جستجوی مقالات مرتبط با کلیدواژه « artificial neural network » در نشریات گروه « مکانیک »

Increasing the depth of microchannels on metallic bipolar plates (BPPs) in PEM fuel cells leads to an increase in the efficiency. In this research, the stamping process has been applied for manufacturing of the BPPs made of commercially pure titanium with a direct parallel flow field. The effect of process parameters including die clearance, forming speed, and sheet/die friction coefficient on the filling rate and thinning of the BPPs was investigated. The required tests were designed via the response surface method (RSM), implemented by a validated finite elements (FE) model, and the desired outputs were extracted. Then, a quadratic equation was presented for predicting the filling rate based on the input parameters using the regression method. In the following, using the artificial bee colony algorithm, the coefficients of the mentioned equation were enhanced and its error was decreased almost by 53%. Finally, an artificial neural network (ANN) was used to predict the filling rate. The results demonstrated that the proposed ANN model is very effective and approximates the filling rate of the microchannel with high accuracy.

In this study, a 4×4 square network structure made of titanium has been optimized under tensile force using relevant parameters. The network structure is a quadrilateral structure with a side length of L and ϴ, and the fracture order of the walls has been compared using MATLAB software and simulation with Abaqus software, and the results of the fracture order of the structure match each other. In the present study, the objective function in optimization is to increase energy absorption and minimize the maximum stress, and the effect of parameters such as side lengths and various angles in this structure has been investigated. 100 different cases have been obtained for values of L and ϴ with output of area under the curve (energy absorbed) and maximum stress and strain using MATLAB software. With input data (L and ϴ) and output data (energy absorbed and maximum stress), a neural network has been trained and a regression model has been used in the neural network to achieve a prediction accuracy of over 99%, which is a high level of accuracy. The relationship function between input and output of the neural network has been obtained using MATLAB software, and the optimization of this 4×4 network structure has been carried out using the genetic algorithm. The objective function in this study is to increase energy absorption and minimize the maximum stress so that the network structure has the highest strength considering the examined parameters.

When working with hardened materials, it's important to control and optimize the surface roughness and machining force. To achieve this, we can use intelligent methods that are based on prediction and optimization models. In this study, an artificial neural network was used to evaluate the surface roughness and machining force of hardened steel 4140 by analyzing cutting speed, feed rate, and machining time. A full factorial method was used to carry out 27 experiments, and an uncoated cemented carbide tool TCMW 16T304 H13A was used to measure surface roughness and machining force during turning. An artificial neural network model with two hidden layers was selected as the optimal architecture for separately predicting surface roughness and machining force. The predicted values were then compared with the experimental results, and the average error percentage for validation data was calculated as 4.25% for surface roughness and 5.11% for machining force. Finally, the optimal cutting parameters were selected to minimize surface roughness and machining force.

The ability to predict tool wear during machining is a very important part of diagnosis, which makes it possible to replace the tool at the appropriate time. Therefore, in this research, the artificial neural network approach was used to predict tool wear. First, hardened steel 4140 was turned with uncoated cemented carbide tool TCMW 16T304 H13A and with input parameters including cutting speed, feed rate and machining time in three different levels and with constant cutting depth, and the amount of tool wear was measured. And the experimental test results were used to train and validate the artificial neural network. The optimal neural network architecture was obtained with 3 nodes in the input layer, two hidden layers with 12 and 36 nodes in the first and second hidden layers, and 1 node in the output layer to predict tool wear. The prediction values of the artificial neural network model were compared with the experimental results and the average error percentage of the validation data was calculated as 3.32%.

Heat transfer is an important and influential factor in the design of internal combustion engines. Accurate prediction of heat transfer has a great impact on the design of cooling systems of internal combustion engines. The simulation of internal combustion engines for their thermal analysis is done using computational fluid dynamics along with boiling simulation. This article investigates the subcooled boiling flow of pure water in a small channel and its dependence on different parameters such as working pressure, heat flux, and mass flux. The subcooling temperature of the inlet fluid is of interest. Numerical simulations were performed using axisymmetric models and the effect of different grid sizes on the accuracy of the results was investigated. The bubble separation diameter was also studied and its relationship with pressure, heat flux, mass flux, and temperature under the inlet coolant was analyzed. Finally, a database of generated simulation results and predictive models for bubble dynamics using machine learning algorithms were presented.

In this paper, an intelligent neural network controller is designed that can balance a quadrotor. After comparing the two types of independent reverse controllers and PID in the simulation environment, their differences are stored as data in the software. By specifying the input data of the controller and the target data and using the feedforward neural network architecture and a Narx architecture, intelligent controllers are designed and the obtained results are shown in several stability detection diagrams. The obtained results show that the equilibrium and vertical flight control are quite acceptable. Finally, the results are tested in a practical model on a real vertical plane. In the practical model, due to the limited amount of training data, the performance of Narx network is much more appropriate than the performance of the feed network. In fact, the complexity of the Narx recursion algorithm helps to interpret the complex dynamic behavior of the system well despite the low data. The disadvantage of this type of network is its very complex hardware implementation, which requires a much more powerful processor and memory to run on the controllers.

The present article deals with the optimal design of the contour or profile of the divergent part of an over-expanded supersonic nozzle in order to achieve the maximum possible thrust while maintaining the length and the exit to throat area ratio of the nozzle. To do so, a reliable and robust tool has been developed by combining computational fluid dynamics (CFD) and artificial intelligence. At first, the original profile is modeled by an innovative method using a third-order B-spline, and then by changing the breakpoints of the profile, a set of possible profiles is produced. This set of profiles has been analyzed by CFD. The geometry of the nozzle along with the thrust force obtained from the profiles have been used to train the artificial neural network. In the next step, the optimal profile was obtained by applying the genetic algorithm. Finally, the prediction of the artificial intelligence for the optimal nozzle thrust is compared with the value obtained from the CFD, which shows the validity of the present approach. The comparison between the original profile and the optimal profile for the nozzle pressure ratio of 14 shows a 36% increase in thrust and a 138% increase in the total pressure recovery factor. The comparison of nozzles for off-design conditions shows that the performance of the optimal nozzle is better than the original nozzle up to a pressure ratio of 30. If the nozzle pressure ratio exceeds 30, using the optimal nozzle instead of the original nozzle will not have priority.

The aerodynamic coefficients of airfoils are usually calculated through numerical and experimental methods causing time and cost consuming as well as depending on the airfoil surface cover. The estimation of aerodynamic coefficients of fabric covered airfoil through fuzzy logic, neural network and fuzzy-neural methods is the innovation of this research in order to determine a low-cost and fast method for optimal design as well as determining the aerodynamic coefficients of airfoils of vehicles having membrane wings. In the models, subsonic velocity was considered. Reynolds number and angle of attack were assumed as input and the values of lift and drag coefficients were assumed as output. Estimations were made on NACA2418 airfoil data, after which the final error of each method was compared. The mean squared error of lift and drag coefficients were 0.8023 and 4.3451e-04 for fuzzy model 0.0012 and 7.5767e-6 for neural network and 0.0697 and 0.0076 for network-fuzzy inference model respectively. The obtained results indicated good fitting of three studied models and best fitting for neural network model, which confirmed by the lift coefficients obtained from experimental tests done for validation.

Modern fabrication is to a large extent based on deformation processing. Plastic deformation process is a technique capable of producing metal products with high strength and good ductility. Using the parameters of load, temperature and the number of passes in twist extrusion, it is possible to produce an alloy with good properties and characteristics. Plastic deformation of AA6061-T6 aluminum alloy by twist extrusion is an important issue. In this study, we investigated the effect of load, temperature and the number of passes of twist extrusion on AA6061-T6. Using the input and output data, the process was modeled by the neural network method. In order to train the neural network, Neuro Solution software was used and for reducing the mean square error, the gradient descent momentum algorithm was implemented. Results showed that the effect of the number of passes and the load on tensile strength and hardness were maximum and minimum respectively.

In this study, a new method for the fault detection of the locomotive engine injector nozzle based on vibration analysis and statistical tests, inside artificial neural networks, is presented. For this point, first the under study received vibration signals in the frequency domain is divided into several smaller ranges and the RMS of each range is then extracted as a frequency property and given as an input to the neural network. Because the high selection of the features reduces the accuracy of the neural network, the extracted feature vector with different levels of significance passes through the T-test filters, firstly, and then enters the neural network as an input. Using of this method, the accuracy of the neural network increases from 78.4 to 94.6%, and also help to detect the frequency ranges. According to the results, the fault of the injector nozzle crack increases the intensity of vibrations in the upper band frequencies of 1500 Hz.

In ultrasonic vibration-assisted turning, an ultrasonic vibration is added to the tool, which leads to the periodical disengagement of the tool and the work-piece. In this research, an experimental study of ultrasonic vibration-assisted turning and conventional turning on Ti6Al4V Titanium alloy is conducted. First, by analyzing different parameters, four parameters are selected as the main affecting input parameters (cutting speed, feed rate, depth of cut, and ultrasonic vibration), and the effects of these four parameters are studied on two output parameters, namely tool wear and surface roughness. After the experimental tests, a statistical analysis is performed on the results and a neural network model is developed to predict the tool wear and surface roughness. The results show that the developed neural network model has a good agreement with the experimental results. In all experiments using ultrasonic vibrations, the tool wear and surface roughness were lower in comparison with the conventional turning. The cause of the tool wear and surface roughness reduction in ultrasonic mode are reducing the average forces applied to the tool, the alternative disengagement between the tool and the workpiece and increased dynamic stability of the process.

Gearboxes may be defective during the manufacturing, assembly or operation process. In gearbox manufacturing factories, it is important to identify defective gearboxes to prevent them from entering the consumption cycle. Audio signals indicate the operation of the gearbox and the condition of its internal components. Therefore, this system is a powerful method to diagnose the gearbox healthy or defective in different gears. In this study, intelligent gearbox troubleshooting using audio signals has been done. After receiving the audio signals, suitable algorithms for troubleshooting were proposed and based on that, the troubleshooting process was performed. In this test, the gearbox of the Thunder 90 passenger car in the assembly line was checked for the health or defect. The received audio signals are first processed in the time, frequency (fast Fourier transform) and time-frequency (discrete wavelet transform) domains. Then the feature was extracted from the processed signals in all three domains of time, frequency and time-frequency; Then, using the error test method and statistical inference method, the desired characteristics were selected for use in classification. Then, using various methods of artificial neural network classification and support vector machine, the intelligent diagnostic process was performed. The results showed that the backup machine classification method on data processed with discrete wavelet transform has an average error of less than 9%. This error is less than the error of other signal processing and classification methods in this experiment.

Load forecasting is one of the most important aspects of the effective management of power systems and depends on various factors such as the means and sources of electricity production by each company, electric load demand, weather factors, economic factors and human activities.. The main goal of this article is the STLF in the regional electricity grid of Guilan province. In this research, after collecting data including wind speed, relative humidity, temperature, dew point, holidays, day length and the effect of corona disease, pre-processing is done on them and they are divided into five clusters using K_Means clustering algorithm. In the following, the relationships between the independent and dependent variables of electricity consumption in Guilan province, using linear regression (LR) algorithms, artificial neural network (ANN) and support vector regression (SVR) along with the Grid search optimization method, were investigated and analyzed in Python software and modeled in the Google Colab environment. The results show that the SVR algorithm has higher accuracy and longer implementation time than the two algorithms of ANN and LR.

Tire online normal force has effects on vehicle safety and performance and dynamic control systems. It is influenced by too many parameters such as vehicle mass and center of gravity position and vehicle instantaneous dynamics states. In this paper, a new estimation algorithm is developed to estimate tires’ online normal forces during a maneuver. The proposed algorithm uses a dynamic measure module to make a hardware-software coupled model which is validated by real test data. The algorithm uses artificial neural networks advantages to estimate the vehicle mass distributions. A combination of real and model-generated data is used to train, test, and validate the artificial neural network structure. By applying two roll and pitch artificial neural network blocks, it estimates tires’ static normal forces. In this respect, the validated vehicle model instantaneously monitors the estimated values. The results show that the proposed algorithm estimates the vehicle total mass with less than 5 percent. In addition, the coupled model uses the estimated static values to estimate the tire's online normal forces with considering the measured vehicle dynamics states by dynamic module. Comparing the obtained results from the proposed method with the outputs from Carsim indicates the acceptable accuracy of this method.