Interpretability is essential for user trust in real-world anomaly detection applications. However, deep learning models, despite their strong performance, often lack transparency. In this work, we study the interpretability of autoencoder-based models for audio anomaly detection, by comparing a standard autoencoder (AE) with a mask autoencoder (MAE) in terms of detection performance and interpretability. We applied several attribution methods, including error maps, saliency maps, SmoothGrad, Integrated Gradients, GradSHAP, and Grad-CAM. Although MAE shows a slightly lower detection, it consistently provides more faithful and temporally precise explanations, suggesting a better alignment with true anomalies. To assess the relevance of the regions highlighted by the explanation method, we propose a perturbation-based faithfulness metric that replaces them with their reconstructions to simulate normal input. Our findings, based on experiments in a real industrial scenario, highlight the importance of incorporating interpretability into anomaly detection pipelines and show that masked training improves explanation quality without compromising performance.

Type-1 Diabetes (T1D) is a chronic autoimmune disease affecting millions of patients worldwide. The pancreas of T1D patients no longer produces the insulin responsible for regulating blood glucose levels. Therefore, T1D patients must inject insulin themselves either through multiple daily injections or continuous subcutaneous insulin infusion (insulin pumps). They must daily manage their diabetes to reach the advised blood glucose targets and avoid short-term and long-term complications. With the technological advances achieved in the last decades in diabetes- and non-diabetes-oriented technologies, a large amount of data has become available. This has encouraged researchers to apply artificial intelligence techniques to improve diabetes management by predicting blood glucose levels or recommending actions for better control of T1D. Our paper reviews recent contributions in this area focusing on data-driven approaches for blood glucose prediction. It highlights future research avenues to better exploit the new advances in machine learning and deep learning models.

In recent years, the wood product industry has been facing a skilled labor shortage. The result is more frequent sudden failures, resulting in additional costs for these companies already operating in a very competitive market. Moreover, sawmills are challenging environments for machinery and sensors. Given that experienced machine operators may be able to diagnose defects or malfunctions, one possible way of assisting novice operators is through acoustic monitoring. As a step towards the automation of wood-processing equipment and decision support systems for machine operators, in this paper, we explore using a deep convolutional autoencoder for acoustic anomaly detection of wood planers on a new real-life dataset. Specifically, our convolutional autoencoder with skip connections (Skip-CAE) and our Skip-CAE transformer outperform the DCASE autoencoder baseline, one-class SVM, isolation forest and a published convolutional autoencoder architecture, respectively obtaining an area under the ROC curve of 0.846 and 0.875 on a dataset of real-factory planer sounds. Moreover, we show that adding skip connections and attention mechanism under the form of a transformer encoder-decoder helps to further improve the anomaly detection capabilities.

In search and rescue operations, optimization-based decision support systems can assist search mission coordinators in planning searches with higher probabilities of success, potentially resulting in more lives saved. However, traditional model-and-solve techniques, such as integer programming, are not easily applicable in maritime searches where there is a need to conduct simulations to compute the value of the objective function. In this paper, we show how we can still use mathematical programming to propose maritime search and rescue plans, even when simulations are used. In addition, we take into account operational constraints such as airspace deconfliction for security reasons. Our model, implemented for a suitable solver using a surrogate to estimate search simulation results on scenarios with two helicopters, proved flexible and fast---we implemented operational constraints in a problem-specific model, solved by a general solver, which provided quality solutions in a short time frame.

OBJECTIVE: To compare the effectiveness of a structured high-intensity program (Rouge & Or [R&O]), which focuses on athletics abilities to facilitate return to sport, with usual care during the later phase of anterior cruciate ligament (ACL) reconstruction rehabilitation, in terms of symptoms and disability, and return to sport among recreational athletes. METHODS: Sixty-six individuals, 3 months post-ACL reconstruction, were randomly assigned to the R&O program or usual care. The primary outcome was symptoms and disability, assessed using the Knee Outcome Survey–Activities of Daily Living Scale. The secondary outcomes were adverse events, pain, perceived level of change, psychological readiness to return to play, lower-limb strength, lower-limb functional performance, and rate of return to sport at preinjury level. Follow-ups were conducted at baseline and at 6, 9, and 12 months. A linear mixed model was used to compare the groups. RESULTS: While both groups showed progress on all outcomes ( P<.01), there were no significant between-group differences ( P>.05). There were no adverse events related to the completion of the 2 programs. CONCLUSION: The structured high-intensity program did not provide additional benefits compared to usual care. However, the R&O program was a standardized and safe intervention to guide recreational athletes following ACL reconstruction. JOSPT Open 2025;3(4):483-491. Epub 18 June 2025. doi:10.2519/josptopen.2025.0091

We present a novel Lie algebra based Iterative Reweighted Least Squares (IRLS) algorithm for robust 3D point cloud alignment. We reformulate the optimal update computation to a compact form which requires only one pass through the data. Although this reformulation does not alter the asymptotic computational complexity, it is well suited for contemporary hardware architectures, yielding significant practical speedups. In extensive experiments on challenging benchmark datasets with added correspondence corruption, the method is consistently at least four times faster than previous literature whilst being mathematically equivalent, demonstrating it is well suited for time-critical applications.

We present a novel artificial intelligence approach that encompasses both predictive and prescriptive aspects for the challenging task of model-based control of industrial wood planers. These sophisticated lumber finishing machines are known for the complexity of their operation, and the available data pertaining to the planing process exhibits complex, non-linear patterns. First, we leverage an ensemble of Gaussian Processes with a specialized weighting scheme named Automatic State Matching, achieving a 39% reduction in prediction error for the thickness of the outgoing board compared to conventional industry methods, as corroborated by real-world data. Subsequently, the predictive strategy is utilized in a novel robust control strategy which exploits the properties of Gaussian Processes to prescribe settings for wood planers. An empirical evaluation on simulated data demonstrated the viability of our prescriptive method, resulting in an 83% reduction in deviation from a predetermined target dimension.

In today's dynamic markets, decision-making relies heavily on simulation models to evaluate different production control methods. Although price-driven production control methods have proven their effectiveness in exploiting price volatility, certain industries are still reluctant to adopt these methods in their operational decision-making. This research demonstrates the relevance of price-driven methods for the wood products industry. A sawmill simulator is used to illustrate this. Since the simulation of the sawmill production process is time-consuming, we propose a probabilistic sampling-based method to rationalize the dataset size. A comparative study shows that exploiting historical and recent price data increases sawmill revenues.

Continuous high-frequency wood drying, when integrated with a traditional wood finishing line, allows correcting moisture content one piece of lumber at a time in order to improve its value. However, the integration of this precision drying process complicates sawmills logistics. The high stochasticity of lumber properties and less than ideal lumber routing decisions may cause bottlenecks and reduces productivity. To counteract this problem and fully exploit the technology, we propose to use reinforcement learning (RL) for learning continuous drying operation policies. An RL agent interacts with a simulated model of the finishing line to optimize its policies. Our results, based on multiple simulations, show that the learned policies outperform the heuristic currently used in industry and are robust to sudden disturbances which frequently occur in real contexts.

Robotized welding processes in the manufacturing industry play a crucial role in enhancing competitiveness through automation, adaptability, and increased productivity. To optimize welding parameters, modeling approaches have gained significance, enabling users to simulate welding experiments and determine appropriate settings. With the growing need for reduced development phases and costs while maintaining quality standards and production volumes, flexible and robust manufacturing technologies are essential. In this paper, we present a literature review highlighting best practices for welding processes. We address five research questions related to welding techniques, planning models, factors affecting welding processes, and performance indicators. Our findings reveal various models and techniques for planning robot operations, focusing on welding robots. Bythis review, we contribute to the development of effective strategies for optimizing robotized welding processes, leading to improved efficiency in manufacturing systems.

A sawmilling process scans a wood log and must establish a series of cutting and rotating operations to perform in order to obtain the set of lumbers having the most value. The search space can be expressed as an and/or tree. Providing an optimal solution, however, may take too much time. The complete search for all possibilities can take several minutes per log and there is no guarantee that a high-value cut for a log will be encountered early in the process. Furthermore, sawmills usually have several hundred logs to process and the available computing time is limited. We propose to learn the best branching decisions from previous wood logs and define a metric to compare two wood logs in order to branch first on the options that worked well for similar logs. This approach (Learn, Compare, Search, or LCS) can be injected into the search process, whether we use a basic Depth-First Search (DFS) or the state-of-the-art Monte Carlo Tree Search (MCTS). Experiments were carried on by modifying an industrial wood cutting simulator. When computation time is limited to five seconds, LCS reduced the lost value by 47.42% when using DFS and by 17.86% when using MCTS.

Predicting the lumber products that can be obtained from a log allows for better allocation of resources and improves operations planning. Although sawing simulators make it possible to anticipate the production associated with a log, they do not allow processing many logs quickly. It was shown that machine learning can be used in place of a simulator. However, prediction quality is still lacking and information rich log representations are seldomly used in the literature for machine learning purposes We compare several log representations that can be used (industry know-how-based features, 2D projections, and 3D point clouds) and several neural network architectures able to process these log representations (multilayer perceptron, residual network and PointNet). We also propose a new way to implement a loss function that improves prediction of sparse object count in regression. This new approach achieves a 15% improvement of F1 score compared to previous approaches.

Sawmills are key elements of the forest product industry supply chain, and they play important economic, social, and environmental roles. Sawmill production planning and control are, however, challenging owing to several factors, including, but not limited to, the heterogeneity of the raw material. The emerging concept of digital twins introduced in the context of Industry 4.0 has generated high interest and has been studied in a variety of domains, including production planning and control. In this paper, we investigate the benefits digital twins would bring to the sawmill industry via a literature review on the wider subject of sawmill production planning and control. Opportunities facilitating their implementation, as well as ongoing challenges from both academic and industrial perspectives, are also studied.

PURPOSE: Getting back to the same level of play as pre-injury is challenging for most athletes following anterior cruciate ligament reconstruction (ACLr). Data show that a third of athletes fail to reach pre-injury level after ACLr. This could be explained by the fact that the rehabilitation process is long, and often not standardized and sufficiently demanding between the 3rd and 6th month. A new intensive standardized program post ACLr (starting 3rd month post-surgery) specifically developed for football varsity athletes seems to lead to good clinical results and satisfying return to sport (RTS) rate and level. The purpose of this study was to compare, in amateur athletes, the rate of RTS at the pre-injury level between athletes completing this new program (EXP group) to the ones receiving usual rehabilitation care (URC) (CTL group). METHODS: Using a single-blind (evaluator) randomized clinical trial, 66 amateur athletes (18-35 years old; 3 months post ACLr; willing to return to sport) were randomly assigned to the EXP (aged 23.3 ± 3.25; 36.3% men) or CTL (aged 24.9 ± 4.92; 36.3% men) group. Participants of the EXP group trained with the varsity program while the CTL group received URC. The RTS rate at 1 year post ACLr was determined using two criteria; a) did they return to the same type of sport they practiced pre-injury; b) did they return at the same level of play as pre-injury. A full RTS was achieved when the answer to these 2 questions (the criteria for a return to sport to the pre-injury level) was yes. RankFD analyses were used to compare groups on full RTS 1 year after ACLr. RESULTS: There was no significant difference (p=: 0.840) in the one-year return to sport at pre-injury level between EXP (39%, 11/28 athletes) and CTL (37% - 11/30 athletes) groups. There was an overall 45% return to type 1 sports (EXP 40%, 10/25 and CTL 50%, 11/22, p = 0.198) and a 42% return to competitive or higher level (EXP 60%, 9/15 and CTL 28%, 5/18, p = 0.067). CONCLUSIONS: A new intensive rehabilitation program specifically developed for football varsity athletes does not seem to provide additional benefit regarding return to sport at pre-injury level for amateur athletes post ACLr. Further research is needed before discarding this new program since the lack of gym access and sport restriction during covid-19 pandemic might have influenced the findings.

In search and rescue operations, an efficient search path, colloquially understood as a path maximizing the probability of finding survivors, is more than a path planning problem. Maximizing the objective adequately, i.e., quickly enough and with sufficient realism, can have substantial positive impact in terms of human lives saved. In this paper, we address the problem of efficiently optimizing search paths in the context of the NP-hard optimal search path problem with visibility, based on search theory. To that end, we evaluate and develop ant colony optimization algorithm variants where the goal is to maximize the probability of finding a moving search object with Markovian motion, given a finite time horizon and finite resources (scans) to allocate to visible regions. Our empirical results, based on evaluating 96 variants of the metaheuristic with standard components tailored to the problem and using realistic size search environments, provide valuable insights regarding the best algorithm configurations. Furthermore, our best variants compare favorably, especially on the larger and more realistic instances, with a standard greedy heuristic and a state-of-the-art mixed-integer linear program solver. With this research, we add to the empirical body of evidence on an ant colony optimization algorithms configuration and applications, and pave the way to the implementation of search path optimization in operational decision support systems for search and rescue.(c) 2022 Elsevier B.V. All rights reserved.

In gas metal arc welding, a weld quality and performance depends on many parameters. Selecting the right ones can be complex, even for an expert. One generally proceeds through trial and error to find a good set of parameters. Therefore, the current experts’ method is not optimized and can require a lot of time and materials. We propose using supervised learning techniques to help experts in their decision-making. To that extent, a two-part recommendation system is proposed. The first step is dedicated to identify, through classification, the number of weld passes. The second one suggests the seven remaining parameter values for each pass: layer, amperage, voltage, wire feed rate, frequency offset, trimming and welding speed. After extracting data from historical Welding Procedure Specification forms, we tested 11 different supervised learning algorithms. The recommendation system is able to provide good results for all the different settings mentioned above even if the data is noisy due to the heuristic nature of the experts’ process. The best classification model is CatBoost with 82.22% average F1 Weighted-Score and the best regression models are Extra Trees or a boosting algorithm with a reduced mean absolute percentage error compared to our baseline.

Wood planers are high speed sophisticated lumber finishing machines that are difficult to operate and for which the available data shows complex, non-linear patterns. We present a machine learning approach to build a control loop for an industrial wood planer. In order to predict the thickness of the outgoing boards with better accuracy than the industry standard whilst allowing dynamic planer adjustments, we use an ensemble of Gaussian Processes with a specialized weighting scheme we call Automatic State Matching. It reduces the prediction error by 39% compared to current industrial practice.

We present a metamodeling approach, based on supervised learning, to estimate the probability of success of maritime search and rescue operations. The objective is to improve search planning in a context where lives are at risk and time is of the essence. The proposed approach has been evaluated both in terms of its predictive performance (Can the probability of success be closely approximated?), and of its added value to the decision support system operationally used by the Canadian Coast Guard (To what extent does the approach improve the current system?). We conducted extensive experimentations to evaluate and compare four machine learning algorithms namely, random forest, k-nearest neighbor, support vector machine regression, and feed forward neural networks with a single layer. Our experimental results, based on real-life data, show that the learned models can approximate the probability of success with sufficient precision and that a heuristic, implementing a k-nearest neighbor model within the existing decision support system, can recommend search plans with higher probabilities of success, which has the potential to save more lives.

Synthetic data generation of industrial processes exhibiting non-stationarity and complex, non-linear dependencies between their inputs and outputs is a challenging task. We argue that vine copula models are particularly well suited for this problem and present a method combining limited available data and expert knowledge in order to generate synthetic data by conditionally sampling from a C-Vine, a type of vine copula. We demonstrate our approach by generating synthetic data for a high-speed, sophisticated lumber finishing machine called a wood planer.