The increase in the use of wood in general, and pellets in particular, for individual heating requires attention to be paid to optimizing the combustion of these pellets, particularly in terms of pollutant emissions such as carbon monoxide (CO). This quest for optimization is complicated by the intrinsic variability of wood and, therefore, pellets. In this context, this paper proposes building a neural network model to predict the CO content in smoke based on pellet characteristics and stoves characteristics and settings. However, experiments are costly and time-consuming, which limits the size of the available dataset. In this context, we propose a methodology aimed at training a multilayer perceptron using a small dataset while reducing the risk of overfitting. This methodology is based on finding the minimal network structure and using a robust learning algorithm. The results show that the robust algorithm effectively limits the risk of overfitting and that the final model retains its generalization capabilities despite the small size of the dataset.

Many challenges have been raised in the scientific literature regarding the development of digital twins that can predict future states of production processes from data streams. This study is concerned with the coordination of several of their submodels to balance precision with computational requirements. A method to use stream-based active learning sampling strategies to couple two such models is proposed. Both models perform the same prediction task but have different advantages and disadvantages. The first is a simulation model that is supposed to have high fidelity level, but to be slow. The second is a machine learning model, which is fast but less accurate and requires many labeled examples to be trained on, which may require a lot of time and effort to gather. The objective is to leverage confidence measures in the predictions of the machine learning model. These measures are used to couple the two models and take advantage of their respective strengths. In particular, the aim is to reduce the digital twin's average prediction error while operating under limited computational capacity. Moreover, an application within the sawmill industry and numerical experiments are presented.

This paper delves into classification tasks, where data is categorized into binary classes, such as fraudulent/non-fraudulent or sick/not sick as example. Employing a statistical approach, this task entails utilizing hypothesis testing. Tuning this test involves selecting an acceptable risk alpha (associated with false positives), thereby implicating a beta risk (related to false negatives). In classification challenges, the principal aim is to mitigate the misclassification rate. However, the determination of these two risks is not be discretionary but rather enforced by the learning process, particularly evident when employing neural networks. This paper seeks to propose a modification of the learning algorithm for multilayer perceptron aimed at effectively balancing these risks. This adaptation hinges on leveraging a weighted criterion to minimize errors, accounting for the signs of different error types. This methodology is assessed across two benchmarks: a simulated dataset and a genuine medical dataset. Keywords: neural network, multilayer perceptron, learning, classification, hypothesis test

This article studies a method to couple two digital models in the context of digital twins. The first model is a simulation model which is supposed to be very accurate but computationally intensive. The second is a fast but approximate machine-learning model of the simulation. Both models serve, therefore, the same prediction task in an online environment but have different advantages and drawbacks. An object-oriented architecture is introduced to implement the proposed coupling strategy. Numerical experiment results on four datasets are also provided to evaluate the performances of the proposed strategy and compare it with a baseline. Three of these datasets originate from the University of California, Irvine machine learning repository. The last one originates from the Canadian forest product industry and contains the outputs of sawing simulation for real wood logs. These experiments demonstrate that the proposed method allows to consistently reduce the average error of the couple predictions.

Short and mid-term production planning and control in the sawmill industry are complicated by several sources of uncertainties. Consequently, it is difficult to predict in advance what set of lumber would be obtained from a specific log. Even if sawmill simulators that can simulate the sawing of a log from a scan of its profile exist, they can be extremely computationally intensive. Several alternative methods, based on machine learning algorithms using different sets of log descriptors were explored in previous works. This paper proposes the usage of multi-layer perceptrons, as well as a vector of features based on the pairwise dissimilarities from the log scans to a set of selected representative logs, chosen as the class medoids. Several MLP architectures are tested and compared on two different datasets with a previously proposed k-Nearest Neighbors algorithm to validate the performance of the proposed set of medoid-based features. While the best-performing architecture depends on the dataset considered, all MLP models demonstrate lower RMSE than the baselines.

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.

Digital Twins (DT) have been introduced as promising decision support tools in many different settings and serve a variety of purposes. Many challenges are raised by their development, including an efficient usage of their computational resources to balance performance on precision, computational cost and speed. This study is, in particular, concerned with Digital Shadows (DS), a concept derived from DT, applied to sawmills sawing production units. A method to combine a computationally intensive sawmill simulation model with a machine learning model is proposed to predict the set of lumbers sawed from logs. Numeric experiments are exposed, and the proposed method demonstrates improvements from 11% to 18% of the monitored couple regret from its baseline.

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.

Digital Twins (DT) is an extremely promising framework developed in the context of Industry 4.0 to facilitate the convergence of the physical and digital spaces. Numerous challenges remain, however, in terms of development, deployment, and self-adaptability of the DT faced with changes from its physical twin. Concerning this last point in particular, the set of Machine Learning (ML) methods known as Active Learning appears promising. This framework allows the DT to play an active role in the selection of the data samples used to train supervised ML models. This paper proposes a use-case inspired from the sawmill industry to illustrate the interest of these method in the presence of various changes in the flow of data gathered by the DT.

: The drying operation is the most energy consuming step of particle board manufacturing process. Even if a great academic and industrial effort has been furnished for last years, the prediction of this energy consumption is still a challenging issue. This paper deals with the energy consumption prediction for industrial wood drying. The study of an European particle board manufacturer’s industrial dryers has provided data sets for two both fresh and recycled wood drying processes. Based on these, MLP Neural network models have been developed and tested. Several tests have been conduced to identify and select the best MLP model’s structure to find a satisfying trade-off between model accuracy and maintenance efficiency. The proposed MLP models have either been distinctly trained on the datasets from both the first and second dryers, and then on their combination, in order to increase data diversity and to reduce training time and model maintenance. Then, the neural network based on the merged dataset has been compared to those developed from the single datasets. This experiment led to the conclusion that, the construction of a global model representing the operation of the two dryers is less accurate than the construction of a dedicated model for each dryer. Yet, the performances of combination model remain acceptable.

HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. A comparison of wood log dissimilarities to predict sawmill output with k-Nearest Neighbor algorithms Sylvain Chabanet, Mathis Dumas, Hind Bril El-Haouzi, Philippe Thomas

The Canadian wood industry use sawing simulators to digitally break a log into a basket of lumbers. However, those simulators tend to be computationally intensive. In some cases, this renders them impractical as decision support tools. Such a use case is the problem of dispatching large volume of wood to several sawmills in order to maximise total yield in dollars. Fast machine learning metamodels were recently proposed to address this issue. However, the approach needs a feature extraction step which could result in a loss of information. Conversely, it was proposed to directly make use of the raw information, available in the 3D scans of the logs typically used by a recent sawmill simulator, in order to retain that information. Here, we improve upon that method by reducing the computational cost incidental with the processing of those raw scans.

Several sawmill simulators exist in the forest-product industry. They are able to simulate the sawing of a log to generate the set of lumbers that would be obtained by transforming a log at a sawmill. In particular, such simulators are able to use a 3D scan of the exterior shape of the logs as input for the simulation. However, it was observed that they can be computationally intensive. Therefore, several authors have proposed to use Artificial Intelligence metamodel, which, in general, can make predictions extremely fast once trained. Such models can approximate the results of a simulator using a vector of descriptive features representing a log, or, alternatively, the full 3D log scans. This paper proposes to use dissimilarity to representative log scans as features to train a Machine Learning classifier. The concept of class Medoids as representative elements of a class will be presented, and a Simlarity Discrimant Analysis was chosen as a good candidate ML classier. This classifier will be compared with two others models studied by the authors.

Predicting the set of lumbers which would be obtained from sawing a log at a specific sawmill is a difficult problem, which complicates short and mid term decision making in this industry. While sawmill simulators able to simulate the sawing of a log from a 3D scan of its outer shape exist, they can be extremely computationally intensive. Several alternative approaches based on machine learning algorithms and different set of features were explored in previous works. This paper proposes the use of one hidden layer perceptrons, and a vector of features build from dissimilarities from the scans to a set of selected wood logs, chosen as the class medoids. Several architectures are tested and compared to validate the pertinence of the proposed set of medoid-based features. The lowest mean squared error was obtained for MISO neural networks with a sigmoid output activation function, to constrain the output value ranges.

We tackle the problem of predicting the lumber products resulting from the break down of the logs at a given sawmill.Although previous studies have shown that supervised learning is well suited for that prediction problem, to our knowledge, there exists only one approach using the 3D log scans as inputs and it is based on the iterative closest-point algorithm.In this paper, we evaluate the combination of neural network architectures (multilayer perceptron, residual network and PointNet) and log representation as input (industry know-how-based features, 2D projections, and 3D point clouds) in the context of lumber production prediction.Our study not only shows that it is possible to predict the output of a sawmill using neural networks, but also that there is value in combining industry know-how-based features and 3D point clouds in various network architectures.

Although digital simulations are becoming increasingly important in the industrial world owing to the transition toward Industry 4.0, as well as the development of digital twin technologies, they have become increasingly computationally intensive. Many authors have proposed the use of machine learning (ML) metamodels to alleviate this cost and take advantage of the enormous amount of data that are currently available in industry. In an industrial context, it is necessary to continuously train predictive models integrated into decision support systems to ensure the consistency of their prediction quality over time. This led the authors to investigate active learning (AL) concepts in the particular context of the sawmilling industry. In this paper, a method based on AL is proposed to combine simulation and an ML metamodel that is trained incrementally using only selected data (smart data). A case study based on the sawmilling industry and experiments are shown, the results of which prove the possible advantages of this approach. (c) 2021 Elsevier B.V. All rights reserved.

The scheduling problem in manufacturing companies with high rework rates remains a complex research area to date. This paper presents a new approach for manufacturing scheduling that combines a predictive schedule with a proactive multicriteria decision-making method based on smart batches and their quality prediction capability. Each batch embeds an algorithm that allows it to predict its quality out of the next workstation. As soon as a batch determines that its process is too hazardous, a collaborative rescheduling decision, using the analytic hierarchy process (AHP), is initiated with its peer. This article details the proposed approach along with the AHP structure and presents the considered decision problem. A simulation model inspired by a lacquering-robot case study is described to validate this proposition. Then, the results of different scenarios are presented and discussed, highlighting the impact of social myopia on smart batches.