A central challenge in wind turbine health monitoring is the scarcity of real-world data due to limited instrumentation, leading researchers to rely on simulation models that often suffer from reduced fidelity. However, even within simulation environments, discrepancies arise because of modeling assumptions, and configuration fidelities, creating domain gaps that limit the transferability of learned representations. To investigate domain translation under controlled conditions, this project explores the use of generative artificial intelligence, specifically cycle-consistent generative adversarial networks (CGANs), to bridge the gap between OpenFAST simulation models representing 1.5 MW and 5 MW wind turbines. A physics-informed CGAN architecture is introduced, where a simplified turbine tower dynamics model is incorporated into the training loss to ensure physically consistent outputs. Quantitative results showed moderate to high agreement in frequency-domain features. Incorporating the physics-informed loss function improved the R2 values by 30%, reduced the RMSE from 1.39 to 1.1 m/s2, and reduced training time by 82%. Furthermore, under increased turbulence intensity (IEC Category A), the RMSE remained stable at approximately 1.1 m/s2. While the present study is entirely simulation-based, it establishes a pipeline for evaluating physics-informed generative domain translation, which may serve as a foundation for future simulation-to-reality validation studies.

Intelligent fault diagnosis utilizing deep learning algorithms is currently a topic of great interest. When developing a new Convolutional Neural Network (CNN) architecture to address machine diagnosis problem, it is common to use a deep model, with many layers, many feature maps, and large kernels. These models are capable of learning complex relationships and can potentially achieve superior performance on test data. However, not only does a large network potentially impose undue computational complexity for training and eventual deployment, it may also lead to more brittleness—where data outside of the curated dataset used in CNN training and evaluation is poorly handled. Accordingly, this paper will investigate a methodical approach for identifying a quasi-optimal CNN architecture to maximize robustness when a model is trained under one set of operating conditions, and deployed under a different set of conditions. Optuna software will be used to optimize a baseline CNN model for robustness to different rotational speeds and bearing Model #’s. To further improve the network generalization capabilities, this paper proposes the addition of white Gaussian noise to the raw vibration training data. Results indicate that the number of trainable weights and associated multiplications in the optimized model were reduced by almost 95% without jeopardizing the network classification accuracy. Additionally, moderate Additive White Gaussian Noise (AWGN) improved the model adaptation capabilities.

Data-driven fault diagnosis utilizing deep learning algorithms is currently a topic of great interest. Without proper training, data-driven models usually fail to generalize on operating conditions different from the ones used in the training set. The majority of domain adaptation research for machinery fault diagnosis focuses on the transfer between limited working conditions for the same machine. In real-life applications, machines operate under a wide range of operating conditions and the data are mostly available for healthy conditions with seldom failures. Hence, models generated from controlled experiments do not usually generalize well under substantial domain shifts. To address this issue, this paper proposes a semi-unsupervised domain adaptation approach for cross-machine fault diagnosis which integrates model optimization and Generative Adversarial Networks (GANs) to bridge the gap between source and target domains. Experiments of transferring between two bearing data sets show that the proposed method is able to effectively train an optimized model that generalizes on both the source and target domains, and train a generator that learns the source domain probability distribution to substitute for larger domain shifts.

Intelligent fault diagnosis utilizing deep learning algorithms has been widely investigated recently. Although previous results demonstrated excellent performance, features learned by Deep Neural Networks (DNN) are part of a large black box. Consequently, lack of understanding of underlying physical meanings embedded within the features can lead to poor performance when applied to different but related datasets i.e. transfer learning applications. This study will investigate the transfer learning performance of a Convolution Neural Network (CNN) considering 4 different operating conditions. Utilizing the Case Western Reserve University (CWRU) bearing dataset, the CNN will be trained to classify 12 classes. Each class represents a unique differentfault scenario with varying severity i.e. inner race fault of 0.007”, 0.014” diameter. Initially, zero load data will be utilized for model training and the model will be tuned until testing accuracy above 99% is obtained. The model performance will be evaluated by feeding vibration data collected when the load is varied to 1, 2 and 3 HP. Initial results indicated that the classification accuracy will degrade substantially. Hence, this paper will visualize convolution kernels in time and frequency domains and will investigate the influence of changing loads on fault characteristics, network classification mechanism and activation strength.

We explore the use of electrochemical sensors toward the creation of a personal carbon monoxide monitoring device. Our ultimate goal is the development of an inexpensive, highly integrated mobile health (mHealth) device for improving the outcomes, and decreasing the treatment costs, for smoking cessation therapy among low socioeconomic status (SES) patients. We will combine a number of proven modalities into a single device, which we call the T-COM. This hand-held personal device will integrate a breath carbon monoxide sensor, multilingual motivational text and audio messaging, gamification features and built-in wireless network connectivity. The requirements relating to measuring the concentration of CO in exhaled breath are especially challenging. They include the following – low cost, minimal thickness, robust to a range of ambient temperatures and humidity, as well as being minimally affected by the high humidity in breath. Concentrations as low as a few ppm of CO must be measurable, with an accuracy of a few ppm. The device must not require field calibration. The sensor must be ready for use at virtually any time with no warm-up period, and must register an accurate measurement in under ten seconds. The required support circuitry must be available with similar constraints on cost, size and accuracy. We experimented with several commercially available CO sensors and circuit topologies, and we report on our findings. In addition we have looked into methods for controlling the level of humidity that reaches the sensor, without otherwise affecting the measurement accuracy. Our results to date suggest that a practical solution is achievable, although further improvements would be valuable. Figure 1

In recent years, advancements in machine learning capabilities have allowed it to see widespread adoption for tasks such as object detection, image classification, and anomaly detection. However, despite their promise, a limitation lies in the fact that a network’s performance quality is based on the data which it receives. A well-trained network will still have poor performance if the subsequent data supplied to it contains artifacts, out of focus regions, or other visual distortions. Under normal circumstances, images of the same scene captured from differing points of focus, angles, or modalities must be separately analysed by the network, despite possibly containing overlapping information such as in the case of images of the same scene captured from different angles, or irrelevant information such as images captured from infrared sensors which can capture thermal information well but not topographical details. This factor can potentially add significantly to the computational time and resources required to utilize the network without providing any additional benefit. In this study, we plan to explore using image fusion techniques to assemble multiple images of the same scene into a single image that retains the most salient key features of the individual source images while discarding overlapping or irrelevant data that does not provide any benefit to the network. Utilizing this image fusion step before inputting a dataset into the network, the number of images would be significantly reduced with the potential to improve the classification performance accuracy by enhancing images while discarding irrelevant and overlapping regions

Devices associated with Internet of Things are typically constrained in their resources and do not have the computational power necessary to analyze their input and detect anomalies that occur. Smart devices or and environmental sensors that measure temperature, air quality, or seismic activity are all built for specific purposes with minimal resources and often do not have enough security in place to protect against infiltration or detect abnormal behavior. Additionally, because these devices and sensors are typically always connected and transmit constant data in near real-time, the high dimensionality of the raw readings are extremely computationally intensive to analyze. A possible solution to reduce the dimensionality of the data while also extracting the most significant features is to use multivariate analysis techniques such as Principal Component Analysis. PCA is a method of multivariate analysis meant to reduce the size of matrices while not only keeping the most significant variables but also learning the interactions between them. In this paper, we propose exploring anomaly detection in IoT using multivariate analysis techniques to reduce the dimensionality of sensor input to reduce the computational complexity of analysis and learning the most significant variables. While the normal conditions of sensor data are often readily available, the size of the data makes it difficult to precisely determine instances of targeted anomalies. In this study, PCA is used to analyze the available features of the data and from them can determine the sensors under normal conditions. Once the normal conditions are determined, outliers which constitute anomalies can be determined through techniques such as Mahalanobis distance to determine the variance of each observation from the normal distribution. Our work can also be expanded to use other methods of dimensionality reduction and feature extraction such as t-Distributed Stochastic Neighbor Embedding.

The problem of proper identification of the color space associated with digital luma and chroma data has been widely reported by video processing professionals.The problem arises from confusing, and sometimes conflicting, statements regarding color space usage and description.Although standards allow for the carriage of descriptive metadata regarding color space, some applications do not require the presence of such metadata.Those standards typically recite assumptions on color space that should be followed in the absence of embedded color space descriptors.The unfortunate result of this approach is a state of confusion in the industry, and consequently the possibility of errors in rendering the output of decoded video and images.Our work represents the first known attempt to determine color space directly from luma/chroma pixel data, and provides an alternative to sole reliance on potentially missing or incorrect metadata, or weakly followed defaults.Although a color space is defined by many parameters, such as primary chromaticity, transfer characteristics and matrix coefficients, we chose to focus on determining which standard for matrix coefficients had been used to create a given luma/chroma image.We addressed the problem via deep convolutional neural networks (DCNNs), trained on millions of images.Our results are encouraging, and suggest that DCNNs can be used to solve this ill-posed problem.

In recent years, an increasing number of devices are being connected to the Internet that encompasses more than just traditional devices. Internet of Things integrates real-world sensors such as smart devices or environment sensors with the Internet allowing for real}-time monitoring of conditions. IoT devices are often constrained in their resources as the sensors involved are designed for specific purposes. Due to these constraints, typical methods of intrusion and anomaly detection cannot be used. Also, due to the amount of raw input data from these sensors, detecting anomalies among the noise and other background data can be computationally intensive. A possible solution to this is by using machine learning models that are trained on both normal and abnormal behavior to detect when anomalies occur. By using techniques such as autoencoders, models can be trained that have learned normal operating conditions. In this study, we explore the use of machine learning techniques such as autoencoders to effectively handle the high dimensionality of sensor datasets while consequently learning their normal operating conditions. Autoencoders are a type of neural network which attempts to reconstruct its input data by combining two NNs, an encoder, and a decoder network. The encoder learns its input by encoding it into a lower-dimensional space while capturing the interactions and correlations between variables. In this paper, we explore the use of techniques such as autoencoders to create a lower-dimensional representation of high dimensional sensor input. Autoencoders encode the data allowing for the network to learn the interactions between parameters in normal conditions which when reconstructed with the decoder represents non-anomalous behavior. When data containing anomalies are input into the network errors will occur within the reconstruction. The error between the reconstructions can be measured using a distance function to determine if an observation is anomalous.

Multi-resolution image decomposition transforms are a popular approach to current image processing problems such as image fusion, noise reduction, and deblurring. Over the past few decades, new algorithms have been developed based on the wavelet transform to remedy its directional and shift invariant shortcomings (undecimated discrete wavelet transform is shift invariant). This study provides a comprehensive analysis of multi-focus image fusion techniques using six different multi-resolution decomposition transforms to determine the optimal transform for an image fusion application. The transforms investigated are the wavelet, double-density wavelet, dual-tree wavelet, curvelet, contourlet, and bandelet. Furthermore, for each transform, seven transform coefficient fusion algorithms are analyzed and the performance is evaluated using eight no-reference objective image fusion metrics. The transforms and algorithms selected are applied to a data set that has 27 pairs of multi-focus source images used for image fusion. By bringing together the transforms, fusion algorithms, and metrics presented in this study as derived separately from different authors, the study seeks to compare these methods. However, a complete comparison amongst the different transforms, algorithms, and metrics has not been found in any of the existing literature. Our goal is to provide useful insight into their applications in image fusion. The summary of the aggregated results indicates that (1) the curvelet is the most robust transform, (2) down-up and linear are the most effect methods of fusion, and (3) Tsallis is the best metric for multi-focus image fusion.

The Convolutional Neural Network (CNN) is a class of deep artificial neural network and has recently gained special attention after demonstrating breakthrough accuracies in various classification tasks.CNNs have shown remarkable performance in machine vision tasks such as image classification, natural language processing and speech recognition.There is evidence that the depth of a CNN plays an important role in performance of CNNs.However, we investigated the feasibility of improving the performance of shallow networks via fusion of the features computed by a homogenous and heterogeneous set of pre-trained networks.We also explored a recently developed framework called the Generative Adversarial Network (GAN), in which we simultaneously train two models, a Generator and a Discriminator.The Generator attempts to produce data that mirrors the probability distribution of the "true" dataset.The Discriminator is trained to distinguish between the true dataset and the counterfeit data produced by the Generator.Our work involves the application of a GAN for generation and fine tuning of synthetic data to be used to train a deep CNN.Specifically, we investigate the use of a synthetic data generator along with a GAN to create an unlimited quantity of labeled training data, without the need for hand-labeling images.We apply this technique to the detection and localization of various vehicles.We attempt to distinguish between military trucks and other types of vehicles.A successful outcome could lead to improvements in addressing security threats rapidly, and cost-effectively.We also investigate an alternative method for generating synthetic data, the Variational Auto-Encoder (VAE).Variational auto-encoders are trained to encode then decode input vectors and can also be useful for generating new training data.VAEs are capable of dimensionality reduction and synthesizing data.Finally, we evaluate our multiplicative fusion method compared to the fusion methods that we investigated previously.

Deep convolutional neural networks have been successfully deployed by large, well-funded teams, but their wider adoption is often limited by the cost and schedule ramifications of their requirement for massive amounts of labeled data. We address this problem through the use of a parameterized synthetic image generator. Our approach is particularly novel in that we have been able to fine tune the generator’s parameters through the use of a generative adversarial network. We describe our approach, and present results that demonstrate its potential benefits. We demonstrate the PSIG-GAN by creating images for training a DCNN to detect the existence and location of weeds in lawn grass.

In recent years, deep convolution neural networks (DCNNs) have won many contests in machine learning, object detection, and pattern recognition. Furthermore, deep learning techniques achieved exceptional performance in image classification, reaching accuracy levels beyond human capability. Malware variants from similar categories often contain similarities due to code reuse. Converting malware samples into images can cause these patterns to manifest as image features, which can be exploited for DCNN classification. Techniques for converting malware binaries into images for visualization and classification have been reported in the literature, and while these methods do reach a high level of classification accuracy on training datasets, they tend to be vulnerable to overfitting and perform poorly on previously unseen samples. In this paper, we explore and document a variety of techniques for representing malware binaries as images with the goal of discovering a format best suited for deep learning. We implement a database for malware binaries from several families, stored in hexadecimal format. These malware samples are converted into images using various approaches and are used to train a neural network to recognize visual patterns in the input and classify malware based on the feature vectors. Each image type is assessed using a variety of learning models, such as transfer learning with existing DCNN architectures and feature extraction for support vector machine classifier training. Each technique is evaluated in terms of classification accuracy, result consistency, and time per trial. Our preliminary results indicate that improved image representation has the potential to enable more effective classification of new malware.

We present a novel approach to video coding that can dramatically reduce decoder DRAM bandwidth requirements while incurring a minimal reduction in compression efficiency, and which may yield an increase in compression efficiency under certain circumstances. Our approach is based on the principle that areas of video pictures where there is high motion are typically captured with significant blur along the direction of motion. This blur permits the judicious use of reduced resolution reference pictures for prediction without significantly reducing prediction quality. Our approach makes it feasible to limit worst-case DRAM bandwidth through the use of reasonably sized on-chip caches for pixel data, which can lead to provably compliant real-time behavior. The reductions in DRAM bandwidth can be expected to yield commensurate reductions in DRAM power dissipation, and consequently improvements in battery life for mobile devices. Although we concentrate our analysis on decoders, our approach can yield even greater advantages for encoders, which require additional bandwidth for motion estimation. We show that the compression efficiency of our encoder can approach that of a standard reference encoder on natural video sequences, but that it may fall short by a modest amount on pathological synthetic sequences.

Prior to the advent of ITU-R Recommendation BT.709 the overwhelming majority of compressed digital video and imagery used the colorspace conversion matrix specified in ITU-R Recommendation BT.601. The introduction of high definition video formats led to the adoption of Rec. BT.709 for use in colorspace conversion by new systems, and this resulted in confusion in the industry. Specifically, video decoders may not be able to determine the correct matrix to use for converting from the luma/chroma representation used for coding, to the Red-Green-Blue representation needed for display. This confusion has led to a situation where some viewers of decompressed video streams experience subtle, but noticeable, errors in coloration. We have successfully developed and trained a deep convolutional neural network to address this heretofore unsolved problem. We obtained outstanding accuracy on ImageNet data, and on YouTube video frames, and our work can be expected to lead to more accurate color rendering delivered to users of digital imaging and video systems.

In recent years, the accurate characterization of the boost phase of a missile's flight has become a more challenging and prominent research topic as the noise level is extremely large relative to the quantity of interest. Reconstructing the boost phase acceleration profile of a ballistic missile from state observation is of interest to the technical intelligence community, ballistic missile defense, as well as the missile warning community. There are methods available such as Tikhonov regularization if the noise level is not too large. However, if the noise environment is very high most algorithms will perform poorly. In this paper, we explore the problem of estimating the thrust of a missile from very noisy estimates of its position over time by using wavelet techniques. Several wavelet basis functions and multi-resolution methods are explored to yield the most effective solution to this problem. These techniques have been successfully used on actual rocket-launch data in the past. Our method can be applied to US boost-phase missile defense such as protection of US homeland against nuclear attacks, other weapons of mass destructions or conventional ballistic missile attacks, military bases, and protecting US allies and partners.

We present a novel approach to video coding that can dramatically reduce decoder DRAM bandwidth requirements while incurring a minimal reduction in compression efficiency, and which may yield an increase in compression efficiency under certain circumstances. Our approach is based on the principle that areas of video pictures where there is high motion are typically captured with significant blur along the direction of motion. This blur permits the judicious use of reduced resolution reference pictures for prediction without significantly reducing prediction quality. Our approach makes it feasible to limit worstcase DRAM bandwidth through the use of reasonably sized onchip caches for pixel data, which can lead to provably compliant real-time behavior. The reductions in DRAM bandwidth can be expected to yield commensurate reductions in DRAM power dissipation, and consequently improvements in battery life for mobile devices. Although we concentrate our analysis on decoders, our approach can yield even greater advantages for encoders, which require additional bandwidth for motion estimation. We show that the compression efficiency of our encoder can approach that of a standard reference encoder on natural video sequences, but that it may fall short by a modest amount on pathological synthetic sequences.

Deep convolutional neural networks (DCNN's) have shown great value in approaching highly challenging problems in image classification. Based on the successes of DCNNs in scene classification and object detection and localization it is natural to consider whether they would be effective for much simpler computer vision tasks. Our work involves the application of a DCNN to the relatively simple task of detecting weeds in lawn grass. We looked at the effects of the choice of CNN hyper-parameters on accuracy and training convergence behavior. In order to obtain a large labeled set of interesting data we generated realistic synthetic imagery. Since our problem is somewhat constrained we were able to run thousands of training experiments and do accurate estimation of the probability density function of the convergence rate. Our results suggest that the use of realistic synthetic imagery is an effective approach for training DCNNs, and that very small DCNNs can be effective for simple image recognition tasks.

One challenge in a video surveillance system is the data rate required to represent digital video. Accordingly, the use of lossy video compression at a compression ratio of 100:1, or higher, is an essential part of any distributed live video system. The ensuing distortion can interfere with the goals of surveillance by confounding both human analysis and computer vision based processing. This paper investigates the interaction between the video coding layer and target detection, and proposes methods for improving overall system effectiveness. Previous related research has focused on joint optimization of the video coding layer where several streams share the same bandwidth. Our work is distinguished from prior studies in several area: we use Gradual Decoder Refresh, rather than the traditional GOP, to enable low delay and similarly avoid the use of B frames, which necessitate frame reordering. We extend the previous work by providing the ROC curves for the detection of foreground object motion, as a function of the quantization parameter. We also consider the H.265 video coding standard, in addition to H.264. We note some surprising findings. We show that H.265 can significantly underperform H.264 in terms of Area Under Curve vs. Bitrate, and that it is possible to produce large "false alarm" blobs for moving object detection, even for a stationary, relatively noise-free source coded at low QP.