machine learning

Bearings are widely used parts in the industry and play a crucial role in the performance of industrial systems and machinery. Therefore, their failure can cause significant damage and even halt production. Due to their importance and widespread use, various methods have been developed for troubleshooting and estimating their remaining useful life. This study introduces a novel approach for predicting the remaining lifespan of industrial bearings. The technique involves initially decomposing the vibration signals from the bearing into intrinsic mode functions (IMFs) using the empirical mode decomposition (EMD) method. Among the calculated IMFs, the one most effective at indicating the degradation rate of the ball bearing is chosen for training the neural network. The selection of the IMF is carried out using a proposed energy-based method. The variance statistical parameter for the selected IMF is calculated after selecting the appropriate Intrinsic Mode Function. Next, the Weibull function is fitted to the resulting data, and the obtained graphs are used to train the neural network. Subsequently, the results are smoothed using the exponential smoothing function. The final performance of the proposed algorithm is evaluated using experimental data. The results indicate that the neural network can effectively track the degradation process of the bearing and assess its remaining lifespan.

The limitation of fossil energy sources and environmental concerns have caused efforts to use clean and sustainable energy sources. The solid oxide fuel cell with intermediate working temperature and ammonia fuel is one of the promising sources to replace conventional energy sources. On the other hand, the use of fuel cell performance prediction methods with appropriate and high accuracy and speed is significantly important. In this research, first, the solid oxide fuel cell with ammonia fuel, electrolyte leakage, and average working temperature are numerically simulated. Then the effective input parameters are selected to calculate the performance of the fuel cell in different conditions. In this regard, after generating a sufficient data set, different machine learning algorithms are used to predict the objective functions, including the power density and the maximum temperature of the fuel cell. The results indicate the complexity of predicting the power density of the fuel cell compared to the maximum temperature. It was also observed that the XG Boosting method with R2 equal to 0.99 has the best efficiency in predicting the parameters of maximum temperature and power density.

This study evaluates the efficiency and productivity of 10 active units in the field of high technology using Data Envelopment Analysis (DEA), Analytic Hierarchy Process (AHP), and machine learning techniques. Input criteria include research investment, number of specialized human resources, and operational costs, while output criteria consist of the number of published articles, registered patents, and commercialized products. AHP was used to calculate the relative weights of the criteria, followed by the application of the CCR and BCC models to assess the relative efficiency of the units. To enhance prediction accuracy, machine learning algorithms, including Artificial Neural Networks (ANN) and Decision Trees, were employed. The results showed that six units in the CCR model and eight units in the BCC model were identified as efficient. The ANN model, with 95% accuracy, predicted unit efficiency more precisely. The analysis revealed that Total revenue and number of published articles had the greatest impact on efficiency. This research presents a combined approach for evaluating and improving performance in high technology, offering valuable tools for managers and policymakers to optimize resource allocation and make strategic decisions.

Combustion instability induced by acoustic waves can lead to various undesired consequences, including thermal stress on the combustion chamber, noise, flame blow-off, flashback, vibrations, and even explosions. This paper employs the Design of Experiments approach to systematically gather reliable experimental data for predicting flame extinction. Considering the substantial cost of pure reaction gases and the potential damage to the acoustic driver under high-pressure conditions, it is imperative to intelligently select extinction test points. Machine learning methods are employed to determine optimal acoustic power levels for these values. The acoustic power required at the moment of extinction is a crucial metric in understanding this phenomenon. Four key features - frequency, equivalence ratio, wall diameter ratio, and Reynolds number - serve as inputs for a machine learning (ML) model. The collected data is utilized to train a selected supervised ML model, specifically the Support Vector Regression (SVR), to accurately predict the acoustic power level required for flame extinction in both methane and propane fuels. Evaluation using the R-squared metric demonstrates the model's accuracy and robust performance across diverse conditions.

Structural health monitoring (SHM) systems are used to ensure the safe operation of infrastructure and making decisions about the structure's maintenance, repair, and retrofitting. This paper presents a new method based on signal processing and machine learning for SHM. The proposed method is based on deep temporal and time-frequency features. The time-frequency representation (TFR) is obtained using continuous time Fourier transform (CWT), then TRF is applied to the convolutional neural network (CNN) to extract deep time-frequency features. To extract the deep temporal features, the intrinsic mode functions (IMF) are obtained using the empirical mode decomposition (EMD), and the long short-term memory (LSTM) network models the IMFs. Due to the large number of deep features, the features with high correlation were removed using the feature reduction algorithm. Finally, using the optimized support vector machine (SVM), the structure's health or the damage's location is identified. The results show that the proposed method is more accurate than similar methods in linear and non-linear scenarios and can be used as a reliable method in SHM applications.

In this study, a comprehensive analysis of driving behavior with an emphasis on fuel consumption and driver categorization is presented. Data from 80 drivers were collected using a custom-designed datalogger connected to the vehicle’s On-Board Diagnostics (OBD) port. Critical features related to driving patterns were extracted through a correlation matrix and concepts in the field of powertrains. Key variables such as acceleration and deceleration were identified and derived. Regression models were applied to predict fuel consumption based on this driving feature. Through this analysis, the most influential factors affecting fuel efficiency were highlighted. Additionally, unsupervised machine learning techniques were employed to cluster drivers into distinct groups based on their driving styles. A comparative study of various algorithms was conducted to evaluate the efficacy of different clustering methods. Valuable insights for automotive manufacturers, policymakers, and drivers are offered by the results, emphasizing the role of driving behavior in fuel efficiency and the potential for tailored driver assistance systems.