanism to provide a formal statistical guarantee on the false negative rate (FNR) of the resulting model according to a user-specified FNR target. Computational experiments on a real-world dataset demonstrate that DISCO outperforms state-of-the-art methods in both classification performance and operational efficiency. To the best of our knowledge, this is the first work to integrate a DML-based representation learning with a formal risk control mechanism for reliable credit card fraud detection. The proposed framework therefore offers a trustworthy and practical solution that addresses the dual imperatives of predictive performance and reliability.
Research paper accepted by Neural Networks!
Deep learning has exhibited a promising performance in a series of task-specific bearing fault diagnosis problems. However, the current literature lacks a versatile model that could serve as a general-purpose feature extractor for a wide range of downstream fault diagnosis tasks. To fill this gap, this paper develops a theoretically well-grounded approach to build an \bl{enhanced} backbone model for various key downstream tasks. The proposed methodology follows a three-step procedure. In the first step, using approximation theory, we prove that incorporating neural heterogeneity by combining linear and quadratic neurons leads to a more efficient approximation of any univariate real coefficient polynomial. Compared to conventional neural networks, HNNs exhibit enhanced representation capability while using fewer model parameters. Secondly, building upon this theoretical foundation, we design an \bl{enhanced} feature extractor termed as heterogeneous neural blind deconvolution (HBD). At a high level, HBD is comprised of two time domain blind deconvolution branches in parallel: one using regular convolutional networks and the other using quadratic convolutional network. The inclusion of diverse neurons in HBD facilitates to learn discriminative features for reliable fault diagnosis that neither neuron type could achieve independently. Following the dual time domain blind deconvolution branches, a frequency-domain BD module complements the feature extraction capability of the time domain blind deconvolution by performing signal filtering in the frequency domain. Finally, to illustrate the general-purpose nature of HBD, we explore the application of HBD across various downstream fault diagnosis tasks, including anti-noise fault diagnosis, cross-domain fault diagnosis, and lightweight model for fault diagnosis on edge devices. Extensive experiments and comparisons with state-of-the-art baselines clearly show the advantage of HBD in enhancing the accuracy and interpretablity for bearing fault diagnosis.
