Peter Karsmakers

This paper presents a constraint-guided deep learning (DL) framework to develop physically consistent health indicators (HIs) in bearing prognostics and health management. Conventional data-driven approaches often lack physical plausibility, while physics-based models are limited by incomplete knowledge of complex systems. To address this, we integrate domain knowledge into DL models via constraints, ensuring monotonicity, bounding output ranges between 1 and 0 (representing healthy to failed states, respectively), and maintaining consistency between signal energy trends and HI estimates. Using constraints eliminates the need for complex loss term balancing to incorporate domain knowledge. The constraint-guided gradient descent algorithm (CGGD) is used to train a DL model that satisfies specific constraints. Using time-frequency representations of accelerometer signals from the pronostia and XJTU-SY bearing datasets, the model learned using constraints generates more accurate and reliable representations of bearing health compared to conventional methods. It produces smoother degradation profiles that align with the expected physical behavior. Model performance is assessed using three metrics: trendability, robustness, and consistency. When compared to a conventional baseline model, the model learned using constraints shows a significant improvement in all three metrics. Another baseline incorporated the monotonicity behavior directly into the loss function using a soft-ranking approach. While this approach outperforms the model learned using constraints in trendability, due to its explicit monotonicity enforcement, the model learned with constraints performed better in robustness and consistency, providing stable and interpretable HI estimates over time. The ablation study confirms the importance of each constraint: the monotonicity constraint improves trendability, the boundary constraint ensures consistency, and the energy–HI consistency constraint enhances robustness. These findings demonstrate the effectiveness of CGGD in producing reliable and physically meaningful HIs for bearing prognostics and health management, offering a promising direction for future prognostic applications.

This article presents a contextual outlier detection technique employing Machine Learning (ML) methods to improve the quality of experimentally obtained data from heavy-ion irradiation campaigns. Artifacts arising from the experimental setup often compromise the integrity and representativeness of the measured data. These artifacts, also known as outliers, can be removed to ensure a consistent dataset that accurately represents the Device Under Test (DUT). Due to the dynamic and variable nature of the DUT properties, outliers may manifest in various forms, rendering conventional outlier detection techniques inadequate. To address this challenge, we present a contextual outlier removal technique that incorporates spatial context by segmenting the tested area and using the Isolation Forest (iForest) algorithm for localized outlier detection. As a case study, this technique is applied to Single-event frequency transient (SEFT) data obtained during heavy-ion microbeam irradiation of a planar spiral inductor within an integrated 65 nm CMOS LC-tank digitally controlled oscillator (DCO) circuit. A graphical side-by-side comparison demonstrates significant improvements in data quality, validated by a proposed metric that shows an average enhancement of 4.4x.

Systems consisting of multiple components can experience downtime during their operational lifetime due to issues with one or more of these components. For instance, accurately predicting the remaining time before a battery depletes its total energy is crucial for systems like a Base Transceiver Station. This paper addresses how long the battery will sustain a Base Transceiver Station under varying load conditions with the associated uncertainty when the external power source is interrupted. A customized approach using battery modeling and load forecasting is designed to predict the remaining usable time of the battery. In total, 7 Base Transceiver Stations are considered and clustered into 3 groups based on their installed battery capacity. First, a battery model is developed employing the modified Shepard model using real-life measured battery discharge data. Since the load currents are unknown beforehand, they are forecasted using a time series modeling approach. The resulting uncertainty from both models is then quantified using a prediction interval. Using our approach, the averaged Mean Absolute Error of the remaining usable time estimates is 12.61 minutes for all discharges. The associated uncertainty with these estimation using the 95% prediction interval is 26.43 minutes. The estimation errors are relatively small compared to true discharge times, as some batteries can provide energy close to 15 hours. The overall estimation results closely matched the measured values, providing valuable information for proactive planning and management of potential failures.

Indoor radar-based human activity recognition (HAR) using machine learning has shown promising results. However, deploying an HAR model in unseen environments remains challenging due to a potential mismatch between training and operational conditions. Such mismatch can be reduced by acquiring annotated training data in more diverse situations. However, since this is time intensive, this paper explores the application of data augmentation and unsupervised domain adaptation (UDA) to enhance the robustness of HAR models, even when they are trained using a very limited amount of annotated data. In the initial analysis, a baseline HAR model was evaluated using a validation set (a) from the same environment as the training data and (b) from a different environment. The results showed a 29.6% decrease in the F1-score when tested on data from the different environment. Implementing data augmentation techniques—specifically, time–frequency warping—reduced this performance gap to 17.8%. Further improvements were achieved by applying an unsupervised domain adaptation strategy, which brought the performance gap drop down to 13.2%. Furthermore, an ablation study examining various augmentation methods and synthetic sample quantities demonstrates the superior performance of our proposed augmentation approach. The paper concludes with a discussion on how environmental variations, such as changes in aspect angle, occlusion and layout, can affect the time-Doppler radar representation and, consequently, HAR performance.

The variation in the acoustic condition of a room presents a major hurdle in the performance robustness of sound event classification. Room impulse response characterizes the way in which a sound wave is propagated from source to receiver and the overall perceptual quality and intelligibility of the recorded sound. This study presents the Room Acoustic Adversarial Neural Network (RAANN) method that can make sound event classification more robust to changes in acoustic condition by exploiting knowledge regarding the room acoustics during learning. With RAANN, the weighted F1 score for the classification task improved by 1.54 percentage points, and the standard deviation in performance dropped from 1.74 percentage points to 1.07 percentage points for acoustic conditions that were harder than those seen during the learning phase. The Clarity Index over the first 25 ms emerged as a good metric for the acoustic estimation in the RAANN training.

The main goal of machine condition monitoring is, as the name implies, to monitor the condition of industrial applications. The objective of this monitoring can be mainly split into two problems. A diagnostic problem, where normal data should be distinguished from anomalous data, otherwise called Anomaly Detection (AD), or a prognostic problem, where the aim is to predict the evolution of a Condition Indicator (CI) that reflects the condition of an asset throughout its life time. When considering machine condition monitoring, it is expected that this CI shows a monotonic behavior, as the condition of a machine gradually degrades over time. This work proposes an extension to Constraint Guided AutoEncoders (CGAE), which is a robust AD method, that enables building a single model that can be used for both AD and CI estimation. For the purpose of improved CI estimation the extension incorporates a constraint that enforces the model to have monotonically increasing CI predictions over time. Experimental results indicate that the proposed algorithm performs similar, or slightly better, than CGAE, with regards to AD, while improving the monotonic behavior of the CI.