جستجوی مقالات مرتبط با کلیدواژه « perceptron neural network » در نشریات گروه « فنی و مهندسی »

This article presents numerical studies on semi-active seismic response control of structures equipped with Magneto-Rheological (MR) dampers. A multi-layer artificial neural network (ANN) was employed to mitigate the influence of time delay, This ANN was trained using data from the El-Centro earthquake. The inputs of ANN are the seismic responses of the structure in the current step, and the outputs are the MR damper voltages in the current step. The required training data for the neural controller is generated using genetic algorithm (GA). Using the El-Centro earthquake data, GA calculates the optimal damper force at each time step. The optimal voltage is obtained using the inverse model of the Bouc-Wen based on the predicted force and the corresponding velocity of the MR damper. This data is stored and used to train a multi-layer perceptron neural network. The ANN is then employed as a controller in the structure. To evaluate the efficiency of the proposed method, three- story, seven- story and twenty-story structures with a different number of MR dampers were subjected to the Kobe, Northridge, and Hachinohe earthquakes. The maximum reduction in structural drifts in the three-story structure are 13.05%, 39.90%, 15.89%, and 8.21%, for the El-Centro, Hachinohe, Kobe, and Northridge earthquakes, respectively. As the control structure is using a pre-trained neural network, the computation load in the event of an earthquake is extremely low. Additionally, as the ANN is trained on seismic pre-step data to predict the damper's current voltage, the influence of time lag is also minimized.

A five-level Power Factor Correction incremental rectifier (PFC) is proposed in this paper. In this topology, the output voltage and current of the rectifier are controlled using the multilevel modulation and smart controller technologies. A multi-carrier pulse width modulation is used to create the switching pulse. In this topology, the number of semiconductor switches is reduced to 3. The smart controller is implemented using a Multilayer Perceptron (MLP) neural network and it is trained using the backpropagation algorithm. This controller is used instead of the well-known PID controller to control the input voltage and current. It should be noted that in this work, the goal is to design an intelligent controller using a neural network instead of a PID controller. The results obtained using this controller as compared to the PID controller show a decrease in the peak voltage, an increase in the rise time, and a ripple reduction in the output voltage. This study is conducted using the Simulink environment in MATLAB and the results suggest that a smart controller can be an alternative to the PID controller.

The current paper presents a robust optimum design of friction stir welding (FSW) lap joint AA1100 aluminum alloy sheets using Monte Carlo simulation, NSGA-II and neural network. First, to find the relation between the inputs and outputs a perceptron neural network model was obtained. In this way, results of thirty friction stir welding tests are used for training and testing the neural network. Using such obtained neural network model, for the reliability robust design of the FSW, a multi-objective genetic algorithm is employed. In this way, the statistical moments of the forces, temperature, strength, elongation, micro-hardness of welded zone, grain size and welded zone thickness are considered as the conflicting objectives. The optimization process was followed by multi criteria decision making process, NIP and TOPSIS, to propose optimum points for each of the pin profiles. It is represented that some beneficial design principles are involved in FSW which were discovered by the proposed optimization process.

With the availability of websites and the growth of comments, reviews of user-generated content are published on the Internet. Sentiment Classification is one of the most common problems in text mining, which applies to categorize reviews into positive and negative classes. Pre-processing has an important role when these textual contextsareemployed by machine learning techniques. Without efficient pre-processing methods, unreliable results will be achieved. This research probes to investigate the performance of pre-processing for the Sentiment Classification problem on three popular datasets. We suggest a high-performance framework to enhance classification performance. First, features of user's opinions are extracted based on three methods:(1) Backward Feature Selection; (2) High Correlation Filter; and (3) Low Variance Filter. Second, the error rate of the primary classification for each method is calculated through the perceptron. Finally, the bestmethod is selected through the fuzzy analytic hierarchy process. This framework is beneficial for companies to observepeople's comments about their brands and for many other applications. The current authors have provided further evidence to confirm thesuperiority of the proposed framework. The obtained results indicate that on average this proposed framework outperformed its counterparts. This framework yields 90.63 precision, 90.89 accuracy, 91.27 recall, and 91.05% f-measure

The main part of the eye is the retina covering the entire back section of the eye. Eye disease is one of the most important cause of disability and even death in developed countries as well as in developing countries. Disorders created in the retina that occur due to special diseases can be detected by specific retinal images. Studying the variations in retinal photos in a special time could help physicians to diagnose the associated diseases. In this paper, the detection of blood veins in retina photos was investigated. For this purpose, first a new method is proposed to promote the quality of retina photos by combining the histogram adjustment and gray level grouping. We use the feature vector to classify the pixels. Next, a method for classifying the images based on the feature extraction vector is required. The use of neural networks is one of the best and most widely used methods of machine learning for classification. We used a 3-layer Perceptron to classify pixels.