Advanced Dimension Reduction for Stochastic Dynamic Systems: rLSTM-AE for Efficient Dynamic Reliability Analysis

The field of dynamic reliability analysis is facing some big challenges these days, mainly because of the complexity involved with high-dimensional systems. In civil engineering, aerospace, and mechanical design, engineers deal with systems that have tons of variables to consider. They have to think about things like material properties, environmental conditions and even unpredictable forces like earthquakes. With all these moving parts it’s essential to understand how uncertainties ripple through the system and affect performance over time. But here’s the problem: working with such high-dimensional data is incredibly complicated. It requires a ton of computing power which make it really difficult to keep the analysis both accurate and efficient. Engineers are constantly trying to find ways to tackle these tough problems but the sheer amount of data they have to process makes it a major challenge. Traditionally, Monte Carlo Simulation (MCS) have been used for reliability analysis where thousands or even sometimes millions of system simulations are performed to evaluate the failure probability and such method is considered accurate, however, it is computationally expensive and impractical for large-scale, high-dimensional systems. On the other hand, surrogate models (also called metamodels) have recently emerged as a potential solution to reduce computational demand and techniques such as Polynomial Chaos Expansion (PCE) and Gaussian Process Regression (GPR) are frequently used to replace expensive simulations with more manageable approximations. Unfortunately, these surrogate models also suffer from the so-called “curse of dimensionality.” As the number of input variables increases, the accuracy of these models tends to degrade, which necessitates more training data to maintain performance that result in worse computational burden. This challenge has stimulated interest in dimension-reduction techniques which aim to reduce the number of variables while preserving the essential characteristics of the system. Commonly used approaches like Principal Component Analysis (PCA) or its variants, while effective in some cases, are often inadequate for capturing the nonlinear dynamics and time-dependent behavior inherent in complex stochastic systems. As a result, there’s a critical need for more advanced methods that can handle both the time-invariant structural parameters and the time-variant external forces in a system, such as seismic excitations.

To this account, recent paper published in Mechanical Systems and Signal Processing Journal and conducted by Yu Zhang, Associate Professor You Dong from the Department of Civil and Environmental Engineering at The Hong Kong Polytechnic University in collaboration with Professor Michael Beer from the Institute for Risk and Reliability at the Leibniz University Hannover in Germany, the researchers developed a more effective solution for dimension reduction and reliability analysis in high-dimensional stochastic dynamic systems. Their goal was to develop a method that could simultaneously account for time-dependent variables and reduce the dimensionality of the input space without sacrificing accuracy. They introduced the rLSTM-AE (Recurrent Long Short-Term Memory with Autoencoder), a novel approach designed to address the limitations of existing methods by combining deep learning techniques with active learning strategies. This study is particularly important because it provides a framework for efficiently analyzing high-dimensional systems, where traditional methods fail due to excessive computational demands or inadequate accuracy.

In their study, the researchers carried out two major experiments to test how well their new rLSTM-AE method could handle the complexities of high-dimensional systems. For the first experiment, they looked at a single-degree-of-freedom (SDOF) model, which is commonly used to test how reliable systems are under dynamic conditions. They applied random seismic forces to the SDOF system and included key factors like mass, stiffness, and damping as random variables. To really push the limits, they also added over 1,000 random variables to represent different aspects of the earthquake’s ground motions. This made the system both time-varying and time-invariant, which created a realistic, but challenging, scenario for the rLSTM-AE method. The goal was to see how well this method could predict extreme responses which is an essential aspect when studying system reliability. What they found was pretty remarkable. The rLSTM-AE method was able to take this overwhelming number of 1,004 variables and narrow them down to just 11 important ones, all while maintaining a high degree of accuracy. Even with this massive simplification, the method stayed accurate in predicting extreme responses, with results closely matching those of traditional MCS. Not only did rLSTM-AE match the accuracy of MCS, but it also did so much more efficiently. By adding active learning-based GPR, the method further improved, delivering accurate failure probability estimates without eating up nearly as much computational power. The second experiment raised the bar even higher. This time, they worked with a 3D reinforced concrete frame, a much more complex and realistic structure, which was subjected to completely random seismic forces. Here, they had seven random factors—like the strength of the concrete and the steel’s yield strength—and, once again, over a thousand variables total. When they applied the rLSTM-AE method, it proved capable of reducing the problem to a smaller set of essential variables while still keeping accuracy high. They compared these results with the gold-standard Monte Carlo simulations. Despite the 3D system’s complexity, rLSTM-AE successfully captured the most important extreme responses and provided failure probabilities that were impressively close to the Monte Carlo benchmarks. It performed particularly well when predicting small failure probabilities, achieving a relative error of under 5% with the help of active learning-based GPR. This demonstrated that rLSTM-AE could be trusted even for low-probability events, which are notoriously tricky to estimate accurately. Overall, these two experiments highlighted the rLSTM-AE method’s potential in managing high-dimensional systems. By zeroing in on only the most critical variables, the method made it possible to conduct reliable, efficient analysis without needing the hefty computational power typically required. And with the addition of active learning, the model could adaptively improve its predictions, making it a valuable tool for engineers dealing with complex and uncertain systems.

The new study of Zhang, Dong and Beer has the potential to really shake things up in engineering by offering a new way to analyze complex systems with tons of variables. The rLSTM-AE method that the researchers developed addresses a major challenge: dealing with systems that have a huge number of unpredictable factors. Typically, traditional methods can’t handle these well without using up a lot of time and computing power. But this new approach changes that by cutting down the number of variables without losing accuracy. It can still accurately predict extreme events and estimate failure risks. This breakthrough is especially relevant in fields like civil, mechanical, and aerospace engineering, where systems face uncertain conditions and forces all the time. The rLSTM-AE method is different from other approaches because it can work with both stable and changing variables at once, capturing complex system behavior that others miss. This allows engineers to tackle systems that were once considered too complex or costly to analyze properly. Plus, with the active learning component, the method can focus only on the most important parts of the system, which means fewer simulations are needed and the process becomes a lot more efficient. This isn’t just about saving on computing resources but also has real-world implications. We can think for instance in areas like earthquake engineering or monitoring the health of structures, accurately predicting potential failures is essential for safety and making well-informed decisions. Additionally, the study also opens the door to new research possibilities and a lot of potential for future work, like combining this method with other advanced techniques or looking into multi-objective optimization. Doing so could make these tools even more flexible and powerful for engineers. Overall, the rLSTM-AE method doesn’t just simplify the analysis of complex systems; it could also lead to better risk assessments, stronger infrastructure, and more reliable engineering solutions across the board.