Visakan Kadirkamanathan

Multidisciplinary design optimisation (MDO) is a field of optimisation where problems are partitioned into a set of subproblems, or disciplines, with interactions between them. Multi-objective (MO)-MDO considers cases where multiple objectives exist at the problem or subproblem level. Of particular interest are MO-MDO architectures and methods for a posteriori decision making—in which the aim of the optimiser is to produce an approximation of the efficient set of Pareto optimal solutions. A variety of a posteriori MO-MDO methods have been proposed, many of which draw on concepts and tools from evolutionary computation and machine learning. However, these approaches have arisen across a fragmented set of literatures, and there is no unified guide available to support practitioners or steer the progressive development of new methods. This survey aims to provide such a unified perspective. The issue of characterising multi-objectivity in MO-MDO is discussed and a typology is introduced to identify the different methods in an accessible and straightforward way. The available benchmark problems for MO-MDO are also surveyed. Key issues and potential paths for future research in MO-MDO are identified and discussed.

Modern aircraft assembly involves drilling holes in multi-material stacks, followed by fastening to create a structural joint. Stringent quality requirements to ensure structural integrity necessitate accurate assessment of the hole quality. Carbon fibre reinforced polymer (CFRP) is widely used in aircraft structures, and delamination and uncut fibres are prominent defects associated with CFRP drilling. Furthermore, metal-CFRP composite structures, owing to their higher cost and hence the need reduced waste, present even more stringent requirements on hole quality. Defects such as delamination and uncut fibres in this composite also differ in how they manifest. An automated evaluation framework and robust hole quality metrics are therefore required to address this challenge. In this paper, an automated image processing based evaluation framework for rapid and objective assessment of delamination and uncut fibre defects is developed. This robust framework introduces combined metrics, leveraging a weighted p-norm approach to aggregate five factors for delamination and five for uncut fibres with p = 1, 2 and ∞. A method to choose weights for the different p-norms is also identified to create robust measures. Grayscale images captured via optical microscopy serve as inputs, and the framework outputs both the combined metrics and ten individual metrics for detailed analysis. The multi-objective space with the associated Pareto front and population based ranking derived from the weighted p-norm based metrics are introduced as a means to identify relative hole quality and to associate these with the machining parameters. A dataset comprising 24 holes from two metal-CFRP composite workpieces produced in a robotic drilling operation is used for demonstration of the framework. The results demonstrate the effectiveness of the proposed framework and its utility for real-world applications.

Interpretability and explainability are at the core of applications developed for control of safety-critical systems, requiring low-complexity models, with physically meaningful insights, and maximum prediction accuracy. This can lead to two very distinct representations of non-linear systems: models purely based on first-principles, highly explainable but extremely difficult to use in practice, or data-intensive, with almost no interpretability but tailored to each specific application. To harness the advantages of both approaches, this paper introduces a novel polynomial linear-parameter-varying framework with stability guarantees to model gas turbine engines, with interpretable dynamical states. The identification problem is split into three stages: (i) identification of the scheduling variable mapping via least squares; (ii) identification of the state dynamics via constrained least squares optimisation involving linear matrix inequalities; (iii) identification of the output equation via least squares. The modelling framework inherits interpretability through the selection of physical variables as dynamical states, while model smoothness is enforced by the use of polynomial functions, which are amenable for control design and optimisation. A unique model for the gas turbine engine is obtained at sea level static, and then extended to wider operating conditions through transformations to referred variables. The effectiveness of the modelling framework is demonstrated on two scenarios, using an engine from the literature, in which low prediction errors were observed, including avoidance of instabilities. Potential applications range from digital-twins and Monte-Carlo simulations, to gain-scheduled and model predictive control, or even economic optimisation, among others.

Understanding how bone adapts to external forces is fundamental for exploring potential biomechanical interventions against skeletal diseases. This can be studied preclinically, combining in vivo experiments in rodents and in silico mechanoregulation models. While the in vivo tibial loading model is widely used to study bone adaptation, the common assumption of purely axial loading may be a simplification. This study quantifies the effect of the loading direction on the strain energy density (SED) distribution in the mouse tibia, a commonly used input for mechanoregulated bone remodelling models. To achieve this, validated micro-finite element (micro-FE) models were used to test the differences in local SED when the bone was loaded along different loading directions. In vivo micro-computed tomography (micro-CT) images were acquired from the tibiae of eleven ovariectomised mice at 18 weeks old before intervention and at 20 weeks old, after six mice underwent external mechanical loading. Micro-CT-based micro-FE models were generated for each tibia at both time points and loaded with a unit load in each Cartesian direction independently. The results from these unit load models were linearly combined to simulate various loading directions, defined by angles θ (inferior-superior) and ϕ (anterior–posterior). The results revealed a high sensitivity of the mouse tibia to the loading direction across both groups and time points. Several loading directions (e.g., θ = 10 ° , ϕ = 205–210 ° ) resulted in lower medians of the top 5% SED values compared to those obtained for the nominal axial case ( θ = 0 ° , ϕ = 0 ° ). Conversely, higher values were observed for other directions (e.g., θ = 30 ° , ϕ = 35–50 ° ). These findings emphasise the importance of considering the loading direction in experimental and computational bone adaptation studies.

Sensed data from high-value engineering systems is being increasingly exploited to optimise their operation and maintenance. In aerospace, returning all measured data to a central repository is prohibitively expensive, often causing useful, high-value data to be discarded. The ability to detect, prioritise and return useful data on asset and in real-time is vital to move toward more sustainable maintenance activities.We present a data-driven solution for on-line detection and prioritisation of anomalous data that is centrally processed and used to update individualised digital twins (DT) distributed onto remote machines. The DT is embodied as a convolutional neural network (CNN) optimised for real-time execution on a resource constrained gas turbine monitoring computer. The CNN generates a state prediction with uncertainty, which is used as a metric to select informative data for transfer to a remote fleet monitoring system. The received data is screened for faults before updating the weights on the CNN, which are synchronised between real and virtual asset.Results show the successful detection of a known in-flight engine fault and the collection of data related to high novelty pre-cursor events that were previously unrecognised. We demonstrate that data related to novel operation are also identified for transfer to the fleet monitoring system, allowing model improvement by retraining. In addition to these industrial dataset results, reproducible examples are provided for a public domain NASA dataset.The data prioritisation solution is capable of running in real-time on production-standard low-power embedded hardware and is deployed on the Rolls-Royce Pearl 15 engines.

Deep learning methods contain powerful tools for modelling nonlinear dynamic systems. However, whilst these models are useful for predicting outputs, they tend to be described by complicated black box equations that lack interpretability. They are therefore not so useful for giving insight into system dynamics, and importantly, insight into why a system produces a certain output in response to a given input. This paper presents a novel method for interpreting and comparing deep learning models for nonlinear system identification, using nonlinear output frequency response functions (NOFRFs). NOFRFs describe nonlinear dynamic system behaviour in the frequency-domain using one-dimensional functions, in a manner similar to how Bode plots are used for analysing linear dynamic systems. This is a classical way of interpreting and understanding system behaviour, e.g. via resonances, and in the case of nonlinear systems, super and sub-harmonics, and energy transfer between frequencies. We also use uncertainty quantification via an ensemble bootstrap method to enhance the model interpretation, by propagating the model uncertainty estimates into the frequency-domain. The approach is demonstrated with gated recurrent unit (GRU) and long short term memory (LSTM) models - both are types of recurrent network used in deep learning that are analogous to nonlinear state space models. The results obtained from both a numerical example (a nonlinear mass spring damper system that exhibits energy transfer between frequencies) and a real-world nonlinear system (a magneto-rheological damper) show that it is possible to gain valuable insight and interpretation of the system dynamics from the NOFRFs in a way that is not possible from analysing the time-domain model equations alone.

Raman spectroscopy shows promise as a biomarker for complex nerve and muscle (neuromuscular) diseases. To maximise its potential, several challenges remain. These include the sensitivity to different instrument configurations, translation across preclinical/human tissues and the development of multivariate analytics that can derive interpretable spectral outputs for disease identification. Nonnegative matrix factorisation (NMF) can extract features from high-dimensional data sets and the nonnegative constraint results in physically realistic outputs. In this study, we have undertaken NMF on Raman spectra of muscle obtained from different clinical and preclinical settings. First, we obtained and combined Raman spectra from human patients with mitochondrial disease and healthy volunteers, using both a commercial microscope and in-house fibre optic probe. NMF was applied across all data, and spectral patterns common to both equipment configurations were identified. Linear discriminant models utilising these patterns were able to accurately classify disease states (accuracy 70.2-84.5%). Next, we applied NMF to spectra obtained from the mdx mouse model of a Duchenne muscular dystrophy and patients with dystrophic muscle conditions. Spectral fingerprints common to mouse/human were obtained and able to accurately identify disease (accuracy 79.5-98.8%). We conclude that NMF can be used to analyse Raman data across different equipment configurations and the preclinical/clinical divide. Thus, the application of NMF decomposition methods could enhance the potential of Raman spectroscopy for the study of fatal neuromuscular diseases.