Abstract: Remaining useful life predictions depend on the quality of health indicators (HIs) generated from condition monitoring sensors, evaluated by predefined prognostic metrics such as monotonicity, prognosability, and trendability. Constructing these HIs requires effective models capable of automatically selecting and fusing features from pertinent measurements, given the inherent noise in sensory data. While deep learning approaches have the potential to automatically extract features without the need for significant specialist knowledge, these features lack a clear (physical) interpretation. Furthermore, the evaluation metrics for HIs are nondifferentiable, limiting the application of supervised networks. This research aims to develop an intrinsically interpretable ANN, targeting qualified HIs with significantly lower complexity. A semi-supervised paradigm is employed, simulating labels inspired by the physics of progressive damage. This approach implicitly incorporates nondifferentiable criteria into the learning process. The architecture comprises additive and newly modified multiplicative layers that combine features to better represent the system’s characteristics. The developed multiplicative neurons are not restricted to pairwise actions, and they can also handle both division and multiplication. To extract a compact HI equation, making the model mathematically interpretable, the number of parameters is further reduced by discretizing the weights via a ternary set. This weight discretization simplifies the extracted equation while gently controlling the number of weights that should be overlooked. The developed methodology is specifically tailored to construct interpretable HIs for commercial turbofan engines, showcasing that the generated HIs are of high quality and interpretable.

@en

In this research, a generalized machine learning (ML) framework is proposed to estimate the fatigue life of epoxy polymers and additively manufactured AlSi10Mg alloy materials, leveraging their failure surface void characteristics. An extreme gradient boosting algorithm-based ML framework encompassing Synthetic Minority Over-sampling TEchnique (SMOTE), categorical data encoding, and external loop cross-validation is developed to evaluate the fatigue life across materials. The influence of different training strategies based on materials, input features, encoding method, and data standardization on the model performance is explored. Additionally, the importance of anti-data-leakage and anti-overfitting measures over the ML model performance is addressed. The result shows that the data-leakage-free, external loop cross-validated model can estimate the fatigue life of selective epoxy polymers and metal alloys with an average R² of 0.71 ± 0.06 using a mere 12 to 27 experimental data points per material category. Whereas the model trained with data-leakage and overfitting results in high R² of 0.9.

@en

In recent years, prognostics gained attention in various industries by optimizing maintenance, boosting operational efficiency, and preventing costly downtime. Central to prognostics is the Remaining Useful Life (RUL), representing the critical time before system failure. Deep learning advancements facilitate RUL forecasting by extracting features from diverse data formats such as time series, images, or sequences thereof, in one, two, or three dimensions, respectively. Yet, predicting RUL from image sequences often relies heavily on resource-intensive techniques like digital image correlation, complicating data acquisition. To address challenges with high-dimensional data and unreliable models, this study introduces ISTRUST, an innovative Transformer-based architecture. ISTRUST (Interpretable Spatiotemporal TRansformer for Understanding STructures) tackles the dual challenges posed by high-dimensional data and the black-box nature of existing models. Leveraging Transformers’ attention mechanism, ISTRUST breaks down the spatiotemporal domain, effectively realizing interpretable RUL predictions under uncertainty using only sparse raw image sequences as input. Evaluated on fatigue-loaded composite samples showcasing crack propagation, ISTRUST interprets the relation between cracks and RUL via the attention mechanism. The results substantiate its capacity to interpret and clarify instances in which predictions may exhibit variability in accuracy. Through the attention mechanism, a strong correlation between the model’s spatiotemporal focus and the RUL predictions is established, making it, to the best of our knowledge, the first model to provide interpretable stochastic RUL predictions directly from sequential images of this nature.@en

A hybrid methodology based on numerical and non-destructive experimental schemes, which is able to predict the structural level strength of composite laminates is proposed on the current work. The main objective is to predict the strength by substituting the up to failure experiments with non-destructive experiments where the investigated specimen is loaded up to 20% of its maximum load. A significant gap exists between the 20% and the 100% load which is proposed to be treated by high fidelity physics-based numerical models, deep learning techniques, and non-catastrophic experiments. Thus, a deep learning algorithm is developed, based on the convolutional neural networks and trained by probabilistic failure analysis datasets which result from the utilization of the stochastic finite element method. Also, the Monte Carlo dropout technique is embedded into the developed convolutional neural network to estimate the uncertainty induced by the investigated variations between the simulated and experimental data. The current paper provides a thorough description of the proposed methodology and a practical example which demonstrates the validity of the method.

@en

Recently, companies all over the world have been focusing on the improvement of autonomous health management systems in order to enhance performance and reduce downtime costs. To achieve this, the remaining useful life predictions have been given remarkable attention. These predictions depend on the proper designing process and the quality of health indicators (HI) generated from structural health monitoring sensors based on prior established multiple prognostic evaluation criteria. Constructing such HIs from noisy sensory data demands powerful models that enable the automatic selection and fusion of features taken from those relevant measurements. Deep learning models are promising to autonomously extract features in scenarios with a huge volume of data without requiring considerable domain expertise. Nonetheless, the features established by artificial neural networks are complicated to comprehend and cannot be regarded as physical system characteristics. In this regard, the goal of this paper is to extend a new model; an interpretable artificial neural network that enables the automatic selection and fusion of features to construct the most appropriate HIs with remarkably fewer parameters. This model consists of additive and multiplicative layers that provide a feature fusion that better reflects the system’s physical properties. Additionally, the weights are discretized in two ways: a) using a ternary form with values {-1, 0, 1}, and b) relaxing the aforementioned ternary form by rounding the weights at the first decimal point in the range of [-1, 1]. Both discretization techniques have the ability to softly control the number of parameters that should be ignored. This trick guarantees interpretability for the neural network by extracting simple yet powerful equations representing the constructed HIs. Finally, the model’s performance is evaluated and compared with other approaches using a practical case study. The results show that the proposed approach's designed HIs are both interpretable and of high quality according to the criteria of the HI's evaluation.@en