A novel extract-transform-load methodology applying computer vision to build electrically equivalent models of electricity grids (i.e., digital twins of smart grids) is proposed. Here, we standardise information from electricity grids in various formats into a single visual input for object and text recognition, transformation and loading into a computer simulator. The proposed method is tested using state of the art computer vision algorithms: YOLOv8 for object detection (re-trained using two generic datasets of geometric figures) and PaddleOCR for text recognition. Validation is performed using two real electricity distribution networks in the centre-south region of Colombia, with a global object detection accuracy of up to 80.4%. Initial results suggest that the proposed method significantly reduces human effort compared to two classic approaches. The benefits and shortcomings of the proposed method are discussed together with future work opportunities.
In recent years, both academia and industry have witnessed significant advancements in Large Language Models (LLMs) research, with models like ChatGPT garnering extensive attention from society. These advancements in LLM technology have exerted a profound influence on the entire AI community, potentially revolutionizing how we design and utilize AI systems. Among Natural Language Processing (NLP) tasks, Question Answering (QA) has gained increasing attention. In this paper, we provide an overview of the advancements in LLMs, covering background, major findings, and fine-tuning experiments conducted from BERT to XLNet models on the SQuAD v1.1 and SQuAD v2.0 datasets for the QA task. Our evaluation of various encoder-only models on SQuAD tasks reveals that RoBERTa consistently demonstrates the best performance, achieving the highest Exact Match (EM) and F1 scores on both the SQuAD 1.1 and SQuAD 2.0 datasets. Additionally, we find that Flan T5-base yields even better results, boasting an EM of 77.9% and an F1 Score of 81.2%.
Technological advancements are increasingly evident across various sectors, including automobiles, industry, and healthcare. In precision agriculture, significant progress has been made, with AgroTICs and Smart Agriculture gaining substantial traction in the market. However, a gap remains between cutting-edge technology and family farming, presenting a challenge from both social and applied research perspectives. However, there is still a gap between cutting-edge technology and family farming, which creates a challenge from a social and applied research point of view. In this context, this paper proposes a monitoring model based on Fuzzy Logic and sensor automation applied to estimate the health of a corn crop. The proposed Fuzzy inference system involves calculating an indicator of nutrients as well as the average color and area of corn plants. The nutrient indicator is automatically computed by an ESP32 microcontroller using sensor readings, while the average color and area inputs are manually entered via a mobile application. Additionally, the Fuzzy inference is integrated into the ESP32. The model underwent experimental validation on the health of the plantation, and the results were evaluated in four areas: one was designated for testing, and three were for validation. The model achieved an accuracy of 97.5% in Scenario 3, categorized as ’Very Favorable’, and an accuracy of 65% in Scenarios 2 and 4, categorized as ‘Unfavorable’. The implications of this research contribute to the advancement of AgroTICs among small producers, with the potential to enhance and automate the monitoring of their harvest production.
It has recently been shown that a typical implementation of Simulated Annealing in continuous domains is effectively the same as random search/Hill Climbing. This result is repeated for Dual Annealing that exists as a popular and highly used optimization method in SciPy. A simpler version of Hill Climbing is then developed, and a restart-based version with local search is able to achieve better overall results on the CEC2022 benchmark functions.
Current software engineering focuses on achieving higher quality and speed in development and generating value for the business. This article proposes combining scenario thinking from requirement analysis techniques, such as Use Case 2.0, with test design and automation using the approaches of "Behavior-Driven Development (BDD)" and "Acceptance Test-Driven Development (ATDD)." This approach involves designing automated functional tests supported by artificial intelligence before development, which provides benefits regarding the time and cost of test automation. Enjisst is a user-language test automation platform aimed at creating different options for functional testing, which improves the understanding of the problem and promotes more effective communication of requirements across the entire team. This article demonstrates how Enjisst implements the described approach, which, supported by artificial intelligence and other sophisticated aspects, improves development speed and quality by over 50%. In summary, applying the required quality concepts in current agile and DevOps methodologies is a practical way.
As software technology continues to advance, ensuring quality and reliability of software products remains a significant challenge. One of the critical aspects of this challenge is the early detection of defects in software modules. In this paper, we research and analyze the use of multicriteria analysis in the detection of defects in software. In this method, a combination of two multi-criteria approaches and steps was used to obtain a quantitative analysis regarding the question addressed. In addition, we demonstrate the possibility of using a simpler method to assist in the detection of defects. To this end, a dataset made available by the National Aeronautics and Space Administration (NASA) was used, in addition, the Analytic Hierarchy Process (AHP) and Elimination and Choice Translating Reality (ELECTRE II) methods were used to classify the dataset source codes from the most complex to the least complex one. Results of the multicriteria analysis show that more complex codes are 70% more likely to have some type of defect, while the less complex ones are 50% less likely to have an error. This approach underscores the significance of employing multicriteria decision-making methods in addressing intricate problems, a theme resonant with current trends in computational intelligence research and its applications.
This paper presents a control approach for turbochargers aimed at maintaining a constant rotational speed and understanding its dynamic characteristics within the operating regime. An important aspect to consider when identifying the engineering characteristics of a turbocharger is the disturbance in the key operating conditions. Variations in the temperature and oil pressure ranges of the turbocharger are examined. A ratio control system is designed for the combustion chamber to maintain stable physical conditions in the turbocharger. This system ensures a constant air-fuel ratio aligned with the stoichiometric combustion requirements, while adjusting the characteristics of the thermal process to generate the energy necessary to operate the turbine. The mathematical model is developed and implemented in Matlab/Simulink (R). The most relevant fault-tolerant control strategies are designed and selected for implementation in the combustion chamber. This paper also presents a neuro-fuzzy internal model controller for maintaining constant turbocharger speed, which is crucial for efficient operation and fault diagnosis.
Proportional, integral, and derivative (PID) controllers have been widely adopted for industrial applications. However, these controllers are not very efficient for non-linear systems. Artificial neural networks (ANN) based on the Multilayer Perceptron (MLP) have great potential to replace PID controllers due to their polynomial structure, allowing complex non-linear systems to be controlled. This article introduces the integration of four MLPs as alternatives to a traditional PID controller. These MLPs were trained through four bioinspired algorithms tailored for following tasks in mobile robots. The bioinspired algorithms employed for MLP network training include Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), Month-Flame Optimization (MFO), and Artificial Hummingbird Algorithm (AHA). A comparative analysis was conducted between these MLPs and a classic PID controller, focusing on parameters such as overshoot (OS), settling time (ST), and steady-state error for different simulated scenarios.
Many studies highlight the importance of automating some steps of volcano monitoring. In particular, the classification of the volcano seismic events is a basic but fundamental task for observatories. The literature shows that artificial neural networks (ANN) are currently the most widely used techniques to develop classifiers for seismic signals from volcanoes. When training neural models, in addition to the quality of the dataset, the representation of the seismic events that feed the ANN is crucial for obtaining good performance. In general the representation is carried out in the time, frequency and both (spectrograms) domains. Spectrograms often need to be normalised for a better comparison between classes, resulting in a loss of the information that time variable typically carries. In previous works we have incorporated the events’ duration at the input or in intermediate layers of the ANN, without significant improvements in classification. In this paper we present a method that incorporates the time variable by adding the signal trace as an image to the spectrogram. Additionally, we also tested adding the spectrum module. Class activation maps (CAM) were applied over the different inputs to observe the sensitive areas that contribute to the neural model’s decisions and determine if the newly added representations are relevant for discriminating between classes. The results show that, for some classes, this additional information is indeed relevant for discriminating between seismic events and CAM brings explainability to the models.
Cold storage facilities play a crucial role in preserving perishable products across various industries, including food, pharmaceuticals, and agriculture. The high energy consumption of refrigeration systems in these facilities necessitates efficient energy management. This article uses Long Short-Term Memory (LSTM) models in a novel way to predict temperature and power demand variations in cold storage systems, leading to more accurate forecasts. The research improves demand response applications and achieves significant energy savings through real-world deployment by combining real-time IoT sensor data and deep learning. The study provides scalable framework that can be applied to a variety of industrial settings, helping to promote more sustainable energy management practices by optimizing power usage and lowering operational costs. Various methodologies and models were evaluated to predict and manage temperature variations and energy demand efficiently. The results demonstrate significant enhancements in energy efficiency, reduced energy consumption, and optimized resource usage. Among the evaluated machine learning techniques, deep neural networks, specifically Long Short-Term Memory (LSTM) models, showed superior accuracy and responsiveness. Mini-ROCKET and DrCIF, along with their precision and lower mean absolute error (MAE), make reliable predictions possible with a MAPE below 0.2 having overall accuracy of this model is 96%. Additionally, the adaptability of these models to abrupt operational changes was analyzed, highlighting their advantages and limitations. The paper concludes with recommendations for future research and development, promoting the adoption of advanced and sustainable technologies in energy management for cold storage systems.
Representing natural language information is a key challenge in Artificial Intelligence and Cognitive Science, requiring the transformation of unstructured data into formats suitable for computational tasks. While logical formalisms offer robust methods for information representation, their complexity often limits widespread adoption. Conversely, transformer architectures provide strong generalization capabilities but struggle with logical inference tasks. To address both the need for generalization and reliable logical inference, we propose a novel approach using deep reinforcement learning, enabling agents to autonomously learn the rules of semantic parsing. Our preliminary results indicate successful generation of appropriate representations for simple queries. Future work will extend the environment to handle a wider range of real-world sentences.
Nowadays the amount of people that is involved in some degree of hearing loss is increasing, and sign languages are useful as a rehabilitation method when there is a complete loss of hearing. However, countries like Colombia where there is not enough infrastructure and interpreters, the teaching of a sign language is affected. On the other hand, start with the learning of basic vocabulary of a sign language is a key factor for building the ability of making sentences. The first signs that are usually learned are: alphabet, colors, numbers, family members, pronouns, and greetings. This is where Computer Vision have brought tools for the translation of various signs. This allows to the development of solutions that reduce the frequent need of interpreters when learning basic vocabulary. This work proposes the creation of a word-based dataset for the signs of colors in the Colombian Sign Language. It was evaluated with three different Neural Network architectures: 1) LSTM, 2) GRU and 3) Conv_2D. Where the input for the network were the landmark coordinates extracted with MediaPipe. On the other hand, the Neural Network was evaluated with a dataset split of 80 % for training and 20 % for testing. Finally, it was seen a higher performance for the architecture based on LSTM, having a precision on detection of 92.6 %. The results, show the importance of Machine Learning tools that consider time-dependent data for developing solutions that support self-teaching on dynamic signs.
In Internet of Things (IoT) applications, data flows are continuous streams of high-dimensional time series that aggregate various data sources. In this context, decision-making processes frequently encompass multiple factors and criteria that demand forecasting these time series. This paper introduces MO-WMVFTS, a novel multiple-input multiple-output (MIMO) fuzzy time series (FTS) method for tackling this complex scenario. MO-WMVFTS is a hybrid forecasting method that fuses weighted multivariate FTS (WMVFTS) with embedding transformations, designed for IoT applications. To assess the performance of the proposed method, it was applied to the prediction of energy consumption in smart homes and air quality in smart cities. Here, three real-world datasets are used to assess the validity of our proposed approach, and the results obtained are highly competitive when compared to other existing methods.
This work presents the analysis of machine learning models applied to predict industrial process variables. More specifically, different models are explored, including Linear Regression, Nonlinear Regression, and Linear Regression with XGBoost. The research highlights the relevance of these models in the context of system identification, providing a comprehensive overview of their characteristics, benefits, and limitations. The work emphasizes the importance of choosing the appropriate model based on the characteristics of the problem at hand and the available information. To validate the effectiveness of these models, detailed comparisons of their performance and results are presented. This comparative analysis allows for a better understanding of the advantages and disadvantages of each approach and, thus, making informed decisions when applying machine learning to system identification.
This paper provides a comprehensive comparative analysis of the performance of four bioinspired algorithms (Differential Evolution, Grasshopper Optimization Algorithm, Moth-flame Optimization, and Particle Swarm Optimization) in designing a flatness-based controller for a ball and beam system simulated in MATLAB/Simulink environment. The assessment involves the tuning of feedback gains utilizing four distinct fitness functions: Integral Square Error, Integral Absolute Error, Integral Time Absolute Error, and an ad-hoc cost function. Hypothesis tests are employed to choose the optimal controllers from each combination for every bioinspired algorithm and a fitness function. Subsequently, a control performance analysis of these controllers for the ball and beam system is conducted, focusing on critical parameters such as overshoot, settling time, and steady-state error. This analysis demonstrates that all the controllers achieve a zero steady-state error. Noteworthy is the flatness-based controller tuned with the bioinspired Grasshopper Optimization Algorithm and utilizing the Integral Time Absolute Error fitness function, which stands out for achieving the fastest response with the lowest overshoot. This particular controller emerges as the most recommended choice for the specified application.
Since cohabitation with robots is transitioning from fiction to reality, ensuring their safe and efficient adoption is essential. However, a robot learning a new task needs to explore the environment, meaning that in certain situations potentially new unsafe actions might be attempted by a robotic agent. Therefore, it is imperative that they can anticipate latent dangers in object usage. A plausible approach for it is the use of object affordances. Affordances are the possibilities for action that an object or environment provides to a person or animal. This paper proposes a convolutional neural network-based affordance model to mitigate the risks posed by robotic agents when interacting with objects. The method uses geometric and 3D features to identify potential hazards such as cuts or impacts. The outcomes encompass the design of an architecture and focus on enhancing the safety of both the environment and those interacting with the robotic agent. The results obtained show that the model is able to identify efficiently the risk related to each object and even recognize the kind of hazards from unknown objects.
The recent rapid advancements in Large Language Models (LLMs) and the increasing availability of AI-powered tools have underscored the need for the current generation of programmers to learn how to effectively collaborate with AI from the early stages of their university education. This paper explores the integration of AI tools into a first-year programming course through the implementation of modified classic games (4 in a row with L-shapes, 3-player Battleship, etc.). The primary objective of this study was to assess the impact of AI assistance on students’ ability to define and adapt requirements for novel software applications, while also fostering an understanding of the power and limitations of AI in the classroom. The results reveal a positive student experience, with participants reporting increased confidence in utilizing AI tools for requirements elicitation and recognizing the potential benefits for their future careers. In addition, it highlighted the need to train students in developing skills for requirement identification, prompt creation, and testing and debugging the code.
Achieving high accuracy in time series forecasting, especially with missing data, is crucial. This paper proposes a novel neural networks approach, in which we use multiple time series datasets, potentially pre-processed using techniques like ARIMA, to capture temporal dependencies. Neural networks are then employed for two key functionalities: (1) data fusion to capture complex relationships between datasets, and (2) robust missing data handling through model training. This method addresses the challenges of incomplete wind farm data and aims to significantly improve forecasting accuracy. It is applied on several wind power datasets for experiments, the results show effectiveness of the proposed model.
This paper explores the integration of Long Short-Term Memory (LSTM) neural networks and Isolation Forest algorithms to predict failures in three-phase electric motors at a cement plant, specifically targeting a hot gas exhaust fan used in the clinkerization process. The LSTM model is employed to capture and analyze complex temporal patterns in vibration, speed, and power data, projecting future values with high accuracy. The Isolation Forest algorithm is then used to detect anomalies in these projections by identifying significant deviations from normal behavior. This combined approach allows for continuous monitoring of equipment, enabling timely interventions before failures occur. The implementation of these machine learning techniques demonstrates significant improvements in predictive accuracy, enabling timely interventions that can prevent critical failures and minimize downtime.
Deep learning models are at the forefront of solving classification tasks by recognizing intricate patterns in data using their non-linear functions. However, most data fed to these models exhibit high dimensionality, leading to substantial computational costs. Fortunately, not all information provided to the model is equally relevant for classifying data samples. By identifying and utilizing only essential information, we can significantly reduce computational costs while maintaining classification performance—an essential consideration for deploying models on resource-constrained edge devices.In this work, we propose a methodology for identifying the most relevant information within the data and feeding only that information to deep learning models. Specifically, we compute the Kullback-Leibler divergence on samples from the MIT-BIH arrhythmia classification task. By presenting the model with a fraction of the scalograms extracted using the Continuous Wavelet Transform (CWT), we achieve the following results: a 79% reduction in latency, 74% fewer Floating Point Operations (FLOPs), and a 44% reduction in model size, all while sacrificing only 4.57% accuracy.
