Search

ABSTRACTRail corrugation is a prominent degradative problem in the health monitoring of railway systems. Monitoring process is dependent on use of a diagnostic trolley, which is expensive and needs the track to be out-of-service. Alternatively, in-service rail vehicles with Axle-Box Acceleration measurement systems installed, have shown success in detecting rail corrugation levels based on physical models, albeit with limitations. Extending this approach, we build a Machine Learning model, represented by a tuned Random Forest regressor, trained on collected accelerometer signals along with other offline and/or static features. We also propose a method to engineer acceleration-based features which nullifies the aggregated acceleration vibrations inherited from the other rail due to dynamically coupled vibrations between the left and right rails. The resulting model is able to recreate the moving RMS irregularity profile at bandwidth 100–300 mm, especially in highly corrugated sections, with an R2score of 0.97–0.98. The results show that the suggested data-driven approach outperforms a state-of-the-art model-based benchmark.

For the safe and reliable operation of battery-driven machines, accurate state-of-charge (SOC) estimations are necessary. Unfortunately, existing methods often fail to identify patterns relevant to long-term SOC estimation due to complex battery cell characteristics such as aging. In this paper, we propose the Uncorrelated Sparse Autoencoder with Long Short-Term Memory (USAL). USAL is a novel neural network that addresses the challenging task of long-term SOC estimation given a limited initial history of a cell’s charge-discharge behavior. USAL uses a multi-task learning strategy to harness the advantages of sparse autoencoders and Long Short-term Memory (LSTM) networks by enforcing correlation penalties. The USAL simultaneously learns to (i) generate a latent space of informative SOC encodings from commonly measured cell characteristics, (ii) penalize for high multicollinearity between encodings, and (iii) identify non-trivial long and short temporal correlations between the encodings using LSTM cells. USAL outperforms benchmarked models in our experiments when trained on five initial charge-discharge cycles across multiple battery cells using three publicly available accelerated aging datasets. Note to Practitioners—This paper proposes USAL, a custom-built deep neural network to address the challenging task of long-term SOC estimations in battery cells. Long-term SOC estimation involves estimating SOC for cycles near End-Of-Life (EOL) given some initial charge-discharge cycles. Three fundamental steps involved in long-term SOC estimations using USAL are (i) exploiting a multi-task learning strategy to learn efficient encodings given limited training data, (ii) penalizing these encodings for high correlations to efficiently transform measured inputs into a space of informative features, and (iii) mapping of aging-related trends to support long-term SOC estimations. USAL is designed to be a data-driven SOC estimation method that is (i) capable of alerting the user to a faulty cell when integrated into a real-life Battery Management System (BMS) and (ii) identifying the relative quality of a battery cell from only a few initial charge-discharge cycles.

Self-driving vehicles promise many safety, mobility, and environmental benefits. However, users’ lack of trust and acceptance may threaten the success and potential of this technology. Monitoring the driver’s emotional state is one way to address this challenge. Empathetic automation can respond to the driver’s state and improve the experience and acceptance of self-driving vehicle drivers. In this study, 24 participants rode in a self-driving vehicle simulator and experienced three automation styles (aggressive, moderate, conservative) and four intersection types (with and without a stop sign, and with and without traffic.) We identified the observed drivers’ emotions from the video data and labeled the video frames using the dimensional and discrete emotion models to examine how automation behavior affects the driver’s emotional state. We used multilevel Bayesian linear regression and multilevel Dirichlet regression to model the continuous and discrete emotions, respectively. The automation driving style effect varied for each participant. The same conditions provoked positive responses for some participants, and negative for others. Furthermore, the results showed that intersection type, the position within the intersection, and their interaction affected the driver’s emotional state. This indicates that personalized driver state monitoring systems might enhance drivers’ experience in self-driving vehicles.

ABSTRACTOperating units, in practice, often suffer from multiple modes of failure, and each failure mode has a distinct influence on the service life cycle path of a unit. The rapid development of sensor and communication technologies has enabled multiple sensors to simultaneously monitor and track the health status of a unit in real time. However, one challenging question that remains to be resolved is how to leverage data from multiple sensors for better degradation modeling and prognostic analysis, especially when there are multiple failure modes. Currently, many of the existing approaches in prognostics either (a) fail to capture the dependency between sensors and instead focus on analyzing each sensor independently or (b) fail to incorporate the failure-mode diagnosis for better degradation modeling and prognostics during condition monitoring. To address the limitations in the existing literature, we propose a data-level fusion methodology to construct a composite failure-mode index, named FM-INDEX, via the fusion of multiple sensor data. Our goal is to utilize the FM-INDEX to better characterize the failure mode of an operating unit in real time, thus leading to better degradation modeling and prognostic analysis. A case study that involves the degradation data set of an aircraft gas turbine engine with two potential failure modes is conducted to numerically evaluate the performance of our proposed method compared to other techniques in the related literature.

The Bayesian Learning for Neural Networks (BLNN) package coalesces the predictive power of neural networks with a breadth of Bayesian sampling techniques for the first time in R. BLNN offers users Hamiltonian Monte Carlo (HMC) and No-U-Turn (NUTS) sampling algorithms with dual averaging for posterior weight generation. A robust implementation of hyper-parameters and optional re-estimation through the evidence procedure gives BLNN high predictive precision. BLNN is compatible with RStan diagnostic tool ShinyStan. BLNN can be used in a wide range of applications which are based on developing statistical models such as multiple linear and logistic regression, classification, and survival analysis.