Data-driven Support Vector Machine Research Articles

Data-driven modelling of the multiaxial yield behaviour of nanoporous metals

Nanoporous metals, built out of complex ligament networks, can be produced with an additional level of hierarchy. The resulting complexity of the structure makes modelling of the mechanical behaviour computationally expensive and time consuming. In addition, multiaxial stresses occur in the higher hierarchy ligaments. Therefore, knowledge of the multiaxial material behaviour, including the 6D yield surface, is required. Surrogate models, predicting the mechanical behaviour of the lower level of hierarchy, represented by finite element beam models, are a promising approach to overcome such challenges, when existing analytical models are not able to describe the material behaviour. Therefore, as a first step, we studied the elastic behaviour and the yield surfaces of representative volume elements with idealised diamond and Kelvin structure in finite element simulations. The yield surfaces showed pronounced anisotropy and could not be described by the Deshpande-Fleck model for isotropic solid foams. Instead, we used data-driven and hybrid artificial neural networks, as well as data-driven support vector machines and compared them regarding their potential for the prediction of yield surfaces. All considered methods were well suited and resulted in relative errors <4.5%. Support vector machines showed the best generalisation and accuracy in 6D stress space and are suitable for extrapolation outside the range of training data.

Small Low-Contrast Target Detection: Data-Driven Spatiotemporal Feature Fusion and Implementation.

Detecting small low-contrast targets in the airspace is an essential and challenging task. This article proposes a simple and effective data-driven support vector machine (SVM)-based spatiotemporal feature fusion detection method for small low-contrast targets. We design a novel pixel-level feature, called a spatiotemporal profile, to depict the discontinuity of each pixel in the spatial and temporal domains The spatiotemporal profile is a local patch of the spatiotemporal feature maps concatenated by the spatial feature maps and temporal feature maps in channelwise, which are generated by the morphological black-hat filter and a ghost-free dark-focusing frame difference methods, respectively. Instead of the handcrafted feature fusion mechanisms in previous works, we use the labeled spatiotemporal profiles to train an SVM classifier to learn the spatiotemporal feature fusion mechanism automatically. To speed up detection for high-resolution videos, the serial SVM classification process on central processing units (CPUs) is reformed as parallel convolution operations on graphics processing unit (GPUs), which exhibits over 1000+ times speedup in our real experiments. Finally, blob analysis is applied to generate final detection results. Elaborate experiments are conducted, and experimental results demonstrate that the proposed method performs better than 12 baseline methods for the small low-contrast target detection. The field tests manifest that the parallel implementation of the proposed method can realize real-time detection at 15.3 FPS for videos at a resolution of 2048×1536 and the maximum detection distance can reach 1 km for drones in sunny weather.

Comparative study of very short-term flood forecasting using physics-based numerical model and data-driven prediction model

Reliable hourly flood forecasting using weather radar rainfall data for early warning system is essential for reducing natural disaster risk during extreme typhoon events. This study proposed a novel approach integrated with physics-based WASH123D and HEC-HMS models to forecast 1 h ahead flood level in the Fengshan Creek basin, northern Taiwan. The comparison was done with data-driven support vector machine (SVM) model, and performances were assessed by using statistical indicators (root mean square error, correlation coefficient, the error of time to peak flood level, the error of peak flood). Four typhoons and two plum rain events (with 620 data sets) were selected for the process of model calibration and validation. The model performs better when it used quantitative precipitation estimate radar data rather than rain gauge data. Results of using 1 h ahead quantitative precipitation forecast (QPF) as input for flood forecasting were encouraging but not feasible to use directly for early flood warning system due to errors in peak flood levels and timing. Therefore, the improvement in accuracy of 1 h ahead flood forecasting was done using physics-based approach and SVM model. The systematic comparison revealed that the SVM model is an attractive way out to improve the accuracy of QPF forecasted flood levels but unable to fully describe the flood level patterns in terms of timings and flood peaks, while the results obtained by the physics-based approach were accurate and much better than the SVM model. The approach fully described the physics of hydrograph patterns and outputs have exactly the same 1 h ahead predictions, in excellent agreement with observations. The reliable and accurate reflections of timing and amount of flood peaks in all selected typhoons by a newly developed physics-based approach with its operational nature are recommended to use by the government in the future for early warning to reduce the flood impacts during typhoon events.

Bringing statistical learning machines together for hydro-climatological predictions - Case study for Sacramento San joaquin River Basin, California

Study regionSacramento San Joaquin River Basin, California Study focusThe study forecasts the streamflow at a regional scale within SSJ river basin with largescale climate variables. The proposed approach eliminates the bias resulting from predefined indices at regional scale. The study was performed for eight unimpaired streamflow stations from 1962–2016. First, the Singular Valued Decomposition (SVD) teleconnections of the streamflow corresponding to 500 mbar geopotential height, sea surface temperature, 500 mbar specific humidity (SHUM500), and 500 mbar U-wind (U500) were obtained. Second, the skillful SVD teleconnections were screened non-parametrically. Finally, the screened teleconnections were used as the streamflow predictors in the non-linear regression models (K-nearest neighbor regression and data-driven support vector machine). New hydrological insightsThe SVD results identified new spatial regions that have not been included in existing predefined indices. The nonparametric model indicated the teleconnections of SHUM500 and U500 being better streamflow predictors compared to other climate variables. The regression models were capable to apprehend most of the sustained low flows, proving the model to be effective for drought-affected regions. It was also observed that the proposed approach showed better forecasting skills with preprocessed large scale climate variables rather than using the predefined indices. The proposed study is simple, yet robust in providing qualitative streamflow forecasts that may assist water managers in making policy-related decisions when planning and managing watersheds.

Are the States United? An Analysis of U.S. Hotels’ Offers Through TripAdvisor’s Eyes

This empirical data-driven research aims to unveil thought-provoking insights on the U.S. hotel offer across its 50 states. Information of more than 30,000 hotels was collected through web scraping from TripAdvisor. Using such data, 50 support vector machine models were trained to model the TripAdvisor score, one per state, to assess the convergent and divergent factors in customer satisfaction across all the U.S. states. A conceptual model is proposed and validated through the data-driven support vector machine models developed for each state to identify convergent features across the states to explain customer satisfaction (here represented by TripAdvisor score). Hotel size, price, and stars are not moderated by the location, expressed by the corresponding state, although these highly influence satisfaction, whereas both hotel number of published photos and the amenities are affected by the location. Thus, adaptation issues were found regarding amenities and published photos within each state’s offer.

Support vector ensemble for incipient fault diagnosis in nuclear plant components

Abstract The randomness and incipient nature of certain faults in reactor systems warrant a robust and dynamic detection mechanism. Existing models and methods for fault diagnosis using different mathematical/statistical inferences lack incipient and novel faults detection capability. To this end, we propose a fault diagnosis method that utilizes the flexibility of data-driven Support Vector Machine (SVM) for component-level fault diagnosis. The technique integrates separately-built, separately-trained, specialized SVM modules capable of component-level fault diagnosis into a coherent intelligent system, with each SVM module monitoring sub-units of the reactor coolant system. To evaluate the model, marginal faults selected from the failure mode and effect analysis (FMEA) are simulated in the steam generator and pressure boundary of the Chinese CNP300 PWR (Qinshan I NPP) reactor coolant system, using a best-estimate thermal-hydraulic code, RELAP5/SCDAP Mod4.0. Multiclass SVM model is trained with component level parameters that represent the steady state and selected faults in the components. For optimization purposes, we considered and compared the performances of different multiclass models in MATLAB, using different coding matrices, as well as different kernel functions on the representative data derived from the simulation of Qinshan I NPP. An optimum predictive model - the Error Correcting Output Code (ECOC) with TenaryComplete coding matrix - was obtained from experiments, and utilized to diagnose the incipient faults. Some of the important diagnostic results and heuristic model evaluation methods are presented in this paper.

Data-driven support vector machine with optimization techniques for structural health monitoring and damage detection

Rapid detecting damages/defeats in the large-scale civil engineering structures, assessing their conditions and timely decision making are crucial to ensure their health and ultimately enhance the level of public safety. Advanced sensor network techniques recently allow collecting large amounts of data for structural health monitoring and damage detection, while how to effectively interpret these complex sensor data to technical information posts many challenges. This paper presents three optimization-algorithm based support vector machines for damage detection. The optimization algorithms, including grid-search, partial swarm optimization and genetic algorithm, are used to optimize the penalty parameters and Gaussian kernel function parameters. Two types of feature extraction methods in terms of time-series data are selected to capture effective damage characteristics. A benchmark experimental data with the 17 different scenarios in the literature were used for verifying the proposed data-driven methods. Numerical results revealed that all three optimized machine learning methods exhibited significantly improvement in sensitivity, accuracy and effectiveness over conventional methods. The genetic algorithm based SVM had a better prediction than other methods. Two different feature methods used in this study also demonstrated the appropriate features are crucial to improve the sensitivity in detecting damage and assessing structural health conditions. The findings of this study are expected to help engineers to process big data and effectively detect the damage/defects, and thus enable them to make timely decision for supporting civil infrastructure management practices.

Optimization of landslide conditioning factors using very high-resolution airborne laser scanning (LiDAR) data at catchment scale

Landslide susceptibility, hazards, and risks have been extensively explored and analyzed in the past decades. However, choosing relevant conditioning factors in such analyses remains a challenging task. Landslide susceptibility mapping employs topological, environmental, geological, and hydrological parameters. Some researchers assume that as the number of conditioning factors increases, the precision of the generated susceptibility map increases. By contrast, other case studies prove that a small number of conditioning factors are sufficient to produce landslide susceptibility maps with a reasonable quality. This study investigates the effects of conditioning factors on landslide susceptibility mapping. Bukit Antarabangsa, Ulu Klang, Malaysia was selected as the study area, because it is a catchment area with a high potential of landslide occurrence. A spatial database of 31 landslide locations was evaluated to map landslide-susceptible areas. Two datasets of conditioning factors were constructed in GIS environment. The first dataset was derived from high-resolution airborne laser scanning data (LiDAR), which contains eight landslide conditioning factors: altitude, slope, aspect, curvature, stream power index (SPI), topographic wetness index (TWI), topographic roughness index (TRI), and sediment transport index (STI). The second dataset was gathered by using the same conditioning factors of the first dataset, but with the addition of other conditioning factors: geological and environmental factors of soil, geology, land use/cover (LULC), distance from river, and distance from road. Two different datasets were constructed to compare the efficiency of one over the other in landslide susceptibility zonation. Three methods were implemented to recognize the importance of different conditioning factors in landslide susceptibility mapping. Three different types of models such as weights-of-evidence (WoE) (bivariate statistical analysis), logistic regression (LR) (multivariate statistical analysis), and data-driven support vector machine (SVM) were used to determine the optimal landslide conditioning factors. The area under curve (AUC) was used to assess the obtained results. The prediction rates of WoE, LR, and SVM obtained from only the LiDAR-derived conditioning factors were 59%, 86%, and 84%, respectively. The prediction rates of the WoE, LR, and SVM obtained from the second dataset were 65%, 66%, and 69%, respectively. The LiDAR-derived conditioning factors were more sufficient in generating an accurate landslide susceptibility map. Using additional factors, such as geology, LULC, and so on, does not significantly increase the accuracy of the map. The findings of this study can be used as reference for future analysis in selecting data for landslide conditioning factors.

STAR-GALAXY CLASSIFICATION IN MULTI-BAND OPTICAL IMAGING

Ground-based optical surveys such as PanSTARRS, DES, and LSST, will produce large catalogs to limiting magnitudes of r > 24. Star-galaxy separation poses a major challenge to such surveys because galaxies---even very compact galaxies---outnumber halo stars at these depths. We investigate photometric classification techniques on stars and galaxies with intrinsic FWHM < 0.2 arcsec. We consider unsupervised spectral energy distribution template fitting and supervised, data-driven Support Vector Machines (SVM). For template fitting, we use a Maximum Likelihood (ML) method and a new Hierarchical Bayesian (HB) method, which learns the prior distribution of template probabilities from the data. SVM requires training data to classify unknown sources; ML and HB don't. We consider i.) a best-case scenario (SVM_best) where the training data is (unrealistically) a random sampling of the data in both signal-to-noise and demographics, and ii.) a more realistic scenario where training is done on higher signal-to-noise data (SVM_real) at brighter apparent magnitudes. Testing with COSMOS ugriz data we find that HB outperforms ML, delivering ~80% completeness, with purity of ~60-90% for both stars and galaxies, respectively. We find no algorithm delivers perfect performance, and that studies of metal-poor main-sequence turnoff stars may be challenged by poor star-galaxy separation. Using the Receiver Operating Characteristic curve, we find a best-to-worst ranking of SVM_best, HB, ML, and SVM_real. We conclude, therefore, that a well trained SVM will outperform template-fitting methods. However, a normally trained SVM performs worse. Thus, Hierarchical Bayesian template fitting may prove to be the optimal classification method in future surveys.