86,633 research outputs found
Simulating the dynamics of the neutron flux in a nuclear reactor by locally recurrent neural networks
No full textAssessment of the ultimate actual strength of rock-climbing protection devices: Extraction tests in the field and the human capability to predict the ultimate strength
No full textBackground. Rock climbing protection devices are crucial for climbing practice safety and for mountaineering in general. The use of these devices, together with appropriate tech-niques, reduces injuries in the critical event of a climber’s fall. Although European stan-dards and rules support the manufacturer in the design, production and laboratory test-ing, a thorough investigation of their behaviour in a real environment and during an actual placement has not yet been performed. Methods. The aim of this work is to present an insight into the strength of such devices through the application of a monitored, quasi-static, increasing force in a field environ-ment. Results from several types of devices (pitons, nuts and cams) are presented and crit-ically evaluated with respect to the values of the loads acting on the anchors due to the fall of the climber. Results. As far as the piton actual strength is concerned, the present activities show that the characteristics requested by EN specifications and rules are functional for product qualification purposes, but of very little use when defining the load holding capabilities once the devices are in place. However, even if the actual strength does not match the requirement of the standard, the comparison with the actual load applied is fairly encour-aging. With regards to nuts and cams, it is worth underlining the importance of a correct placement: when placed correctly, the actual strength achieved by the device in the field complies and is higher than the classification of the EN standard. Moreover, an investigation of human capability to predict the ultimate strength of rock-climbing protection devices placed in the field has been carried out, with the aim of verifying the reliability of the climber’s judgement, and, possibly, improve the safety of the in-field decision-making process. Conclusions. The lesson learned from the experiments is that modern equipment shows one step better behaviour and, similarly to pitons, the device-rock coupling dictates the pairs actual strength, assuming of course a sound placement. To the author’s best knowl-edge, the present work represents the first attempt to investigate the human capabilities to assess the reliability of a protection placement in-field
Recurrent Neural Networks for Dynamic Reliability Analysis
No full textA dynamic approach to the reliability analysis of realistic systems is likely to increase the computational burden, due to the need of integrating the dynamics with the system stochastic evolution. Hence, fast-running models of process evolution are sought. In this respect, empirical modelling is becoming a popular approach to system dynamics simulation since it allows identifying the underlying dynamic model by fitting system operational data through a procedure often referred to as ‘learning’. In this paper, a Locally Recurrent Neural Network (LRNN) trained according to a Recursive Back-Propagation (RBP) algorithm is investigated as an efficient tool for fast dynamic simulation. An application is performed with respect to the simulation of the non-linear dynamics of a nuclear reactor, as described by a simplified model of literature
Global reliability sensitivity analysis by Sobol-based dynamic adaptive kriging importance sampling
No full textThe stochastic uncertainties affecting the models used to describe the behavior of structural/mechanical systems may give rise to unfavorable scenarios leading to failures. In this framework, the quantification of the failure probability is a recognized fundamental task for structural safety and reliability analyses. Unfortunately, the estimation of the failure probability of structural/mechanical systems is a computationally demanding task, especially when the failure is a rare event and the computer codes used to model the system response require large computational efforts. One major issue further complicates the estimation process, i.e., the parameters of the probability distributions of the random variables used to describe the uncertainties involved can, in turn, be imprecise, since they are typically estimated by means of statistical inference based on observations and engineering judgment. In this context, reliability sensitivity analysis aims at estimating the influence of this additional source of uncertainty on the system failure probability in order to assess the robustness of the system to the modeling of uncertainties. Intuitively, reliability sensitivity analyses may easily become prohibitive by standard sampling-based methods (e.g., Monte Carlo method), since a nested, second level of uncertainties is involved. To overcome this issue, in this work we embed the efficient AK-IS algorithm for estimating small failure probabilities within an original computational framework that allows to perform a Sobol-based, global sensitivity analysis of the failure probability at an affordable number of computer model evaluations. The algorithm is demonstrated with reference to two case studies of literature of structural/mechanical reliability, often used in the literature as benchmark tests
Validation of Infinite Impulse Response Multilayer Perceptron for Modelling Nuclear Dynamics
No full textArtificial neural networks are powerful algorithms for constructing nonlinear empirical models from operational data. Their use is becoming increasingly popular in the complex modeling tasks required by diagnostic, safety, and control applications in complex technologies such as those employed in the nuclear industry. In this paper, the nonlinear modeling capabilities of an infinite impulse response multilayer perceptron (IIR-MLP) for nuclear dynamics are considered in comparison to static modeling by a finite impulse response multilayer perceptron (FIR-MLP) and a conventional static MLP. The comparison is made with respect to the nonlinear dynamics of a nuclear reactor as investigated by IIR-MLP in a previous paper. The superior performance of the locally recurrent scheme is demonstrated
Particle Filters and Auto-Encoders Combination for Damage Diagnosis on Hysteretic Non-Linear Structures Subject to Changing Environmental Conditions
No full textDamages may naturally arise in structures within their life span due to the insurgence of phenomena related to normal operation. Their occurrence might also be favored by external boundary conditions the systems experience during their lifetime, such as time-varying environmental and operating conditions. Standard maintenance activities, such as scheduled non-destructive testing (NDT) and corrective maintenance, are typically carried out to improve the health and longevity of such systems, typically entailing long downtimes with significant economic impacts. In recent decades, condition-based maintenance strategies (CBM) or even predictive ones (PM) have increasingly gained popularity since, in principle, they allow to optimally intervene on the structure only when really required by its current conditions. These maintenance schemes require that a deep knowledge of the system current state of health and, possibly, of the main degradation mechanisms be available, which may rely on advanced structural health monitoring (SHM) systems being installed on the structures for performing real-time diagnosis and prognosis. Many approaches to SHM have been formulated, with several applications to mechanical, aeronautical, space, and civil structures. Particle Filters (PFs) have been proposed as a model-based, time-domain tool for estimating hidden, not observable system states, including those normally affected by damage, in particular, when the structure behavior is non-linear and affected by non-Gaussian disturbances and noises. Yet, in case of varying operating and environmental conditions, the SHM task often still turns out to be quite challenging, since the diagnostic features associated with damage can be significantly distorted. To overcome this issue, auto-encoders have successfully been employed to extract damage-related features in presence of such varying external conditions. Thus, this work aims at combining these two methods for developing an original approach to damage detection and localization in structures, robust with respect to changing environmental and operating conditions, capable of leveraging the specific benefits provided by the two aforementioned methodologies. The proposed algorithm is demonstrated with reference to the problem of damage diagnosis on a vibrating n-degrees of freedom system, featuring a non-linear stiffness component characterized by a Bouc-Wen hysteretic behavior and subject to varying temperature conditions
Convolutional Neural Networks for Ultrasonic Guided Wave-Based Structural Damage Detection and Localisation
No full textAmong the many methods proposed in the literature to perform structural health monitoring (SHM) of thin-walled structures, two of them appear to be particularly promising and complementary. On the one hand, integrating Machine Learning techniques into this field seems a remarkable solution, since these methods have been shown to be effective in recognising usually hard-to-detect recurring patterns in the measured signals related to the presence of damages in structures, thus improving the diagnostic performances of SHM frameworks. In particular, in the past years, Deep Learning algorithms have gained much importance in this field due to their capability of processing high-dimensional inputs (such as images), thus making it possible to automatically identify onsetting structural damages. On the other hand, ultrasonic guided wave-based approaches are commonly adopted to assess the structural integrity of plate-like structures and pipelines. These approaches, coupled with tomographic algorithms, typically allow performing damage detection and localisation with satisfactory results. However, such reconstruction algorithms are significantly sensors layout-dependent and, as such, they come with some still unsolved issues, leading, for example, to artifacts creation and unsatisfactory tomographic damage localisation performances in case of unevenly distributed network of sensors or when few sensors are installed on the structure. In this work, convolutional neural networks (CNNs) and ultrasonic guided waves are combined into a unique framework, which leverages on the advantages of the two methods to perform damage detection and localisation in plate-like structures. Guided waves are excited and sensed by a network of sensors permanently installed on the structure. The information acquired is then converted into grayscale image as is, without performing any prior feature extraction procedure, which is further analysed by a set of CNNs. First, a classifier is employed to perform damage detection. In case damage is identified, the grayscale image is then analysed by two regression CNNs to localise the damage. The framework is tested using experimentally validated numerical simulations of guided waves propagating in a metallic plate available in the literature
Multiple local particle filter for high-dimensional system identification
No full textNonlinearity and high dimensionality emerge as two primary challenges in the realm of system identification within the context of structural health monitoring (SHM) applications. Particle filter (PF) has been demonstrated efficient for nonlinear identification, but it suffers from the curse of dimensionality and behaves poorly in high-dimensional problems. The idea of state and measurement partitioning has been used in many PF algorithms to simplify high-dimensional identification problems into the identification of several lower-dimensional subgroups, but with very few applications to SHM problems. In this context, by combining multiple particle filters (MPF) with the decay of correlations property, this paper develops a novel multiple local particle filter (MLPF) for high-dimensional identification problems. A whole state vector is partitioned into several state subgroups, each consisting of fewer state components and then estimated by one PF through a novel likelihood including the local state and measurement vectors. The feasibility and efficiency of the proposed method are tested through a benchmark toy example, a case study of a twenty-story Bouc-Wen frame structure under ground motion, and an experimental study of fatigue delamination growth in composites
Towards a deep learning-based unified approach for structural damage detection, localisation and quantification
No full textUltrasonic guided waves have been extensively employed for characterising structural damage thanks to their sensitivity to defects. Although they are easy to excite and acquire, heavy processing is often required to extract single-valued indicators of damage presence, or damage indices, from the acquired signals. Traditionally, damage indices have been elaborated through tomographic algorithms to generate damage probability maps, even though limitations affect the performance of such approach. Recently, the potentialities of machine learning have been leveraged to improve the accuracy of frameworks processing guided waves for damage diagnosis. However, most methods still require extracting damage indices from the acquired signals, which may bring to loss of diagnostic information and reduced accuracy. Furthermore, damage position and extent are usually roughly estimated through classification, while regression should be employed instead. In this context, this work aims (i) to test the capabilities of different supervised machine learning algorithms to localise and quantify damage through regression and (ii) to carry out a critical discussion about possible limitations of using damage indices instead of unprocessed signals. Results are compared to identify which algorithm performs better and if machine learning can improve the accuracy of damage diagnosis compared to traditional imaging methods. An experimentally validated numerical case study was used to test the capabilities of the proposed machine learning-based framework and to bring evidence of the accuracy of the algorithms involved to characterise damage with properties not seen during training
- …
