The 19th International Conference on Soft Computing Models in Industrial and Environmental Applications SOCO 2024

This paper presents a comprehensive study on the application of computer vision, specifically Convolutional Neural Networks (CNNs), for quality control in automotive interior manufacturing using Injection Molding Machines (IMMs). Despite the high precision and productivity of IMMs, a small percentage of defects persist. Addressing this, the paper compares four popular CNN algorithms—ResNet-50, SegNet, FCN, and PSPNet—on their ability to detect and analyze defects in post-manufacturing. The study outlines the advantages of IMMs, the challenges in defect management, and the potential of CNNs to revolutionize quality control processes. Experimental results demonstrate the efficacy of these algorithms, paving the way for enhanced quality management and reduced defect rates in the automotive sector.

This paper presents a proposed taxonomy for collaborative robotics, encompassing a detailed classification based on various technical and functional characteristics applicable to robotic equipment in the context of collaborative tasks. The proposed taxonomy includes thirteen categories and arises from the need for an initial reference framework for the analysis and classification of robots, allowing not only the identification of yet uncovered research topics but also serving as a starting point for classifying different types of current robots. This will help determine whether they can be used in shared spaces with people or not, given that today there are robots with unique characteristics that make it difficult to ascertain their suitability for use in shared spaces. This is due to the fine line between what can be considered a collaborative robot and one for general or industrial use.

This work focuses on proposing alternatives for the $$\gamma ^\varphi $$ γ φ function used to perform positional encoding in volumetric rendering methods based on radiance field optimization. The approach is to replace the standard sinusoidal function with other proposed alternatives having similar properties. Alternatives that combine several $$\gamma ^\varphi $$ γ φ functions are also tested. Experiments with quantitative and qualitative results are included to evaluate the results.

Significant advancements in machine learning in recent years have revolutionized multiple sectors. The Segment-Anything Model (SAM) is a notable example of state-of-the-art image segmentation. Despite claims of zero-shot generalization, SAM exhibits limitations in specific scenarios like medical mammography images. SAM generates three segmentation masks per image to address this and recommends selecting the one with the highest confidence score. However, this is not always the optimal choice. This paper introduces a system that extends SAM’s segmentation capabilities by automatically selecting the correct mask, leveraging few-shot learning methods and an Out-of-Distribution threshold strategy. Several backbones were subjected to experimentation, highlighting the relationship between the support set size and the model’s accuracy.

Breast cancer is one of the most significant issues today. Doctors use methods like mammograms for its detection. Early recognition of this disease is crucial, so automation using convolutional neural networks is proposed, classifying mammograms according to BI-RADS categories. The proposed methodology seeks to streamline image analysis through the automation capabilities of convolutional neuronal networks, presenting a swifter and more reliable alternative to manual assessment. Multiple class distributions were conducted on the dataset to enhance the model’s performance. These divisions provided a deeper understanding of the model and facilitated the identification of optimal parameter combinations and strategies to optimize its effectiveness. Experimental findings affirm the method’s proficiency in accurately classifying mammography images according to BI-RADS standards. This innovative approach holds promise for the development of automated systems to aid radiologists in the early detection of breast cancer.

In this study, we propose a novel feature selection technique to improve the classification of simultaneous anomalies in the context of predictive maintenance. We leverage three publicly available datasets collected from measurements and characterised by high dimensions. In this context, the use of machine learning algorithms with 5 multi-label classifiers has proven effective in anomaly classification tasks. Therefore, we implement a self-adaptive evolutionary strategy algorithm tasked with selecting the most relevant features. We assess the cost and the effectiveness of the classification task with and without feature selection. Overall, our results show that the self-adaptive evolutionary strategy is able to drastically reduce the number of features required for training, while improving or maintaining the performance of the classification task. Our study underscores the potential of feature selection for predictive maintenance multi-label classification tasks, contributing to more efficient and effective predictive maintenance strategies in industrial settings.

This study presents a novel architecture for Human Digital Twins (HDT) applied in manufacturing, designed to address several challenges in the current industry, including data management, safety, intelligent assistance, and operator training. We analysed various existing proposals, highlighting the need for adaptable and scalable architectures that support the complexity of industrial processes. Our proposed architecture integrates and processes multimodal data to facilitate the incorporation of human related information, as emotion detection, fatigue identification, and intelligent monitoring, among others. These technologies that are embedded in our proposal are crucial for improving operational efficiency and safety in the workplace. The proposed architecture stands out for its ability to adapt to various industrial applications, making it a viable solution for implementing digital twins in cyber-physical environments.

Sanitary professionals often experience burnout due to long working hours and the stress associated with handling sensitive situations. Consequently, in recent years, significant efforts have been made to automate various tasks within their roles using deep learning methods. This study proposes an automated system for generating summaries of medical consultations, aiming to alleviate doctors’ workloads. The architecture takes audio recordings of medical consultations as input, which are then transcribed and diarized for further processing. Subsequently, the entire transcription is summarized using state-of-the-art transformer-based models. For this study, a dataset comprising fourteen consultations, along with two summaries provided by different sanitary professionals, was utilized. In order to address the limitations of the available data, a data augmentation technique was implemented which makes use of the capabilities of large language models. The proposed models achieved a result of 0.52 in the ROGUE-1 metric, demonstrating the effectiveness of the proposed methodology.

This paper focuses on the application of a computer vision system for the detection of interlacing bricks within a robotic handling system, highlighting the challenges and solutions related to transitioning these systems from controlled laboratory environments to dynamic industrial manufacturing conditions. By addressing practical problems encountered in industrial settings, we emphasize the reliability issues of image-based object detection systems when integrated into automated robotic manipulation processes. The study includes the development and testing of various modified U-Net architectures for brick detection and evaluates their performance in both laboratory and industrial environments.

A computer-vision-based industrial algorithm is proposed in this study for the detection of the dimensions and the spatial positioning of fruit and vegetables on a conveyor belt for their movement to a packing machine with a robotic arm. The principal purpose of the algorithm is to identify the dimensions (length, width), and position (angle of inclination, and Cartesian coordinates) of the vegetable mesh netting without taking the product label into account. The proposed algorithm has four functions. A convolutive neuronal network model is applied for object detection, with which all objects are identified in the image while the product label is suppressed, so that recognition of the vegetable mesh netting and its dimensions is not impaired. Moreover, the Canny edge detector and the Border following algorithms are applied to perform image pre-processing and edge improvements, respectively, yielding optimal results with more clearly defined noise-free objects. The feature extraction function for dimensions yielded detection precision results of 99.60% for length, 99.00% for width, and 97.80% for inclination angle, showing optimal performance in the test phases.

Early identification of plant diseases plays an important role in reducing the financial losses for farmers and enhancing the productivity and quality of crops. The present work proposes five machine learning (ML) algorithms: K-Nearest Neighbors, Random Forest, Iterative Dichotomizer 3, Adaptive Boosting and Logistic Regression for plant disease identification. In order to test the proposed ML algorithms, we used a large data set containing 74798 images mainly from PlantVillage dataset. The achieved experimental results show that the proposed ML techniques can be used to reliably identify various plant diseases.

The article explores the role of proximity and collaborative networks in shaping smart specialization strategies for regional development. The study introduces a mathematical model for proximity contagion, which considers population, gross value added, and distance between localities to quantify the influence of proximity on smart specialization. The article discusses experimental insights from applying the model to the North-Western region of Romania, highlighting the dynamic interplay between local strengths, collaborative networks, and proximity influences. The findings underscore the need for adaptive strategies and supportive policies to leverage proximity advantages effectively and enhance innovation-driven economic growth at the regional level.

In this study, we tackle a local collection challenge, employing VRP to address the problem of collecting fresh agricultural products from administrative territorial units. This paper proposes a comparative description of two solution approaches: a haploid genetic algorithm (HGA) and a pseudo-diploid genetic algorithm (P-DGA). We compare the achieved results by the considered genetic algorithms (GAs) on a set of benchmark instances existing from the literature, analyze their performance against state-of-the-art algorithms, and validate them on a second set of instances provided by the Maramureş County Directorate for Agriculture, that contains information regarding the production of tomatoes, onions, and garlic from the administrative units within Maramureş County.

This study use machine learning (ML) to reveal the intrinsic economic specializations in different localities by analyzing a comprehensive dataset of economic indicators, such as gross added value (GVA) and revenue to the local budget. The results of our research demonstrate that machine learning algorithms may accurately detect both well-established and growing economic specializations, hence offering more profound understanding of regional economies. Moreover, supervised learning models exhibit the ability to forecast forthcoming economic patterns, so showcasing the predictive prowess of machine learning in economic analysis. Ultimately, this will assist policymakers in making well-informed decisions.

The rise of e-commerce, driven by digital business models, has made online shopping easier and more convenient. While this benefits both businesses and shoppers, the increase in online transactions also attracts fraudsters. In this work, we address the identification of fraudulent transactions using a machine learning approach. For this purpose we analyze the DataCo Supply Chain Dataset and apply a two-step preprocessing workflow to identify non-informative features. The predictive power of the resulting dataset is tested and compared with classification models induced after applying a feature selection approach. Results suggest that a small number of factors are helpful to build highly predictive models capable of identifying fraudulent transactions.

This paper presents a novel approach for calculating average weights in Federated Learning. This technique integrates data-related measures, such as Data Quality, Data Freshness, and Data Importance. These changes provide a more accurate and pertinent averaging procedure, resulting in improved models and more relevant model output. This technology is tested and validated within an agricultural scenario, that employs a Low-Communication paradigm. Additionally, the metadata serves as a backup alerting mechanism to detect any issues inside the system.

Land bonitation is a critical process for assessing land quality, particularly for agricultural and sustainable use, based on factors like soil quality, agricultural potential, and management practices. This article presents an innovative approach that integrates Geographic Information Systems (GIS) with Romanian bonitation standards to improve agricultural land assessment accuracy. The study leverages GIS with crop yield estimation models, remote sensing imagery, and crop simulation, demonstrating robust performance in yield estimation and optimizing water efficiency. The proposed algorithm uses GIS data to calculate bonitation scores based on 17 class indicators, facilitating the evaluation of land suitability for agricultural purposes. Experimental results highlight how grid point density affects computation time and demonstrate the efficacy of machine learning techniques in predicting environmental factors such as humidity. The study underscores the importance of GIS-based bonitation in tackling land fragmentation and environmental variability. Future research will focus on integrating advanced machine learning techniques for more precise assessments and improved crop production.

Crop yield prediction is a critical aspect of agricultural research, with implications for decision-making in areas such as import-export, pricing, and distribution of specific crops. Machine learning algorithms have emerged as effective instruments for enhancing the accuracy of crop yield prediction models in recent years. This article presents a state-of-the-art overview of smart agriculture and the latest machine learning algorithms used for crop yield prediction with emphasis on European Union. The article concludes by highlighting the significance of machine learning algorithms in precision crop management systems and their potential to revolutionize agricultural decision-making. Overall, this comprehensive review provides valuable insights into the latest advancements in machine learning algorithms for crop yield prediction, offering a foundation for future research and practical applications in agriculture.

This paper presents a the development and optimization of a Convolutional Neural Network (CNN) model, designed for Tomato Plant disease and pest Classification (TPC_Net). Using images from the PlantVillage and TomatoVillage datasets, the model was trained on a balanced subset created through augmentation techniques to enhance accuracy and generalizability. We compared TPC_Net with established models adapted for tomato disease classification, demonstrating its superior accuracy, precision, and recall in identifying 11 distinct classes of diseases and pests. The model’s streamlined architecture facilitates deployment in mobile applications, promising significant advancements in agricultural technology for effective disease management.

This article introduces a novel Data Quality Assessment Methodology (DQAM) tailored to the challenges of Big Data and Machine Learning (ML), particularly in the context of Federated Learning. DQAM offers a standardised, automated approach to evaluate data quality, emphasising completeness, information entropy, data symmetry, and time-series continuity. Implemented in Python and compatible with Pandas DataFrames, DQAM provides rapid, scalable assessment capabilities crucial for modern ML pipelines. Practical validation of DQAM in agriculture industry use cases demonstrates its effectiveness in identifying and quantifying data quality issues, paving the way for more reliable insights and decision-making. By addressing the lack of standardization and automation in existing approaches, DQAM contributes to enhancing the reliability and accuracy of ML models.

Machine Learning and Data Mining imply notable privacy vulnerabilities since they have the potential to expose confidential details about people or collectives that have contributed to the data. This paper proposes a Generalised Privacy Assessment Methodology that can be universally applied to any Machine Learning model and its corresponding Data Source. The suggested technique facilitates the quantification of the Vulnerability Score for any given data source or model. We demonstrate the practicality and efficacy of the concept by implementing it in several scenarios, ranging from a small-scale IoT project to a hypothetical extensive database, like those used by governmental entities.

With the serious danger to nature and humanity that forest fires are, taken into consideration, this work aims to develop an artificial intelligence model capable of accurately predicting the forest fire risk in a certain region based on four different factors: temperature, wind speed, rain and humidity. Thus, three models were created using three different approaches: Artificial Neural Networks (ANN), Random Forest (RF), and K-Nearest Neighbor (KNN), and making use of an Algerian forest fire dataset. The ANN and RF both achieved high accuracy results of 97%, while the KNN achieved a slightly lower average of 91%.

This study explores the enhancement of API security testing against Broken Object Level Authorization (BOLA) and Authentication vulnerabilities through the integration of an LLM (Large Language Model)-enhanced framework. By incorporating the OpenHermes 2.5 Mistral 7B model, fine-tuned for domain-specific vulnerabilities, into the Karate testing framework, we demonstrate a novel approach to generating comprehensive and accurate test scenarios. Our methodology emphasizes prompt refinement, dynamic data handling, and endpoint context recognition to address the complexities of API security testing. A case study utilizing the VAmPI API Specification reveals significant improvements in test case generation and complexity, highlighting the potential of fine-tuned LLMs in identifying and mitigating API vulnerabilities. While promising, our exploration uncovers challenges in dataset development, token allocation, and mock data reliance, suggesting areas for future research, including the adoption of Retrieval Augmented Generation (RAG) and Retrieval Aware Fine-Tuning (RAFT) methods. Our findings underscore the transformative impact of LLMs on API security testing, paving the way for more robust and efficient testing frameworks.

Genetic Algorithms are a valuable combinatorial optimization tool, inspired by evolutionary principles, which offer broad prospects for improvement and research. The present paper addresses the impact of two natural biological foundations, namely the age-based selection of individuals and the concept of sexuate crossover. By combining these two concepts, the exploration of the solution space is promoted through age-dependent self-adaptive mutation or the transmission of promising genes to offspring. These approaches were applied and validated through benchmark functions that demonstrated the effectiveness of this algorithm, providing improved results in terms of runtime and quality of solutions found.

Accelerometry data are useful in the diagnosis and treatment of mental illnesses. However, the correct application of these data depends on their quality, and in the case of mental health data, it is essential to characterize the state of euthymia in order to distinguish critical states such as depression or mania. To do this, it is necessary to delimit quality time periods in the follow-up, especially those periods in which the patient is free of seizures. In this work, an algorithm suitable for isolating quality data useful for the characterization of the state of euthymia is presented. The algorithm is based on the robustness of the classification results of the data vectors.

In this paper, a controller based on a soft computing technique, specifically fuzzy logic, is presented and applied to a wind turbine. In the case of floating offshore wind turbines (FOWT), environmental disturbances, especially strong wind and waves, produce structural loads that result in fatigue and damage. In this paper, a fuzzy controller is proposed that, in addition to optimizing power generation in the sub-nominal wind speed operation region of a FOWT, also reduces tower vibrations. The control strategy uses the angular velocity of the generator and the acceleration of the top of the tower to obtain the output torque needed to achieve system performance and improve stability. This fuzzy control is compared with the OpenFAST speed-torque control, achieving a suppression rate of 32% while maintaining the output power.

The urgent need to reduce carbon dioxide emissions in energy generation poses a significant challenge for modern society. Consequently, it is of paramount importance to intensify efforts to enhance clean energy sources, such as wind energy. In this context, addressing the inherent variability of wind power through prediction methodologies is a compelling field of study. Additionally, this clean energy should be accessible to everyone, even on a small scale, by leveraging affordable devices like the Raspberry Pi and other low-cost hardware platforms. This study aims to evaluate the effectiveness of machine learning (ML) algorithms, with a particular focus on deep learning models, in accurately forecasting wind turbine power output. Specifically, the research uses Long Short-Term Memory (LSTM) networks that are designed and trained using conventional computing resources, but evaluated on an affordable computing system like the Raspberry Pi 3 with 2 GB of RAM to facilitate the management of energy generation. Through a comparative analysis, considering precision and real-time performance, the study identifies the optimal parameters for accurately modeling time series data related to wind energy production and evaluates its implementation. Our findings demonstrate that effective wind power forecasting can be achieved on low-cost hardware platforms, highlighting the potential for widespread adoption and personal management of wind energy generation.

Wind energy has emerged as a crucial solution in the search for sustainable and environmentally friendly energy sources. As global energy demand grows rapidly and the need to reduce greenhouse gas emissions increases, wind turbines have become a key technology for the transition to a cleaner and more sustainable energy future. This paper focuses on the application of autoencoder neural networks for the accurate and efficient detection of anomalies in wind turbines. The main objective is to significantly improve monitoring and predictive maintenance of wind turbines and thus enhance wind energy availability. The dataset used includes records of wind turbines with both, healthy and faulty conditions. Healthy records serve as a baseline to illustrate the expected amplitude of the turbine vibrations under optimal operation conditions. In contrast, failure records arising from blade surface erosion, blade unbalanced, and blade twist are used to evaluate the failure detection model. The proposed algorithm demonstrates a precision of 95.42% for blade unbalanced failures, 94.63% for blade surface erosion, and 87.54% for blade twist faults.

The increasing concern about climate change has contributed to promoting renewable energy technologies. Mini-eolic turbines are a common solution for domestic energy supply self-consumption installations. Due to this technology’s strong dependency on climate conditions and corresponding variability, it is important to develop intelligent systems to model and estimate its behavior. This paper uses three different feature selection methods, a clustering algorithm, and a regression technique to predict the power generated by a small wind turbine located in a bioclimatic house. Different configurations are tested, evaluating the impact on the model performance.

This paper introduces a novel approach to ranking-based feature selection. Numerous metrics exist to assess univariate filters. This contribution identifies the top best features by employing two distinct methods, computes the global weights of each feature, and sorts these weights in decreasing order for those which are common to both procedures. According to the first part, this work is a feature selection proposal and concerning the last, it is besides a feature sorting approach. The benchmark data sets originate from NIPS 2003. This approach has been tested on problems with up to 20000 features and up to 7000 samples. The test results are very promising since, in some cases the outcomes significantly outperform the raw scenario, and in others, they surpass those of an approach which may be considered to be related to the current contribution. Finally, this study opens new avenues for research, particularly in examining the importance of specific feature sorting methods for certain supervised machine learning algorithms.

Battery technology is advancing due to the demand of safe energy storage in systems such as Electric Vehicles (EVs) and portable electronics systems. Estimating battery conditions and increasing the range of electric vehicles requires specific models and optimize control strategies. Anticipate battery performance is complicated due to several variables, such as temperature, deterioration, and usage models. Knowing battery performance entails an advanced understanding of time series analysis, particularly forecasting. This study proposes the use of Artificial Neural Networks (ANN) to create a Lithium Ion Battery (LIB) prediction model. The experiments were carried out in a 59.2 VDC, 120 Ah LIB, recording the voltage and temperature while simulating the charge and discharge. A hybrid model is developed with internal local models using clustering. The model prognostic the State of Charge (SOC) evolution and reach a Mean Squared Errors with values between $$10^{-3}$$ 10 - 3 and $$10^{-4}$$ 10 - 4 . This research shows that hybrid models are adecuated instrument for predict the battery SOC.