یادگیری عمیق

One of the most crucial steps in ore exploration is the recognition of geological patterns and features. These features contain mineralogy, lithology, alteration, rock texture, etc. This stage has always been associated with many challenges. Among the challenges of this stage, we can mention the time-consuming and costly nature of this stage and the need for high expertise and human resources to recognize these patterns and features. In recent years, deep learning and machine learning have been adopted in earth sciences. In this research, by using the architecture of U-net, ore and waste were separated, and the grade pattern was identified using the core box images. For this purpose, iron minerals were segmented using binary image segmentation, trial-and-error methods were used to optimize the network, and finally, the model's accuracy for identifying ore was 91%. The IoU metric was utilized for further evaluation; this metric is a suitable criterion for the final evaluation of the image segmentation model, which has reached 75% in recognition of iron ores. For the final evaluation of the obtained model, the grade outputs of the model and the XRF analysis results of one core were compared. The network error was evaluated at 9%, which shows the good accuracy of the obtained model according to the real data.

Calculation of variograms and spatial continuity is one of the first and most important processes in geostatistical modeling, which is a long and experience-oriented process. Due to the complexities of calculating experimental variograms, interpretation and fitting the appropriate model are always the main challenges in this field. This article presents an intelligent variogram modeling method using deep learning that can increase the speed of variogram modeling and also prevent common errors in manual variogram model fitting. In this method, two convolutional neural networks are used. The first CNN network converts the initial data into a 2D simulated map based on various variogram models. For this purpose, it is necessary to train the first network with initial data and their corresponding simulations. The output of this model is entered into the second convolutional neural network as input, and the variogram parameters (including range, azimuth, ratio, and nugget effect) are predicted. In this article, the proposed algorithm is implemented on synthetic 2D data and the parameters of the CNN models are optimized. The accuracy of the proposed model was 97 %, and then the proposed algorithm was used for variogram modeling of Nouchon area geochemical data, which included the elements Cu, Zn, and Pb. the accuracy of the obtained model compared to manual fitting was 90%.

This study uses data modeling and text mining techniques for oil price predictions. To improve the model's explanatory capability, text features from internet news articles on crude oil are automatically extracted using convolutional neural networks. Additionally, various time series models employ a state analysis approach called convolution. The years 2021 to 2011 saw the collection of almost 13000 news items, and it was discovered that text mining and data from large Internet-based apps perform better for prediction than other approaches. This means that it is pretty fair to say that there is a parallel link between news headlines, those headlines, and searches in the Google search engine. This relationship is highly appropriate for correctly forecasting the price of crude oil.

The operation parameters of EPB-TBM have always been significant factors in tunnel constructions. So it is crucial to estimate the cutterhead torque and thrust force of the machine. In this study, by employing the multilayer perceptron artificial neural network (ANN-MLP) and multivariate regression (MVR) methods, the empirical models were developed to estimate the EPB operating parameters, including cutterhead torque and thrust force, in the rock section of the Tehran metro line 6, South extension (TML6-SE) project. In this section, the excavation was performed in a strong, blocky to massive rock. The machine was equipped with the disc cutters on the cutterhead as a cutting tool instead of rippers and drag bits. The mechanized excavation in this situation is unusual with using the EPB machines. The input data included the performance parameters such as penetration rate, earth pressure, cutterhead rotation speed, and cutter load. The statistical indices were used to verify the developed models. The results confirmed the accuracy of the models. The MAE loss function determined for torque in both training and testing stages predicted by the ANN was 0.0001 and 0.005, respectively. The MAE loss function determined for thrust force in both training and testing stages predicted by the ANN was 0.00016 and 0.010, respectively. The relationships between parameters in the dataset were investigated to obtain and offer new equations using the multivariable regression statistical method (MVR). The MAE loss function determined for cutterhead torque and thrust force was 0.0018 and 0.0010, respectively.