Industrial plants are major sources of environmental noise, producing complex and high-intensity acoustic emissions that vary across different operational conditions.Automatically characterizing the sources that generate harmful acoustic emissions is crucial to take the necessary actions to reduce them.However, manually labeling these sounds is impractical due to their volume and variability.In this study, we employ an unsupervised deep learning framework for clustering industrial sound emissions in steelmaking plants, focusing on areas such as the hot rolling mill, Electric Arc Furnace, and scrapyard.The approach integrates Variational Autoencoders with Gaussian Mixture Models to learn compact latent representations from Mel-spectrogram features of raw, unlabelled audio data.We compare this approach to traditional clustering techniques such as K-means and GMM, as well as Deep Embedding Clustering.The results demonstrate that the approach significantly outperforms traditional methods, offering reliable and interpretable clustering of industrial acoustic events.This research contributes to the development of automated, efficient, and sustainable noise-monitoring systems for industrial operations, addressing key challenges in environmental noise monitoring.* The work described in the present paper was developed within the project Real-time acoustic sensorS and artificial Intelligence appLications for the rEduction of local eNvironmental impaCt due to noise Emissions (SILENCE) (Grant agreement No. 101112516) -that has received funding from the RFCS programme, which is gratefully acknowledged.The sole responsibility of the issues treated in the present paper lies with the authors; the Commission is not responsible for any use that may be made of the information contained therein.

Abstract Within the design of a machine learning-based solution for classification or regression problems, variable selection techniques are often applied to identify the input variables, which mainly affect the considered target. The selection of such variables provides very interesting advantages, such as lower complexity of the model and of the learning algorithm, reduction of computational time and improvement of performances. Moreover, variable selection is useful to gain a profound knowledge of the considered problem. High correlation in variables often produces multiple subsets of equally optimal variables, which makes the traditional method of variable selection unstable, leading to instability and reducing the confidence of selected variables. Stability identifies the reproducibility power of the variable selection method. Therefore, having a high stability is as important as the high precision of the developed model. The paper presents an automatic procedure for variable selection in classification (binary and multi-class) and regression tasks, which provides an optimal stability index without requiring any a priori information on data. The proposed approach has been tested on different small datasets, which are unstable by nature, and has achieved satisfactory results.

This article presents the application of a recent neural network topology known as the deep echo state network to the prediction and modeling of strongly nonlinear systems typical of the process industry. The article analyzes the results by introducing a comparison with one of the most common and efficient topologies, the long short-term memories, in order to highlight the strengths and weaknesses of a reservoir computing approach compared to one currently considered as a standard of recurrent neural network. As benchmark application, two specific processes common in the integrated steelworks are selected, with the purpose of forecasting the future energy exchanges and transformations. The procedures of training, validation and test are based on data analysis, outlier detection and reconciliation and variable selection starting from real field industrial data. The analysis of results shows the effectiveness of deep echo state networks and their strong forecasting capabilities with respect to standard recurrent methodologies both in terms of training procedures and accuracy. Supplementary Information: The online version contains supplementary material available at 10.1007/s00521-021-05984-x.

In the era of European Green Deal, steelworks are committed to reduce their CO2 emissions by preserving their competitiveness.One of the options to achieve such aim is the valorization of process off-gases.Methane and Methanol production can be obtained by coupling novel reactors with an advanced control system that dispatches these gases after enrichment with green hydrogen.Knowing in advance the gases availability and composition is fundamental.The paper present Echo State Networks based-models that are applied to this aim and achieve adequate forecasting accuracy also in case of highly dynamic processes.

Abstract The paper proposes an approach to the design of the chemical composition of steel, which is based on neural networks and genetic algorithms and aims at achieving a desired hardenability behavior possibly matching other constraints related to the steel production. Hardenability is a mechanical feature of steel, which is extremely relevant for a wide range of steel applications and refers to the steel capability to improve its hardness following a heat treatment. In the proposed approach, a neural-network-based predictor of the so-called Jominy hardenability profile is exploited, and an optimization problem is formulated, where the optimization function allows taking into account both the desired accuracy in meeting the target Jominy profile and other constraint. The optimization is performed through genetic algorithms. Numerical results are presented and discussed, showing the efficiency of the proposed approach together with its flexibility and easy customization with respect to the user demands and production objectives.

The paper describes an approach for the preprocessing of the dataset to be used for the training of an arbitrary classifier in the context of binary classification of unbalanced datasets. The proposed method is based on the reduction of the number of frequent samples that are included in the training dataset in order to mitigate the unbalance rate and to promote the correct detection of the rare samples, which are often the most interesting ones in relation to the application the classifier is designed for. The proposed technique exploits two Self Organizing Maps that clusterize the rare and frequent samples in the original training dataset. The obtained clusterizations are exploited in order to determine the frequent samples whose removal is most convenient. The approach has been successfully tested and compared to other well known resampling techniques by exploiting different datasets coming from both the UCI repository and real industrial applications.

Sequential Forward Selection is one of the most popular variable selection methods and is widely adopted in many applications, among which the design of classifiers. The main problem related to this approach regards the stability of the selection, namely the exploitation of different training data set can lead to the selection of different variable subsets. This problem is particularly critical in applications where variable selection is used to interpret the behaviour of the process or phenomenon under consideration, i.e. to understand which among a potentially huge list of variables actually affect the classification. The paper proposes a method that improves the stability of the sequential forward selection procedure while preserving and possibly improving the classification performance.

The estimate of the probability density function or probability mass function of an unknown stochastic process is a very important preliminary step for any further elaboration. Most of the traditional approaches to this problem perform a preliminary choice of a parametric mathematical model of the function to estimate and a subsequent fitting on its parameters. To this aim some a-priori knowledge and/or assumptions on the phenomenon under consideration are needed. In this paper an alternative approach is proposed, which does not require any assumption on the available data, as it extracts the probability density function from the output of a neural network, that is trained with a suitable database including the original data and some ad hoc created data with known distribution. The results of the tests performed on synthetic and industrial databases are described and discussed in the paper.

Data preprocessing is a main step in data mining because real data can be corrupted for different causes and high performance data mining systems require high quality data. When a database is used for training a neural network, a fuzzy system or a ne

The paper deals a novel resampling method in order to cope with imbalanced dataset in binary classification problems. Imbalanced datasets are frequently found in many industrial applications: for instance, the occurrence of particular product defects or machine faults are rare events whose detection is of utmost importance. In this paper a new resampling method combining an oversampling and an undersampling techniques is treated. In order to prove the effectiveness of the proposed approach, several tests have been developed. Two classifiers based on Support Vector Machine and Decision Tree have been designed, which are applied for binary classification on four datasets: a synthetic dataset, a widely used public dataset and two industrial datasets. The obtained results are presented and discussed in the paper; in particular, the performance that is achieved by the two classifiers through our resampling approach is compared to the ones that are obtained without any resampling and through the classical SMOTE approach, respectively.

Recently in many industrial fields the exploitation of vision systems for quality control had a considerable increase, which is mainly due to the technological progress experienced by such systems, that, with respect to the past, made their performance more appealing and more reliable while the associated costs are decreased. The advantages of these kind of systems in terms of savings in human resources and improved quality monitoring have become far more evident, by encouraging their adoption in a wide variety of production cycles. The present paper deals with the elaboration and information extraction from images, that represent portions of the surface of flat steel products, and describes an algorithm for defect detection and classification. The overall classification procedure is composed of a preliminary part that is mostly related to image processing and analysis, which aims at pointing out the defect (independently on the class it belongs to), as well as to the extraction of relevant features of the detected defect; the second part exploits a fuzzy inference system in order to analyze the type of defect and solves a classification problem that presently can be addressed only with the support of a human operator. Fuzzy inference systems are suitable to this application because they are able to mimic and reproduce the human reasoning.

Distribution fitting is a widely recurring problem in different fields such as telecommunication, finance and economics, sociology, physics, etc. Standard methods often require solving difficult equations systems or investments in specialized software. The paper presents a new approach to distribution fitting that exploits Genetic Algorithms in order to simultaneously identify the distribution type and tune its parameters by exploiting a dataset sampled from the observed random variable and a set of distribution families. The strength of this approach lies in the easiness of the expansion of this set: in fact distributions are simply described by means of their probability density functions and cumulative distribution functions, which are well-known. This approach employs two different score metrics, the Mean Absolute Error and the Kolmogorov-Smirnov test, that are linearly combined in order to find the best fitting distribution. The results obtained in an industrial application are presented and discussed.

The paper presents an application of genetic algorithms to the problem of input variables selection for the design of neural systems. The basic idea of the proposed method lies in the use of genetic algorithms in order to select the set of variables to be fed to the neural networks. However, the main concept behind this approach is far more general and does not depend on the particular adopted model: it can be used for a wide category of systems, also non-neural, and with a variety of performance indicators. The proposed method has been tested on a simple case study, in order to demonstrate its effectiveness. The results obtained in the processing of experimental data are presented and discussed.

When designing a neural or fuzzy system, a careful preprocessing of the database is of utmost importance in order to produce a trustable system. In function approximation applications, when a functional relationship between input and output variables is supposed to exist, the presence of data where the similar set of input variables is associated to very different values of the output is not always beneficial for the final system to design. A method is presented which can be used to detect anomalous data, namely non-coherent associations between input and output patterns. This technique, by mean of a comparison between two distance matrix associated to the input and output patterns, is able to detect elements in a dataset, where similar values of input variables are associated to quite different output values. A numerical example and a more complex application in the pre-processing of data coming from an industrial database were presented.

The paper analyses the issues behind strategies optimization of an existing automated warehouse for the steelmaking industry. Genetic algorithms are employed to this purpose by deriving a custom chromosome structure as well as ad-hoc crossover and mutation operators. A comparison between three different solutions able to deal with multiobjective optimization are presented: the first approach is based on a common linear weighting function that combines different objectives; in the second, a fuzzy system is used to aggregate objective functions, while in the last the strength Pareto genetic algorithm is applied in order to exploit a real multiobjective optimization. These three approaches are described and results are presented in order to highlight benefits and pitfalls of each technique.

Wavelet theory is a very powerful tool for the analysis and synthesis of signals. In particular, the Haar basis is adopted, the computational effort required by the wavelet transform, is not high and thus suitable for real-time implementations even when low computing power is available. We investigate the application of a wavelet-based controller to a mechatronic drill for orthopaedic surgery. The penetration velocity of the drill is generated on the basis of the wavelet analysis of the thrust force signal and the controller is capable of fulfilling three different tasks corresponding to different specifications of the hole to be done in a long bone. Results from real experimental data are presented and discussed.