Search

In almost all PHM applications, driving highest possible performance (prediction accuracy and robustness) of PHM models (fault detection, fault diagnosis and prognostics) has been the top development priority, since PHM models’ performance directly impacts how much business value the PHM models can bring. However, recent research work in other domains, e.g., computer vision (CV), has shown that machine learning (ML) models, especially deep learning models, are vulnerable to adversarial attacks; that is, small deliberately-designed perturbations to the original samples can cause the model to make false predictions with high confidence. In fact, adversarial machine learning (AML) targeting security of ML algorithms against adversaries, has become an emerging ML topic and has attracted tremendous research attention in CV and NLP. Yet, in the PHM community, not much attention has been paid to adversarial vulnerability or security of PHM models. We contend that the economic impact of adversarial attacks to a PHM model might be even bigger than that to hard perceptual problems and thus securing PHM models from adversarial attacks is as important as the PHM models themselves. Also, PHM models, since the data used by the models are primarily time-series sensor measurements, have their own unique characteristics and deserve special attention in securing them. In this paper we attempt to explore the adversarial vulnerability of PHM models by conducting an initial case study. More specifically, we consider several unique characteristics associated with streaming time-series sensor measurements data in developing attack strategies for attacking PHM models. We hope our initial study here can shed some light on and stimulate more research interests in the area of PHM models’ security.

Deep learning (DL), regarded as a breakthrough machine learning technique, has proven to be effective for a variety of real-world applications. However, DL has not been actively applied to condition monitoring of industrial assets, such as gas turbine combustors. We propose a deep semi-supervised anomaly detection (deepSSAD) that has two key components: (1) using DL to learn representations or features from multivariate, time-series sensor measurements; and (2) using one-class classification to model normality in the learned feature space, thus performing anomaly detection. Both steps use normal data only; thus our anomaly detection falls into the semi-supervised anomaly detection category, which is advantageous for industrial asset condition monitoring where abnormal or faulty data is rare. Using the data collected from a real-world gas turbine combustion system, we demonstrate that our proposed approach achieved a good detection performance (AUC) of 0.9706 ± 0.0029. Furthermore, we compare the detection performance of the proposed approach against that of other different designs, including different features (i.e., the deep learned, handcrafted and PCA features) and different detection models (i.e., one-class ELM, one-class SVM, isolation forest, and Gaussian mixture model). The proposed approach significantly outperforms others. The proposed combustor anomaly detection approach is effective in detecting combustor anomalies or faults.

Cyber-physical systems (CPSs) security has become a critical research topic as more and more CPS applications are making increasing impacts in diverse industrial sectors. Due to the tight interaction between cyber and physical components, CPS security requires a different strategy from the traditional information technology (IT) security. In this paper, we propose a machine learning-based attack detection (AD) scheme, as part of our overall CPS security strategies. The proposed scheme performs AD at the physical layer by modeling and monitoring physics or physical behavior of the physical asset or process. In developing the proposed AD scheme, we devote our efforts on intelligently deriving salient signatures or features out of the large number of noisy physical measurements by leveraging physical knowledge and using advanced machine learning techniques. Such derived features not only capture the physical relationships among the measurements but also have more discriminant power in distinguishing normal and attack activities. In our experimental study for demonstrating the effectiveness of the proposed AD scheme, we consider heavy-duty gas turbines of combined cycle power plants as the CPS application. Using the data from both the high-fidelity simulation and several real plants, we demonstrate that our proposed AD scheme is effective in early detection of attacks or malicious activities.

Clustering is a challenging problem when many features are irrelevant to separate clusters. Also, different clusters may relate to various feature subsets. This work proposes a deep clustering algorithm that localizes the search for instance clusters and their relevant features. The relevant features of each cluster are defined as those with high associations (dependency) within that cluster. Given the number of clusters $K$ , we formulate the problem as $K$ . -parallel auto-reconstructive learning, where low-rank graph learning, rooted in graph Laplacian theory, is used to explore the unknown feature associations of each cluster. The model performs automatic feature weighting on residuals to minimize loss from the corresponding cluster. Through such design, different feature subsets can be learned to calculate the loss from different clusters. Subsequently, the associations between features and clusters can be acquired, and better clustering result can be achieved. Moreover, the associated features of each cluster can be used to interpret the clustering patterns.

Aircraft sensors generate multivariate time series during flights, where each sensor corresponds to one variable. During normal operation mode, the associations (dependencies) among variables are mainly stationary. One type of flight anomaly that is of interest relates to variable association change. Detection and diagnosis of such type of anomaly need to pinpoint the time series, i.e., variables related to association change, which helps in understanding the underlying mechanisms of anomalies. However, it is hard to detect such change because the variable associations are usually unknown and complicated, and the anomalous samples are usually insufficient for learning the substandard association. In this work, we present a neural network that can 1) detect this type of anomalies given multivariate time series as input; 2) locate the association change by learning the nonlinear variable associations from both normal data and the detected anomalies. Specifically, we leverage the learned model from normal data to learn the faulty association of the anomalies. Experiments using simulated and real-world flight data show that our model outperforms existing methods in flight anomaly detection and diagnosis.

The discovery of ‘event signatures’ and useful insights from very large historical Phasor Measurement Unit (PMU) datasets is predicated on offline Big Data analysis approaches that rely on the generation of predictive features on a massive scale. This paper presents lessons learned from a data platform perspective towards reducing barriers to adoption of Big Data analytics against a real dataset of almost half a trillion data points drawn from over 400 PMUs distributed across the North American power grid. We demonstrate software abstractions and targeted performance optimizations that can lead to significant productivity gains for power systems researchers seeking to perform offline exploratory temporal analysis and modeling tasks, with a focus on feature generation. We describe how our optimized approach goes beyond a naive application of mainstream Big Data technologies, enabling feature generation tasks, that previously took days or even weeks, to now be completed in just a few hours.

This paper presents the design of a novel choke ring center-fed antenna. The proposed antenna consists of two layers of microstrip patches, a folded ground plane, two layers of horizontal coaxial corrugated grooves and two layers of longitudinal radial corrugated grooves. Compared with the existing antenna that only adopts the choke ring structure in the horizontal direction, the antenna can effectively reduce the back lobe and improve the backward ratio of the antenna. Simulated results show that the reflection coefficients for the two frequency bands are less than -10 dB, covering GPS L2 (1.195 GHz~ 1.248 GHz) and GPS L1 (1.552 GHz ~ 1.599 GHz). In the entire working bandwidth, the forward to backward ratio of the antenna is greater than 30 dB.

Based on the sparse distribution of targets, compressed sensing theory has been widely used in the field of SAR imaging technology. The FMCW radar can be miniaturized. Based on the FMCW SAR imaging technology of compressed sensing, this paper analyzes the echo model of the signal and the theory of compressed sensing. The sparse representation of the signal is given, and the base matrix is constructed, and the distance of the target is obtained by OMP algorithm. The two-dimensional image of the target is obtained by combining the matching filtering method, which verifies the effectiveness of the algorithm.

Identifying the faulty class of multivariate time series is crucial for today's flight data analysis. However, most of the existing time series classification methods suffer from imbalanced data and lack of model interpretability, especially on flight data of which faulty events are usually uncommon with a limited amount of data. Here, we present a neural network classification model for imbalanced multivariate time series by leveraging the information learned from normal class, which can also learn the nonlinear Granger causality for each class, so that we can pinpoint how time series classes differ from each other. Experiments on simulated data and real flight data shows that this model can achieve high accuracy of identifying anomalous flights.

This paper mainly introduces the moving target detection experiment using a linear frequency modulated continuous wave (LFMCW) radar. First, the target detection theory is presented. Then, the designed system including radar, VCO generation circuit and DSP implement are introduced. Finally, the system is used to detect the corner reflectors, vehicles and pedestrian targets to verify the effectiveness of the algorithm. The test results show that the speed measurement accuracy is less than 2km/h, the angle measurement accuracy is less than 1 degree, and the distance measurement accuracy is less than 1 meter.

This paper presents the design of a wideband dual-polarized microstrip antenna array working at C-band. The antenna adopts a multi-layer structure using additional parasitic patch, in order to achieve better port isolation and wide-band characteristics. The 1*4 array antenna is simulated and optimized, which is fed by the coplanar microstrip lines. The results show that the VSWRs of the antenna ports are less than 2 in the frequency range of 4.9 GHz to 6.1 GHz, with a relative bandwidth of about 21.8%. The designed antenna also has a gain of over 13.4 dB at the frequency of 5.5 GHz.

The range and DOA estimation of a micro-Doppler target is verified using a linear frequency modulated continuous wave (LFMCW) radar. The signal processing algorithms are implemented by DSP. The target range results of two radars are compared, and the consistency is high. The test results indicate that the DOA error is less than 1 degree, and the range error is less than 0.25 m.

Human activity recognition (HAR) is an important task in many internet of things (IoT) applications. In recent years, significant efforts have been made towards achieving the highest possible recognition performance (accuracy and robustness) by using advanced machine learning techniques, including deep learning. However, to the best of our knowledge, the adversarial vulnerability of the Doppler sensor-based HAR systems has not been studied. In other domains such as computer vision, the vulnerability of deep learning algorithms to adversarial samples has attracted tremendous research interests in the past few years. In this work, we investigate the adversarial vulnerability of the Doppler-based human activity recognition system. Using a case study we demonstrate that the adversarial examples can significantly degrade the performance of the human activity recognition. Specifically, the basic iterative method (BIM) attack can reduce classification accuracy by as much as 85%. We also discuss different types of attacks, e.g., data poisoning attacks and potential strategies of protecting the Doppler-based HAR systems against adversarial attacks.

In this paper, we present a Rotman lens-fed Fabry-Perot resonator antennas (FPRA) for generating multi-mode OAM beams with low beam divergence. The Rotman lens works as a beamforming network in the FPRA for the multi-mode OAM operation, significantly simplifying the feed network of the antenna array. The partially reflecting metasurfaces (PRMS) composed of honeycomb complementary hexagonal patches and apertures are designed to construct the Fabry-Perot resonant cavity and enhance the directivity and meliorate the beam orientation of the generated OAM beam. The simulated and measured results verify the effectiveness of the proposed FPRA.

Aiming at addressing the problem of interactive gesture recognition between lunar robot and astronaut, a novel gesture detection and recognition algorithm is proposed. In gesture detection stage, based on saliency detection via Graph-Based Manifold Ranking (GBMR) algorithm, the depth information of foreground is added to the calculation of superpixel. By increasing the weight of connectivity domains in graph theory model, the foreground boundary is highlighted and the impact of background is weakened. In gesture recognition stage, Pyramid Histogram of Oriented Gradient (PHOG) feature and Gabor amplitude also phase feature of image samples are extracted. To highlight the Gabor amplitude feature, we propose a novel feature calculation by fusing feature in different directions at the same scale. Because of the strong classification capability and not-easy-to-fit advantage of Adaboosting, this paper applies it as the classifier to realize gesture recognition. Experimental results show that the improved gesture detection algorithm can maintain the robustness to influences of complex environment. Based on multi-feature fusion, the error rate of gesture recognition remains at about 4.2%, and the recognition rate is around 95.8%.

This paper shows a millimeter wave (MMW) antenna based on low temperature co-fired ceramic (LTCC) technology with high polarization purity. By using the LTCC technology with multilayer parasitic patches in vertical direction, a wide bandwidth and high polarization purity can be obtained for the designed antenna. The "V" type patch is designed based on LTCC multi-layer technology, which is shaped by changing the distance of patch between different layers. By building this multi-layer structure, the polarization purity and bandwidth of antenna can be improved significantly. The simulation results show that the reflection coefficient of the antenna is smaller than −10 dB from 24.28 GHz to 27.84 GHz with a relative bandwidth of about 13.6%, and the cross-polarization isolation is higher than 50 dB within the operation band. The simulated results verify that the proposed antenna can be a promising candidate for 5G communication systems.

This paper presents an improved ultra-wideband antenna design. The proposed antenna has with the characteristics of miniaturization and ultra-wideband. The proposed antenna is composed of the microstrip feed line, the dielectric substrate, and a modified ground plane with an exponential function gradation line. The size of the antenna can be reduced by half on the basis of the original antenna through introducing a mirror metal plane to achieve a miniaturized design without affecting the performance. By etching the L-shaped slot on the ground plane, the bandwidth of the antenna can also be increased. The improved antenna has good radiation characteristics with a stably directional radiation pattern in the whole working bandwidth.

Sensor based Physical Activity (PA) recognition is an imperative research topic due to its wide applications in many areas such as mobile healthcare, elderly care and personalized recommendation. However, PA recognition based on single sensor suffers from issues such as limited spatial coverage, imprecision and uncertainty. In contrast, fusion of heterogeneous sensor sources brings improved resolution, precision and robustness. In this paper, a multi-sensor fusion method based on variants of Extreme Learning Machine (ELM) is presented to improve the recognition performance in terms of speed and accuracy. The Kernel ELM, Weighted ELM, Regularized ELM are tasked directly to handle the multi-activity classification problem using the real-world test data with 46 test subjects. Their performance has been evaluated against previously published methods based on Support Vector Machines (SVM) and K-Nearest Neighbor (KNN) using a two-step subject-wise cross validation scheme. The multi-sensor fusion based Kernel ELM has shown favorable performance with 88.4% average testing accuracy, outperforming SVM by 1.1%. And for real-time PA recognition, Kernel ELM is 30% faster than SVM. The ELMs trained with the two-step, subject-wise cross validation approach, combined with feature level data fusion also led to a better cross-person generalization than SVM and KNN.

This paper presents a substrate integrated circularly polarized horn antenna based on low temperature cofired ceramic (LTCC) technology. The integrated circularly polarized horn antenna is fed by aperture-coupled strip-line and is obtained with sequential rotation technology. The impedance bandwidth is from 33.426 GHz to 40.446 GHz, and the simulated gain is about 10.2 dB at 35 GHz. The simulated results show that this circularly polarized antenna has good axial ratio with the working bandwidth, and the size is smaller than the traditional horn antenna. The proposed circularly polarized horn antenna could be applied into the wireless communications or satellite systems.

A dual-polarized transmit-array lens (TAL) is presented for the generation of dual-mode convergent orbital angular momentum (OAM) beams in C-band. The proposed TAL consists of dual-polarized metamaterial unit cells, which are composed of four stacked patches coupled by cross aperture, to complete the phase correction for both polarizations and stimulate the dual-mode convergent OAM beams. The simulated results verify the effectiveness of the proposed design.

An electronically reconfigurable unit cell with 1-bit phase resolution (0°/180°) is presented for programable transmit-array lens (TAL) in C-band. The I-bit phase resolution is obtained by using two PIN diodes integrated in the transmitting layer of the unit cell. The simulated results verify that the proposed unit can achieve good 1-bit phase tuning, with minimum insertion loss of 0.2 dB and 2-dB transmission bandwidth of about 10%.

A novel hexagonal metamaterial unit cell is presented to construct the honeycomb transmit-array lens (TAL) generating the convergent orbital angular momentum (OAM) beams in C-band. Compared with the most common square-lattice arrays, the proposed honeycomb TAL can provide a wider operating band, a larger phase-shift range and stable frequency responses as the function of angle and polarization of the incident EM wave. The simulated results verify the effectiveness of the proposed design.

A steady-state 3-D large-signal model of the beam–wave interaction (BWI) in folded-waveguide (FW) traveling-wave tubes (TWTs) has been developed. The model used a three-port network representation of the circuit and a set of discrete rays representation of the 3-D electron beam. Values of the impedance elements that characterize the model are determined by a 3-D finite-element code, high-frequency circuit simulator (HFCS). Besides, the RF fields inside the beam tunnel are represented with the digitized field profile obtained from HFCS. A new circuit attenuation model is described by introducing only one parameter. The model has been implemented in the BWISFW-3-D code. The results of the code predictions agree well with the measured data for an FW-TWT operating in the $W$ -band. The effect of uniform focusing magnetic field on BWI is also studied.

A three-dimensional (3-D) nonlinear theory model of beam–wave interaction for common traveling wave tubes (TWTs) is described. The equation governing the amplitude of electromagnetic wave is derived analogously to Poynting’s theorem. The electron dynamics are treated using the 3-D Lorentz force equations. RF space charge fields are obtained from solutions of the Helmholtz equations. In the model, the RF field profiles for cold structure are represented by the digitized RF field calculated by a finite-element software, HFSS. Because the digitized RF field can be obtained in the same way, the 3-D simulation code can be used to simulate common TWTs, such as helix TWTs, coupled-cavity TWTs, and even the folded waveguide TWTs. Results from the 3-D code are compared with those from 1-D code, experiment and the existing analytical theories for three types of slow wave structures.

Data-driven predictive analytics are in use today across a number of industrial applications, but further integration is hindered by the requirement of similarity among model training and test data distributions. This paper addresses the need of learning from possibly nonstationary data streams, or under concept drift, a commonly seen phenomenon in practical applications. A simple dual-learner ensemble strategy, alternating learners framework, is proposed. A long-memory model learns stable concepts from a long relevant time window, while a short-memory model learns transient concepts from a small recent window. The difference in prediction performance of these two models is monitored and induces an alternating policy to select, update and reset the two models. The method features an online updating mechanism to maintain the ensemble accuracy, and a concept-dependent trigger to focus on relevant data. Through empirical studies the method demonstrates effective tracking and prediction when the steaming data carry abrupt and/or gradual changes.

Data-driven predictive analytics are in use today across a number of industrial applications, but further integration is hindered by the requirement of similarity among model training and test data distributions. This paper addresses the need of learning from possibly nonstationary data streams, or under concept drift, a commonly seen phenomenon in practical applications. A simple dual-learner ensemble strategy, alternating learners framework, is proposed. A long-memory model learns stable concepts from a long relevant time window, while a short-memory model learns transient concepts from a small recent window. The difference in prediction performance of these two models is monitored and induces an alternating policy to select, update and reset the two models. The method features an online updating mechanism to maintain the ensemble accuracy, and a concept-dependent trigger to focus on relevant data. Through empirical studies the method demonstrates effective tracking and prediction when the steaming data carry abrupt and/or gradual changes.

The multi-mode OAM beams generated by a circular phased antenna array are always in divergent directions, which restrict the practical application. To relieve this drawback, the Fabry-Perot (FP) cavity is employed to meliorate the beam orientation, and enhance the directivity of the OAM array. The Fabry-Perot cavity consists of a ground plane, an air cavity and a partially reflecting surface (PRS) composed of periodic complementary circular patches and apertures. The simulated results verify the effectiveness of the proposed design.

We propose a simple but strong baseline for time series classification from scratch with deep neural networks. Our proposed baseline models are pure end-to-end without any heavy preprocessing on the raw data or feature crafting. The proposed Fully Convolutional Network (FCN) achieves premium performance to other state-of-the-art approaches and our exploration of the very deep neural networks with the ResNet structure is also competitive. The global average pooling in our convolutional model enables the exploitation of the Class Activation Map (CAM) to find out the contributing region in the raw data for the specific labels. Our models provides a simple choice for the real world application and a good starting point for the future research. An overall analysis is provided to discuss the generalization capability of our models, learned features, network structures and the classification semantics.

Gas turbine combustor anomaly detection plays a critical role in reducing operation and maintenance costs in power plant operations. One-class classification, to learn a model that properly describes the normal samples, is one of many anomaly detection approaches that have shown promising performance in real-world applications. Extreme learning machines (ELMs), a recent developed machine learning technique, have outperformed SVMs and other machine learning methods for many machine learning problems. ELMs have been originally used for binary or multiclass classification and regression as well. Using ELM for one-class classification has emerged very recently, but has not been widely explored for real world applications. In this paper, we adopt ELMs to a new application domain - industrial machine condition monitoring. More specifically, we apply one-class ELMs for more accurate anomaly detection of gas turbine combustors. We evaluate different one-class ELMs and compare them against other one-class classifiers. Using a real-world gas turbine combustor anomaly detection as the case study, we demonstrate that one-class ELMs can be more effective than other one-class classification algorithms in early detecting gas turbine combustor anomalies.

The impact-echo (IE) method is a popular non-destructive testing (NDT) technique widely used for measuring the thickness of plate-like structures and for detecting certain defects inside concrete elements or structures. However, the IE method is not effective for full condition assessment (i.e., defect detection, defect diagnosis, defect sizing and location), because the simple frequency spectrum analysis involved in the existing IE method is not sufficient to capture the IE signal patterns associated with different conditions. In this paper, we attempt to enhance the IE technique and enable it for full condition assessment of concrete elements by introducing advanced machine learning techniques for performing comprehensive analysis and pattern recognition of IE signals. Specifically, we use wavelet decomposition for extracting signatures or features out of the raw IE signals and apply extreme learning machine, one of the recently developed machine learning techniques, as classification models for full condition assessment. To validate the capabilities of the proposed method, we build a number of specimens with various types, sizes, and locations of defects and perform IE testing on these specimens in a lab environment. Based on analysis of the collected IE signals using the proposed machine learning based IE method, we demonstrate that the proposed method is effective in performing full condition assessment of concrete elements or structures.

Aircraft engine fault diagnosis plays a crucial role in cost-effective operations of aircraft engines. However, designing an engine fault diagnostic system with the desired performance is a challenging task, because of several characteristics associated with aircraft engines. Geared towards achieving the highest possible performance of fault diagnosis, this paper explores strategies on improving diagnosis performance. Specifically, we introduce flight regime mapping and a two-level multiple classifier system as means to improve classification performance. By designing a real-world aircraft fault diagnostic system, we demonstrate that the strategies adopted in this study are effective in improving the performance of aircraft engine fault diagnostic systems.

Many fault detection algorithms deal with fault signatures that manifest themselves as step changes. While detection of these step changes can be difficult due to noise and other complicating factors, detecting slowly developing faults is usually even more complicated. Tradeoffs between early detection and false positive avoidance are more difficult to establish. Often times, slow drift faults go completely undetected because the monitoring systems assume that they are ordinary system changes and some monitoring schemes may adapt to the changes. Where redundant sensors are used, a drifting sensor may cause the logic to latch on to the ldquobadrdquo sensor. Another problem may be intermittent sensors faults where the detection logic is too sluggish to recognize a problem before the sensor has returned to seemingly normal behavior. To address these classes of problems, we introduce here a set of algorithms that learns to avoid the bad sensor, thus indirectly recognizing the aberrant sensor. We combine advanced sensor validation techniques with learning. The sensor validation is inspired by fuzzy principles. The parameters of this algorithm are learned using competing optimization approaches. We compare the results from a particle swarm optimization approach with those obtained from genetic algorithms. Results are shown for an application in the transportation industry.

Under customer service agreements (CSA), engine operational data are collected and stored for monitoring and analysis. Other data sources provide damage assessments that are either provided post-maintenance or analytically assessed. This paper takes advantage of these data and investigates local fuzzy models to determine the remaining useful life (RUL) of an engine or engine component. Local fuzzy models are related to both kernel regressions and locally weighted learning. The particular local models described in this paper are not based on individual models that consider the track history of a specific engine nor are they based on a global average model that would consider the collective track history of all the engines. Instead, for a given engine or component, this local fuzzy model defines a cluster of peers in which each of these peers is a similar instance to this given engine with comparable operational characteristics; the RUL prediction for this given engine is obtained by a fuzzy aggregation of its peers’ RUL. We combine the fuzzy instance-based approach with an evolutionary framework for model tuning and maintenance. This evolutionary tuning process is repeated periodically to automatically update and improve the fuzzy models such that they can be updated to date with the latest collection of data. This fuzzy instance-based approach is applied to predicting the RUL of a commercial engine validated with post-maintenance assessment.

The design of a V-band helix-folded waveguide (H-FWG) cascaded traveling wave tube (TWT) is presented. In this cascaded structure, a high-gain helix TWT is carefully designed as a first stage amplifier followed by a FWG TWT to obtain high power. Simulations predict that a peak power of 800W with saturated gain of 55 dB from 55GHz to 60 GHz can be achieved.

Low strain pile integrity testing (LSPIT), due to its simplicity and low cost, is one of the most popular NDE methods used in pile foundation construction. While performing LSPIT in the field is generally quite simple and quick, determining the integrity of the test piles by analyzing and interpreting the test signals (reflectograms) is still a manual process performed by experienced experts only. For foundation construction sites where the number of piles to be tested is large, it may take days before the expert can complete interpreting all of the piles and delivering the integrity assessment report. Techniques that can automate test signal interpretation, thus shortening the LSPIT’s turnaround time, are of great business value and are in great need. Motivated by this need, in this paper, we develop a computer-aided reflectogram interpretation (CARI) methodology that can interpret a large number of LSPIT signals quickly and consistently. The methodology, built on advanced signal processing and machine learning technologies, can be used to assist the experts in performing both qualitative and quantitative interpretation of LSPIT signals. Specifically, the methodology can ease experts’ interpretation burden by screening all test piles quickly and identifying a small number of suspected piles for experts to perform manual, in-depth interpretation. We demonstrate the methodology’s effectiveness using the LSPIT signals collected from a number of real-world pile construction sites. The proposed methodology can potentially enhance LSPIT and make it even more efficient and effective in quality control of deep foundation construction.

Over the past few decades, application of artificial neural networks (ANN) to time-series forecasting (TSF) has been growing rapidly due to several unique features of ANN models. However, to date, a consistent ANN performance over different studies has not been achieved. Many factors contribute to the inconsistency in the performance of neural network models. One such factor is that ANN modeling involves determining a large number of design parameters, and the current design practice is essentially heuristic and ad hoc, this does not exploit the full potential of neural networks. Systematic ANN modeling processes and strategies for TSF are, therefore, greatly needed. Motivated by this need, this paper attempts to develop an automatic ANN modeling scheme. It is based on the generalized regression neural network (GRNN), a special type of neural network. By taking advantage of several GRNN properties (i.e., a single design parameter and fast learning) and by incorporating several design strategies (e.g., fusing multiple GRNNs), we have been able to make the proposed modeling scheme to be effective for modeling large-scale business time series. The initial model was entered into the NN3 time-series competition. It was awarded the best prediction on the reduced dataset among approximately 60 different models submitted by scholars worldwide.

A steady-state large signal model of the beam-wave interaction in a folded or serpentine waveguide traveling-wave tube is described. The model used a three-port network representation of the circuit and a 1-D disk model of the beam. Values of the admittance elements that characterize the model are identified with a 3-D finite element code, HFSS. This model can be used to simulate the beam-wave interaction in other types of slow-wave structures.

In this paper we cast a vision for Security Prognostics (SP) for critical systems, promoting the view that security related protections would be well served to integrate fully with Monitoring and Diagnostics (M&D) systems that assess the health of complex assets and systems. To detect complex Cyber threats we propose combining system parameters already in use by M&D systems for Prognostics and Health Monitoring (PHM) with security parameters. Combining system parameters used by M&D to detect non-malicious faults with the system parameters used by security schemes to detect complex Cyber threats will improve: (a) accuracy of PHM (b) security of M&D, and (c) availability and safety of critical systems. We also introduce the notion of Remaining Secure Life (RSL), assessed based on the propagation of “security damage,” to create the prospect for Security Prognostics. RSL will assist in the selection of appropriate response(s), based on breach or compromise to security component's and potential impact on system operation. An example of M&D data is provided which is normally associated with non-malicious faults providing input to detect Malware execution through time series monitoring.

Feature dimensionality reduction is a critical task in various machine learning applications including prognostics and health management (PHM) applications. Linear transformations, most popularly principal component analysis (PCA) and linear discriminant analysis (LDA), are the most widely-used methods for feature dimensionality reduction. For classification problems, LDA, being a supervised linear transformation that aims at maximally retaining class discriminant information, is generally considered to be a better method than PCA, an unsupervised method. However, LDA suffers from the singularity or small sample size problem. Attempting to address this problem, in this paper we propose a cluster-based LDA (cLDA) for feature dimensionality reduction. It first partitions features in distinct clusters and then performs cluster-wise LDA transformation. We demonstrate the effectiveness of the proposed cLDA on reducing the number of features by using a real-world PHM application - partial discharge diagnosis.

The Gamma Test has attracted the attention of many researchers in the nonlinear modeling field, especially with Artificial Neural Networks. In theory, the test should provide a modeler with valuable information to find the best input variables without extensive model development for each potential input combination. However, it has been found that the Gamma Test does not always point to the best input combination as validated by the cross validation method. This paper presents a study of using the generalized regression neural network (GRNN) to estimate evaporation. Both the Gamma Test and cross validation are used to find the best model input combination. It has been found that the Gamma Test is not able to identify the best input variables, but the best result is included in the top Gamma value group. The standard error has very valuable information for choosing the group members. This demonstrates that the Gamma Test is still a valuable tool in significantly reducing the modeling workload. The reason for this phenomenon is discussed under the relationship between the Gamma estimate and its stand error. Further research is still needed to explore this relationship in more efficient model input selections.

The hierarchical mixtures of experts (HME) model is a flexible model that stochastically partitions the feature space into sub-regions, in which simple surfaces can be fitted to data. However, there are issues with model selection during the learning of the HME model using Expectation-Maximization (EM) inference. In addition, the EM algorithm suffers from the well-known problem of local minima. In this paper, we present a hybrid generative-discriminative approach that inherits the flexibility of the HME while decomposing the learning process into a few simple steps. The proposed algorithm solves the model-selection problem and empirical experiments on public benchmark datasets show its advantages in classification accuracy and efficiency.

Polymorphic attacks threaten to make many intrusion detection schemes ineffective [15] . In order to address the threat of advanced attacks, model based techniques are required. In this paper we improve our Grammar Based Modeling techniques [1]–[5] to be more resilient to attacks that change in form by using advanced classification techniques. Similarity distances from known models are input as features input to Support Vector Machines and other advanced classification techniques to provide improved classification performance. Results indicate promise for intrusion detection and response against polymorphic attack with minimal false alarms.

Gaussian process (GP), as one of the cornerstones of Bayesian non-parametric methods, has received extensive attention in machine learning. GP has intrinsic advantages in data modeling, given its construction in the framework of Bayesian hieratical modeling and no requirement for a priori information of function forms in Bayesian reference. In light of its success in various applications, utilizing GP for time-series forecasting has gained increasing interest in recent years. This paper is concerned with using GP for multiple-step-ahead time-series forecasting, an important type of time-series analysis. Utilizing a large number of real-world time series, this paper evaluates two different GP modeling strategies (direct and recursive) for performing multiple-step-ahead forecasting.

Jet engine gas path fault diagnosis is not only important in modern condition-based maintenance of aircraft engines, but also a challenging classification problem. Exploring more advanced classification techniques for achieving improved classification performance for gas path fault diagnosis, therefore, has been increasingly active in recent years in PHM community. To that end, in this paper, we apply a recently developed dynamic fusion scheme to gas path fault diagnosis. Through designing a real-world gas path fault diagnostic system, we demonstrate that dynamic fusion of multiple classifiers can be effective in improving classification performance of gas path diagnosis.

Locomotives are complex electromechanical systems. Continuously monitoring the health state of locomotives is critical in modern cost-effective maintenance strategy. A typical locomotive is equipped with the capability to generate fault messages or incidents based on logical rules in the control system. In the mean time, sensor readings and operational state variables are also collected. The goal is to detect faults early to provide lead-time for maintenance actions and trip planning based on the collected fault log and parametric data. In this paper, we present a model-based anomaly detection strategy. In this method, the inputs-outputs relationship of a locomotive subsystem is modeled using a neural network model based on normal operational data. The residuals between measurements and model outputs are calculated. A reasoning module based on these multiple residuals is used to generate an overall health indicator of the subsystem at each instance of times, which is further used to determine whether the subsystem is abnormal. Statistical testing, Gaussian mixture model and support vector machine are used to generate this healthy index and their performances are compared. We demonstrate the effectiveness of the anomaly detection strategy using real-world operational data from locomotives.

Two industrial deployments of multi-criteria decision-making systems at General Electric are reviewed from the perspective of their multi-criteria decision-making component similarities and differences. The motivation is to present a framework for multi-criteria decision-making system development and deployment. The first deployment is a financial portfolio management system that integrates hybrid multi-objective optimization and interactive Pareto frontier decision-making techniques to optimally allocate financial assets while considering multiple measures of return and risk, and numerous regulatory constraints. The second deployment is a power plant management system that integrates predictive modeling based on neural networks, optimization based on multi-objective evolutionary algorithms, and automated decision-making based on Pareto frontier techniques. The integrated approach, embedded in a real-time plant optimization and control software environment dynamically optimizes emissions and efficiency while simultaneously meeting load demands and other operational constraints in a complex real-world power plant

Both experimental and theoretical studies have proved that classifier fusion can be effective in improving overall classification performance. Classifier fusion can be performed on either score (raw classifier outputs) level or decision level. While tremendous research interests have been on score-level fusion, research work for decision-level fusion is sparse. This paper presents a particle swarm optimization based decision-level fusion scheme for optimizing classifier fusion performance. Multiple classifiers are fused at the decision level, and the particle swarm optimization algorithm finds optimal decision threshold for each classifier and the optimal fusion rule. Specifically, we present an optimal fusion strategy for fusing multiple classifiers to satisfy accuracy performance requirements, as applied to a real-world classification problem. The optimal decision fusion technique is found to perform significantly better than the conventional classifier fusion methods, i.e., traditional decision level fusion and averaged sum rule

Diagnostics and prognostics is particularly challenging in systems with a restricted suite of sensors; e.g., in aircraft engines where harsh operating conditions, weight considerations, and regulatory concerns limit the number of sensors. In this paper, we investigate anomaly detection techniques subject to these constraints. Specifically, we use as input to these techniques only controller-generated, log-data for the system. While such log-data is not designed to carry predictive information related to system health, we show that it is possible to extract early warning signals related to the failure of the system by looking for the presence or onset of anomalous or novel patterns in the log-data. We present preliminary results obtained by the application of this approach to some complex systems. We also provide a roadmap for extending this approach by the incorporation of minimal amount of system-specific knowledge of the kind that is typically available for complex systems. This extension is expected to strengthen the applicability of the approach to diagnostic and prognostic analysis at the level of the system components, as well as to the estimation of the root cause of a detected system anomaly.Copyright © 2007 by General Electric Company

This paper describes how damage propagation can be tracked and modeled for a range of fault modes in some modules of commercial high bypass aircraft engines. To that end, response surfaces of all sensors are generated via a thermo-dynamical simulation model for the engine (cycle deck) as a function of variations of flow and efficiency of the modules of interest. These surfaces are normalized and superimposed. Next, sensor readings are matched to those surfaces and -using an optimization approach -the corresponding flow and efficiency pair is found that best explains the sensor data. This flow and efficiency pair is then compared to previous pairs and the direction of the change as well as the rate of change is determined. The whole trajectory is then projected into the time domain. An extrapolation of the curve to the limit (which is established via operational margins) yields the remaining life. In a backward mode, the extrapolated curve is discretized and estimated future flow and efficiency pairs are retrieved. These pairs are then input to the cycle deck to produce future anticipated sensor readings as well as confirmatory trips of operational margins. Changes of the future sensor readings with real readings are used to adjust the remaining life calculations. The method is demonstrated on time series of historical engine faults.

This paper describes an approach to estimate the deterioration level of aircraft engines using engine monitoring data and a physics-based engine model. The estimation process is carried out by a neural network, which is trained by data generated using a physical-based engine model complemented with an empirically derived engine deterioration model. The deterioration model allows manipulation of several engine health parameters, such as module efficiency and flow capacity, to simulate engine deterioration. Simulated sensor outputs are used to build independent transfer functions relating the sensor values to a deterioration level. A calibration model corrects the sensor readings to a reference condition so that the effect of variation of operating condition is minimized. The proposed approach can be used to assess engine deterioration level in real time. The proposed deterioration estimation approach is validated using real-world engine data.Copyright © 2007 by General Electric Company