Wyniki wyszukiwania - BazTech

1

Decoding human movement intentions from EEG signals using recurrent neural networks

Kleczyński Daniel, Kluska Jacek

Zeszyty Naukowe Politechniki Rzeszowskiej. Elektrotechnika

|

2025

|

Vol. 31

43--53

EN

The present study develops a model for recognizing movement intentions from Electroencephalography (EEG) signals using various Recurrent Neural Network (RNN) configurations, with a focus on Long Short-Term Memory (LSTM) networks. Experiments demonstrate that, with proper sample allocation and class balance, the model achieves an average accuracy of 0.9815±0.0025 and an average Receiver Operating Characteristic area (AUC) of 0.9989±0.0004 when training and test data include the same subjects. The best-performing LSTM model - augmented with a fully connected layer - was configured with a hidden layer size of 233, learning rate of 3.872 × 10−4, 3 layers, dropout of 0.3773, and sequence length of 457. However, when test subjects were completely excluded from training, the model’s accuracy did not exceed 50%, suggesting significant inter-subject variability or limitations in generalization. This work contributes to advancing Brain-Computer Interfaces for applications such as prosthetic control and provides insights into the prerequisites for effective EEG signal utilization.

2

Study of AQI Monitoring System of Indoor Environment Using Machine Learning Model and IoT Device

Talati Isha, Shah Kunjan, Patel Om, Tanna Jaymin, Jain Akshat, Oza Ankit D., Yadav Amruta Arun, Alshayeb Mohammed J., Khan Mohammad Amir, Islam Saiful

Rocznik Ochrona Środowiska

|

2025

|

Tom 27

152--163

EN

Indoor air quality has a direct impact on human health. Thus, it's essential to comprehend the various aspects of indoor air quality. It supports both the implementation of preventative measures and the monitoring of indoor air pollution. Monitoring and forecasting air pollution is extremely essential, especially in developing countries like India. This study proposes a system that employs ESP8266 (NodeMCU) data sent to the cloud to monitor the levels of air pollutants such as ozone, particle matter, carbon monoxide, carbon dioxide, temperature, and total volatile organic compounds. Our sensors include the ozone sensor MQ-131, the dust sensor GP2Y1010-AU0F, the TVOC sensor AGS02MA, the carbon monoxide sensor MQ-9, the carbon dioxide sensor MQ-135, and the humidity sensor DHT11. The IoT device continuously shows the indoor air quality level (IAQL). The next step was to accurately anticipate the Internal Air Quality Level (IAQL) and pollution levels from dangerous gases for the next seven days using the LSTM, Seasonal ARIMA, and Linear Regression models. The Authors could accurately predict the observations of the following seven days after using data from the previous ninety days to create our best model. This implies that our model can accurately predict the values for each parameter with an accuracy of at least 95%. Therefore, we believe such a solution would be advantageous if a large-scale installation were implemented. If consumers can remotely verify the air quality in their homes, the pollution in the interior atmosphere will decrease. This has the potential to make civilization healthier.

3

Methods of classifying voltage surges using Deep Neural Networks

Krawczyk-Borysiak Zuzanna, Kluge Paweł, Łasica Andrzej, Sul Przemysław, Ciuba Maciej

Przegląd Elektrotechniczny

|

2025

|

R. 101, nr 3

211--214

EN

The paper focuses on exploring the potential application of neural networks for the classification of voltage surges compliance with the norm. Three potential neural network architectures were considered for the task - a convolutional neural network (further referred to as CNN), a model combining convolutional and LSTM layers (CNN+LSTM) and a transformer model. The best results were achieved by the simple transformer model (accuracy of 93% on the test dataset), followed by CNN+LSTM model (accuracy: 81%), and CNN (accuracy: 69%).

PL

Artykuł koncentruje się na badaniu potencjalnego zastosowania sieci neuronowych do klasyfikacji zgodności udarów napięciowych z normą. Do tego zadania rozważono trzy potencjalne architektury sieci neuronowych - konwolucyjną sieć neuronową (CNN), model łączący warstwy konwolucyjne i LSTM (CNN+LSTM) oraz model transformatora. Najlepsze wyniki uzyskał prosty model transformatora (dokładność 93% w zestawie danych testowych), następnie model CNN+LSTM (dokładność: 81%) i CNN (dokładność: 69%).

4

Data-driven prediction of asphalt pavement temperature profiles

Sudyka Jacek, Mazurek Grzegorz, Buczyński Przemysław

Materiały Budowlane

|

2025

|

nr 12

142--156

EN

Reliable knowledge of the temperature distribution within asphalt pavements is essential for maintenance and structural diagnosis. To forcast the asphalt layers temperature the recurrent neural networks (RNN, including LSTM and BiLSTM) and gradient-boosted decision trees (XGBoost) have been used based on a multi-month field dataset (March – October) with multi-depth temperature measurements and meteorological variables. RNNs captured both diurnal fluctuations and seasonal trends with high predictive accuracy. While the classical XGBoost setup was slightly less precise, it offered very short training times and greater interpretability; its depth-generalized experimental variant enabled interpolation across the full depth range with an error of ~0,97°C (R² ≈ 0,988). The findings support hybridization (RNN + XGBoost) to combine temporal-pattern extraction with efficient regression on static features (e.g., depth, time-of-day).

PL

Utrzymanie i diagnostyka nawierzchni asfaltowych wymagają wiarygodnej informacji o rozkładzie temperatury w czasie i w głąb konstrukcji. Do prognozowania temperatury warstw asfaltowych zastosowano rekurencyjne sieci neuronowe (RNN, w tym LSTM i BiLSTM) oraz gradientowe modele drzew decyzyjnych (XGBoost) na podstawie wielomiesięcznych danych terenowych (marzec – październik) obejmujących pomiary na wielu głębokościach oraz parametry meteorologiczne. Modele RNN wiernie odwzorowały zarówno wahania dobowe, jak i sezonowe. XGBoost, choć w wariancie klasycznym nieco mniej precyzyjny, zapewnił bardzo krótki czas obliczeń i większą interpretowalność; jego wariant eksperymentalny z uogólnieniem po głębokości umożliwił interpolację temperatury w całym zakresie badanych głębokości z błędem rzędu ~0,97°C (R² ≈ 0,988). Wyniki wskazują na zasadność hybrydyzacji podejść (RNN + XGBoost), łączącej identyfikację wzorców czasowych z efektywną regresją po cechach statycznych (m.in. głębokość, pora doby).

5

Application of neural networks for drought forecasting based on the standardised precipitation index

Fathiyana Rana Z., Haryadi Deny, Ananda Naufal, Hidayat Dinda J., Wicaksana Haryas Subyantara, Azahra Rahma Wafii, Saputra Hakim Giraldi

Journal of Water and Land Development

|

2025

|

no. 65

113--121

EN

Based on data from the National Disaster Management Agency, South Sumatra is one of the provinces with a reasonably large drought-affected area, totalling 8,853,691.009 ha. Drought is a hydrometeorological disaster, characterised by anomalous rainfall below normal levels. Reduced rainfall can lead to decreased soil moisture, reduced river flows, and a general scarcity of water, which limits availability of water both on the surface and in the soil. To anticipate and mitigate the impacts of drought, an accurate forecasting system is essential for effective disaster management and mitigation. This research focuses on forecasting drought using the standardised precipitation index (SPI) based on Long Short-Term Memory (LSTM) and Multilayer Perceptron (MLP) algorithms. It compares LSTM and MLP algorithms by integrating rainfall data from the FY-4A satellite and observational rain gauges, which are processed to generate SPI values. These data are employed to train and test MLP and LSTM models in predicting future drought conditions. The results indicate that drought can be effectively predicted using both MLP and LSTM. However, the MLP outperforms the LSTM, as reflected by a higher Nash-Sutcliffe efficiency (NSE) value, a lower error rate, and a predicted date trend that more closely aligns with actual observations.

6

ILP Optimized LSTM-based Autoscaling and Scheduling of Containers in Edge-cloud Environment

Singh Shivan, Narayan D.G., Mujawar Sadaf, Hanchinamani G.S., Hiremath P.S.

Journal of Telecommunications and Information Technology

|

2025

|

nr 2

56--68

EN

Edge computing is a decentralized computing paradigm that brings computation and data storage closer to data sources, enabling faster processing and reduced latency. This approach is critical for real-time applications, but it introduces significant challenges in managing resources efficiently in edgecloud environments. Issues such as increased response times, inefficient autoscaling, and suboptimal task scheduling arise due to the dynamic and resource-constrained nature of edge nodes. Kubernetes, a widely used container orchestration platform, provides basic autoscaling and scheduling mechanisms, but its default configurations often fail to meet the stringent performance requirements of edge environments, especially in lightweight implementations like KubeEdge. This work presents an ILP-optimized, LSTM-based approach for autoscaling and scheduling in edge–cloud environments. The LSTM model forecasts resource demands using both real-time and historical data, enabling proactive resource allocation, while the integer linear programming (ILP) framework optimally assigns workloads and scales containers to meet predicted demands. By jointly addressing auto-scaling and scheduling challenges, the proposed method improves response time and resource utilization. The experimental setup is built on a KubeEdge testbed deployed across 11 nodes (1 cloud node and 10 edge nodes). Experimental results show that the ILP-enhanced framework achieves a 12.34% reduction in response time and a 7.85% increase in throughput compared to the LSTM-only approach.

7

Forecasting of short-term power demands in Polish Power System using ensemble of LSTM networks

Ciechulski Tomasz

Journal of Automation, Electronics and Electrical Engineering

|

2025

|

Vol. 7, No. 1

29--37

EN

The article presents and discusses the results of the research of forecasting power demands in Polish Power System with time horizon of one hour ahead in conditions of limited availability of forecasting model input data, covering only three months. The prediction was carried out using deep neural networks - LSTM (Long Short-Term Memory) connected to an ensemble. The performance of the ensemble is much more efficient than individual networks working separately. The numerical experiments were conducted using MATLAB computing environment. The accuracy of the predictions was estimated using such statistical measures as MAPE, MAE, RMSE, Pearson correlation coefficient R.

8

Intelligent Electricity Load Forecasting Method using ARIMA-LSTM-Random Forest

Aziukovskyi Oleksandr, Hnatushenko Volodymyr, Kashtan Vita, Polyanska Alla, Jamróz Andrzej

Inżynieria Mineralna

|

2025

|

Vol. 1, no. 1

123--132

EN

The instability of energy systems caused by internal economic factors and external challenges, including geopolitical conflicts, significantly complicates the process of planning and managing energy resources. An essential tool for implementing energy-saving measures is introducing modern computer technologies, including artificial intelligence systems, in the energy sector. Intelligent technologies make it possible to use methods for predicting electrical load, including artificial intelligence algorithms. This paper proposes a combined ARIMA-LSTM-Random Forest model for forecasting electric load. The combination of the approaches allows considering both linear and nonlinear dependencies in the data, which is critical to ensure the accuracy of forecasts. Using data for the previous seven days provides enough information to identify seasonal trends and fluctuations, which makes this a promising prospect for medium-term forecasting in energy monitoring tasks. Thus, combining the ARIMA, LSTM, and Random Forest methods achieves high accuracy in forecasting electricity consumption. The proposed approach is an optimal solution since it combines the advantages of each model and compensates for their shortcomings. The proposed ARIMA-LSTM-Random Forest method significantly improved the results: MSE = 0.27, RMSE = 0.23, MAPE = 0.35%. The method minimized absolute and relative errors, confirming its advantage for this forecasting task. The results are promising for practical application in the load management of electric networks.

PL

Niestabilność systemów energetycznych, spowodowana wewnętrznymi czynnikami ekonomicznymi i wyzwaniami zewnętrznymi, w tym konfliktami geopolitycznymi, znacząco komplikuje proces planowania i zarządzania zasobami energetycznymi. Niezbędnym narzędziem wdrażania działań na rzecz oszczędności energii jest wprowadzenie do sektora energetycznego nowoczesnych technologii komputerowych, w tym systemów sztucznej inteligencji. Technologie inteligentne umożliwiają wykorzystanie metod prognozowania obciążenia elektrycznego, w tym algorytmów sztucznej inteligencji. W niniejszym artykule zaproponowano połączony model ARIMA-LSTM-Random Forest do prognozowania obciążenia elektrycznego. Połączenie tych podejść pozwala na uwzględnienie zarównoliniowych, jak i nieliniowych zależności w danych, co jest kluczowe dla zapewnienia dokładności prognoz. Wykorzystanie danychz ostatnich siedmiu dni dostarcza wystarczających informacji do identyfikacji trendów i wahań sezonowych, co czyni to obiecującą perspektywą dla prognozowania średnioterminowego w zadaniach monitorowania energii. Zatem połączenie metod ARIMA,LSTM i Random Forest pozwala osiągnąć wysoką dokładność prognozowania zużycia energii elektrycznej. Proponowane podejściejest rozwiązaniem optymalnym, ponieważ łączy zalety każdego modelu i kompensuje ich wady. Zaproponowana metoda ARIMA--LSTM-Random Forest znacząco poprawiła wyniki: MSE = 0,27, RMSE = 0,23, MAPE = 0,35%. Metoda zminimalizowała błędybezwzględne i względne, co potwierdza jej przewagę w tym zadaniu prognostycznym. Wyniki są obiecujące pod kątem praktycznychzastosowań w zarządzaniu obciążeniem sieci elektroenergetycznych.

9

Artificial Intelligence Application for Wind Power Forecasting for Wind Turbine Generators: A Case Study in Vietnam

Pham Huu Chien, Ngo Xuan Cuong, Do Nhu Y.

Inżynieria Mineralna

|

2025

|

Vol. 1, no. 2

7--16

EN

Accurate short-term wind power forecasting plays a critical role in enhancing the reliability and efficiency of wind farm operations, especially in regions with high wind variability such as coastal Vietnam. This study investigates the application of three advanced artificial intelligence (AI) Long shortterm memory (LSTM), Gated recurrent unit (GRU), and Xtreme gradient boosting (XGBoost) for wind power prediction using real-world SCADA data collected from the WT01 turbine -4MW at the wind farm Ninh Thuan over a 360-day period. The forecasting performance of these models is evaluated under three feature scenarios: (i) wind speed only, (ii) wind speed combined with rotor speed, and (iii) a comprehensive set of five features including wind speed, rotor speed, pitch angle, vibration level, and internal temperature. Model performance is assessed using standard metrics such as RMSE, NMAPE, and training time. Results show that LSTM achieved the highest accuracy, with RMSE and NMAPE of 654.15 kW and 12.82%, respectively, when trained with all five features. GRU delivered comparable results with shorter training time, while XGBoost exhibited superior computational efficiency but slightly lower accuracy. These findings highlight the importance of feature richness in enhancing prediction accuracy and suggest that the choice of forecasting model should consider both accuracy requirements and computational constraints. The study offers practical insights for AI-based forecasting system design tailored to the operational needs of wind power facilities in Vietnam and similar regions.

10

Deep Learning in motion analysis for false start detection in speedway racing

Krakowian Jacek, Jeleń Łukasz

International Journal of Electronics and Telecommunications

|

2025

|

Vol. 71, No. 3

art. no. 9, s. 1-8

EN

Accurately identifying false starts in speedway racing is a very challenging task due to the subtle nature of pre-start movements. Manual detection methods, often dependent on the judgment of race officials, are prone to errors and subjectivity, leading to inconsistencies in decision-making. This paper introduces an automated approach that leverages computer vision methods to enhance detection precision. Here, we have expanded its use to detect false starts in speedway racing. The proposed approach introduces image processing techniques with 3D Convolutional Neural Networks (CNNs) and Long-Short-Term Memory (LSTM) networks to analyze rider movements during the starting procedure. Unlike manual detection, which often misses fine movements at the start line, our method uses 3D CNNs to monitor racer movements and applies LSTM networks to assess time-based motion patterns that signal false starts. The presented results show that the 3D CNN achieved an accuracy of 86.36% with a higher precision when compared to traditional methods. This automated process not only enhances fairness in competitive racing, but also illustrates the broader capability of emerging technologies to refine decision-making in sports.

11

Leveraging Artificial Intelligence for Cyanobacterial Bloom Prediction: A Hybrid Deep Learning and Generative Adversarial Network Framework for Accurate Forecasting and Proactive Management

Dendani Nadjette, Saoudi Amel, Amara Nour Djihane, Azizi Nabiha, Dugdale Julie, Hadiby Rayenne

Foundations of Computing and Decision Sciences

|

2025

|

Vol. 50, No. 4

427--450

EN

This study presents an Artificial Intelligence-based system designed to predict cyanobacterial harmful algal blooms (CyanoHABs). The system utilizes Long Short-Term Memory (LSTM) networks to predict the timing of bloom occurrences and One-Dimensional Convolutional Neural Networks (1D-CNNs) to estimate cyanobacterial density. Additionally, Generative Adversarial Networks (GANs) are employed for data augmentation to enrich the database. The system's performance was validated using the Algerian Mexa database, achieving an R-squared (R²) value of 98% and a root mean square error (RMSE) of 9% for cyanobacterial density prediction, and an R-squared value of 88% with a root mean square error of 31% for bloom timing prediction. These results highlight the system's robust predictive capabilities, enabling proactive monitoring and management of CyanoHABs to mitigate their adverse impacts on health and the environment.

12

The statistical-based diagnosis with usage of acoustic sound decomposition and projected LSTM network of induction motors

Zastępa Marek

Eksploatacja i Niezawodność

|

2025

|

Vol. 27, no. 3

art. no. 205651

EN

The use of acoustic signals in the diagnosis of electrical machines allows for non-invasive and rapid diagnostics. The author proposed the novel approach of acoustic diagnosis of single-phase induction motors, which is 98.67% accurate on the test set and allows for fault detection in circa 0.042 s, and 97.33% accurate for 0.021 s long samples similarly. The research includes five classes of faults. In this method, intrinsic mode functions (IMFs) gained from the empirical mode decomposition (EMD) of the motor sound are used to calculate the following statistical parameters: mean, mean square, root mean square, standard deviation, energy, and norm. Next, these parameters are organized from a prepared matrix to a vector of parameters one IMF by one, suitable for neural network input. Such prepared data is then passed to the proposed architecture of the projected LSTM neural network. The training processes were fast - they took only 12 and 13 seconds selectively. The presented novel method is useful for acoustic fault diagnosis of electric motors and could be used for other motors.

13

Forecasting the required quantity of cement manufacturing materials using time series and Q-Network techniques

Madjour Hassina, Zermane Hanane, Benaicha Sonia

Ecological Chemistry and Engineering. S

|

2025

|

Vol. 32, nr 3

415--426

EN

In the era of Industry 4.0, accurate prediction of industrial process parameters is essential for optimising operations, lowering costs, and enhancing product quality. Traditional statistical methods often struggle to capture the complex temporal dependencies within industrial processes. This study explores the use of Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Q-Network models to predict material quantities in an industrial dataset. The dataset was pre-processed to address missing values and outliers, and the models were evaluated based on Mean Squared Error (MSE), R2, and accuracy. The results show that the LSTM model achieved an MSE of 14.253 and an R2 of 0.700. The BiLSTM model greatly outperformed it, with an MSE of 0.714 and an R2 of 0.985. The Q-Network model produced an MSE of 0.005 and an R2 of 0.992. These findings demonstrate the Q-Network’s superior ability to capture temporal dependencies within the data.

14

Data mining and analysis of flight control system based on temporal features and improved LSTM algorithm

He Dingjun, Fan Liping, Xu Lijun

Diagnostyka

|

2025

|

Vol. 26, No. 4

art. no. 2025408

EN

Due to the structural characteristics of multi-redundancy and multi-closed loops in flight control systems, their fault propagation modes are complex, and the internal physical structure is closely coupled with system components, which poses challenges for analysis and modeling. To improve the accuracy and predictive ability of flight control system fault diagnosis, this study proposes a flight control system fault diagnosis method built on an improved bidirectional long short-term memory network. By integrating convolutional neural networks and bidirectional long short-term memory networks to extract local and temporal features of the data space, the classification and regression problems of flight control system state prediction have been solved. The results indicated that the proposed fault diagnosis algorithm had the highest recognition accuracy for the four modes. Compared with single convolutional neural networks and long short-term memory networks, the accuracy has increased by 2.11% and 1.32%, and the fault diagnosis accuracy has reached 99.49%, which could accurately identify various types of faults. The improved network proposed this time significantly improves the accuracy of flight control system fault diagnosis and reduces false alarm and missed alarm rates.

15

Construction of intelligent transformer anomaly detection model based on LSTM combined with optical flow features

Zhao Shuzong, Pattanadech Norasage

Diagnostyka

|

2025

|

Vol. 26, No. 1

art. no. 2025115

EN

Transformers are essential for the transmission and distribution of electricity, but due to changes in load and the influence of the working environment, various faults may occur in transformers. To accurately and quickly detect faults in transformers and conduct effective fault diagnosis and equipment maintenance, this study solves the problems of data imbalance and temporal data in transformers by introducing a long short-term memory network with fatigue factors. In addition, a fusion model is ultimately constructed by combining the recursive all-pair field transformation streamer method to achieve more accurate and robust optical flow estimation in the model. The experiment indicated that the maximum accuracy of the predicted values combined with the model was around 95% and the minimum was around 35%. Compared to other models, the maximum accuracy of actual values was around 80% and the minimum accuracy was better at 10%. In the application experiment, the frequency of insulation faults was the least obvious, with only 10 faults. The resistance fault was evident, with a total of 100 faults. The combined model could well reflect the fluctuation of fault current and the collection of fault numbers by different sensors. Therefore, the proposed model has high accuracy, good precision, and outstanding application effects, which can provide new ideas for constructing intelligent transformer anomaly detection models.

16

Preliminary study on artificial intelligence methods for cybersecurity threat detection in computer networks based on raw data packets

Ogonowski Aleksander, Żebrowski Michał, Ćwiek Arkadiusz, Jarosiewicz Tobiasz, Klimaszewski Konrad, Padee Adam, Wasiuk Piotr, Wójcik Michał

Computer Science

|

2025

|

T. 26 (SI)

45-–68

EN

Most of the intrusion detection methods in computer networks are based on traffic flow characteristics. However, this approach may not fully exploit the potential of deep learning algorithms to directly extract features and patterns from raw packets. Moreover, it impedes real-time monitoring due to the necessity of waiting for the processing pipeline to complete and introduces dependencies on additional software components. In this paper, we investigate deep learning methodologies capable of detecting attacks in real-time directly from raw packet data within network traffic. Our investigation utilizes the CICIDS-2017 dataset, which includes both benign traffic and prevalent real-world attacks, providing a comprehensive foundation for our research.

17

Intrusion detection with machine learning: a two-step federated approach.using the CIC IoT 2023 dataset

Jakotiya Komal, Shirsath Vishal, Inamdar SharanBasava

Computer Science

|

2025

|

T. 26 (2)

77--95

EN

The main objective of the planned effort is to provide analytical analyses of current intrusion detection systems grounded on ML algorithms. Furthermore, examined in this work are the useful data sets and several techniques already in use to develop an effective IDS using single, hybrid, and ensemble machine learning algorithms. The approaches in the literature have then been investi-gated under several criteria to provide a clear road and direction for the next projects that will be successful. Nowadays, companies of all kinds include an intrusion detection system (IDS), which inhibits cybercrime to protect the network, resources, and private data. Many strategies have been suggested and implemented up till now to prevent uncivil behaviour. Since machine learning (ML) approaches are successful, the proposed approach applied several ML models for the intrusion detection system. The CIC IoT 2023 Dataset is the one applied in this paper, and a two-step process for Intrusion detection was proposed. Tested with several techniques including random forest, XGBoost, logistic regression, MLP model, and RNN. Following fine-tuning, the federated learning model using neural networks had the best accuracy—99.84%.

18

Detection and forecasting of Parkinson disease progression from speech signal features using multilayer perceptron and LSTM

Ali Majid, Shakir Hina, Samreen Asia, Ahmed Sohaib

Computer Science

|

2025

|

T. 26 (2)

31--48

EN

Accurate diagnosis of Parkinson′s disease, especially in its early stages, can be a challenging task. The application of machine learning (ML) techniques has helped improve the diagnostic accuracy of Parkinson′s disease (PD) detection but integration of diagnostic features in ML models for the prediction of disease progression has remained an unexplored research avenue. In this research work, Long Short Term Memory (LSTM) was trained using diagnostic features on Parkinson patients speech signals, to predict the disease progression while a Multilayer Perceptron (MLP) was trained on the same diagnostic features to detect PD. Diagnostic features were selected using two well known feature selection methods named Relief F and Sequential Forward Selection method. The integration of feature selection methods in LSTM model has resulted in PD progression forecast with an accuracy of 88.7%. Further more, with the application of input diagnostic features on MLP, PD stage was accurately detected with an accuracy of 98.63%, precision of 97.64% and recall of 98.8% showing model robustness and efficiency for its potential application in health care.

19

Identification and classification of turn short-circuit and demagnetization failures in PMSM using LSTM and GRU methods

Lale Timur, Yüksek Gökhan

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2025

|

Vol. 73, nr 1

art. no. e151958

EN

In an extremely broad range of industrial applications, especially in electric vehicles, permanent magnet synchronous motors (PMSMs) play a vital role. Any failure in PMSMs may cause possible safety hazards, a drop in productivity, and expensive downtime. Therefore, their reliable operation is essential. Accurate failure identification and classification allow for addressing problems before they escalate, which helps ensure the seamless operation of PMSMs and reduces the likelihood of equipment failure. Therefore, in this paper, novel failure identification methods based on gated recurrent unit (GRU) and long short-term memory (LSTM) from recurrent neural network (RNN) methods are proposed for early identification of stator inter-turn short circuit failure (ISCF) and demagnetization failure (DF) occurring in PMSMs under multiple operating conditions. The proposed methods use three-phase current signals recorded from the experimental study under multiple operating conditions of the motor as input data. In the proposed methods, both feature extraction and classification are executed within a unified framework. The experimental outcomes obtained demonstrate that the proposed methods can identify a total of six unique motor conditions, including three ISCF variations and two DF variations, with high accuracy. The LSTM and GRU approaches predicted the identification of failures with 98.23% and 98.72% accuracy, respectively. Compared to existing methods, the success of the proposed approaches is satisfactory. In addition, LSTM and GRU-based failure identification methods are also compared in detail for accuracy, precision, sensitivity, specificity, and training time in this study.

20

Prediction of photovoltaic energy generation using recurrent and transformer neural networks

Siwek Krzysztof, Świder Stanisław

Archives of Electrical Engineering

|

2025

|

Vol. 74, nr 2

311--330

EN

Precise prediction of photovoltaic (PV) energy generation is essential for optimal, profitable and ecological management of electric energy resources all over the world. As a result, attempts are being made to develop more accurate prediction algorithms. This paper compares the application of Long Short-Term Memory (LSTM, a subtype of Recurrent Neural Networks), PatchTST (a type of Transformer Neural Network – TNN) and ensemble models (making use of these two approaches) for estimating PV energy production 24 hours ahead. The results indicate that both analysed single methods have comparable prediction accuracy, though the hybrid approach outperforms them. The experiments were conducted on data from PV sites deployed across campuses at Australian La Trobe University. However, future studies could verify this approach using different datasets. Algorithms and results presented in this study may especially contribute to the development of Recurrent and Transformer Neural Networks as prediction methods of PV energy production.