Structural Safety最新文献

The design live loads are determined by the probability distribution of the maximum live load, which is influenced by the amplitudes and time intervals of various sustained and extraordinary loads. If the relative impact of different input variables on the maximum can be clarified, more targeted load surveys and modeling can be achieved. However, there is currently no global sensitivity analysis that simultaneously considers all input variables. In this study, the probability density function of the maximum live load is determined using the load coincidence principle and probability density evolution method. The relative entropy is employed as a measure for conducting a global sensitivity analysis across five common building occupancy types. The results indicate a significant imbalance in the impact of different input variables. The load amplitudes have a much greater effect than the time intervals. Among various load amplitudes, those related to the extraordinary loads often have the most significant impact. Regarding the time intervals, the occurrence intervals corresponding to the extraordinary loads caused by furniture stacking and normal crowding consistently have the least influence. For the time intervals with minimal impact, it is suggested to treat them as deterministic values in the live load modeling. This treatment has a negligible impact (not exceeding 10%) on the mean and upper fractile of the maximum, which are generally used for design in load codes.

We present a new Subset Simulation approach using Hamiltonian neural network-based Monte Carlo sampling for reliability analysis. The proposed strategy combines the superior sampling of the Hamiltonian Monte Carlo method with computationally efficient gradient evaluations using Hamiltonian neural networks. This combination is especially advantageous because the neural network architecture conserves the Hamiltonian, which defines the acceptance criteria of the Hamiltonian Monte Carlo sampler. Hence, this strategy achieves high acceptance rates at low computational cost. Our approach estimates small failure probabilities using Subset Simulations. However, in low-probability sample regions, the gradient evaluation is particularly challenging. The remarkable accuracy of the proposed strategy is demonstrated on different reliability problems, and its efficiency is compared to the traditional Hamiltonian Monte Carlo method. We note that this approach can reach its limitations for gradient estimations in low-probability regions of complex and high-dimensional distributions. Thus, we propose techniques to improve gradient prediction in these particular situations and enable accurate estimations of the probability of failure. The highlight of this study is the reliability analysis of a system whose parameter distributions must be inferred with Bayesian inference problems. In such a case, the Hamiltonian Monte Carlo method requires a full model evaluation for each gradient evaluation and, therefore, comes at a very high cost. However, using Hamiltonian neural networks in this framework replaces the expensive model evaluation, resulting in tremendous improvements in computational efficiency.

The state of the practice in blast-resistant applications against explosions is to design the structural components for a prescribed combination of explosive mass and location, namely the design basis threat. In this context, the blast source is represented by specific scenarios, mostly associated with expert judgement, rating systems or code provisions. While offering a useful basis for practical applications, the level of detail can be significantly enhanced within a probabilistic framework for risk assessment. In research practice, plenty simplified probabilistic approaches have been proposed on external explosions in order to perform risk assessment. A rigorous methodology for such an assessment is presented herein, using tools and techniques derived from seismic risk-assessment applications. Specifically, the mean annual frequency of different explosive mass hazards is represented via a recurrence law, while the potential locations are modelled as a 2D spatial distribution, accounting for the various layers of defense that exist around the structure, i.e., perimeter protection, landscape, public or private spaces, etc. The methodology is finally substantiated with the case study of a typical building subjected to potential blast loadings from external aggressors.

Structural and multidisciplinary design under uncertainty for high reliability or equivalently small probability of failure is a challenging task owing to the high computational cost associated with generating the samples at the extreme (tail) of the underlying distribution. Among other approaches, statistics of extremes based techniques are usually suitable for small probability estimation. However, typically only 10% of the samples generated that correspond to the tail of the distribution are used for probability estimation. If apriori information about regions in the design space that corresponds to the tail is available, additional samples in the identified region permit better tail fit and hence better probability estimation. In the current work, we propose iSOM (interpretable Self-Organising Map) to identify region/s in the design space, that corresponds to the extremes. An initial sample is used to map (visualize) the limit state function and random/design variables using iSOM which permits the designer to identify the region(s) that corresponds to the tail of the response. Adaptive sampling is performed in the identified region of interest to obtain additional samples. Next, the cumulative distribution function (CDF) of the response using initial as well as adaptive samples is evaluated for probability estimation. The effectiveness of the proposed approach is evident from its successful implementation on benchmark examples, real-world engineering examples, and a multi-objective reliability-based design optimization (MORBDO) case. The proposed method showcases the capability of iSOM to perform adaptive sampling for limit-state functions characterized by non-linearity and multiple modes. iSOM-enabled sampling in conjunction with log-TPNT provides better estimates of small failure probabilities than log-TPNT alone. The results from the proposed approach is compared with results from state-of-the-art (SOTA) sampling and surrogate-based techniques. For a given number of limit state evaluations, the proposed approach estimates probabilities of the order 1e−4, with lesser variance, compared to other SOTA approaches. Hence, the proposed approach is likely to encourage further research into employing iSOM-assisted sampling for other reliability estimation methods as well.

Sparse polynomial chaos expansion (PCE) combined with the bootstrap resampling method is a viable alternative to obtain an active learning algorithm for reliability analysis. The existing learning functions in PCE-based active learning algorithms do not consider the joint probability density function (PDF) information. The present study explores a sparse PCE-based active learning algorithm based on a newly proposed learning function that maintains a balance between the misclassification probability and the joint PDF information of sample points. In doing so, the coefficients of the sparse PCE are estimated using a Bayesian compressive sensing regressor, as it is noted to be one of the best-performing regression solvers for PCE, irrespective of sampling schemes. The proposed learning function considers the weight of the joint PDF with the local accuracy measure of bootstrap PCE (bPCE) to add new samples iteratively in the existing training set. The convergence is achieved when the ten consecutive failure estimates are within a negligible discrepancy and also checks the confidence bounds of the bPCE estimates. The effectiveness of the proposed approach is demonstrated using two structural engineering examples and one well-known analytical test function and is found to be quite efficient and accurate in estimating reliability.