System for predicting geotechnical instability using ai

Abstract

The proposed invention relates to a system that predicts geotechnical instabilities, which comprises the use of artificial intelligence for pattern detection, wherein the system autonomously analyses information captured by capture devices and, using different artificial intelligence and machine learning techniques, delivers an estimation of the future movement.

Description

[0001] GEOTECHNICAL INSTABILITY PREDICTION SYSTEM USING AI

[0002] Field of Invention

[0003] The present invention covers the field of mining, operations in mining areas, prediction of earth movements in mining operations, as well as the field of safety in mining operations, for workers and operators of mining machinery.

[0004] Background

[0005] In a mining operation, any instability can pose a catastrophic threat to the lives of workers, disrupting the mine's production chain, and resulting in millions of dollars in lost business costs.

[0006] From a mechanical perspective, a rock mass is a discontinuous, anisotropic, and heterogeneous medium composed of different types of rocks, loose materials, and discontinuities. The geomechanical behavior of a rock mass depends not only on the strength of each of its constituent materials, but also, and to a greater extent, on the interaction between them, thus determining its resilience and stability under the extraction and operation of a mining pit.

[0007] Despite technological advances in the tools available to obtain quality information and increasingly convenient access to rock mass characteristics, mining projects face a significant degree of uncertainty regarding slope stability during mining operations. Currently, massive rock slope failure in surface mining (pit mining) is continuously monitored by operators who review individual sensors. This is a manual process involving the information captured by different instruments, requiring constant human supervision. Identifying instrument trends is a process determined by the operator's experience. Based on this analysis, expert judgment determines whether or not the instruments indicate a possible rock mass failure.

[0008] For example, in document CL 201602304, entitled "A METHOD FOR ACTIVELY DETERMINING A LOAD WEIGHT FOR MINING EXCAVATING EQUIPMENT, COMPRISING ANALYZING AN EXCAVATION SURFACE, DETERMINING AN EXCAVATION SURFACE SHAPE, IDENTIFYING AN EXCAVATION PATH, CALCULATING A VOLUME AND CALCULATING A WEIGHT OF A MATERIAL IN A BUCKET OF THE EQUIPMENT, AN APPARATUS", a method is described in which an excavation surface is analyzed, the shape of the excavation surface is determined based on the analyzed excavation surface, an excavation path is identified from a mining excavation equipment, a volume of a material excavated by the mining excavation equipment is actively calculated based on at least the shape of the excavation surface and the excavation path of the mining excavation equipment and a weight of the material excavated by the equipment. Mining excavation is actively calculated based on at least one density factor.In document CL 202203750, entitled “SYSTEM AND METHOD FOR MEASURING ROAD ROLLING RESISTANCE”, reference is made to a system and method for measuring the rolling resistance coefficient for road maintenance by means of at least one aerial drone with photographic capacity and configured to take images of a road network, together with computational means configured with an artificial intelligence trained to analyze the images taken by said at least one drone and determine the rolling resistance coefficient CRR of a plurality of road segments, wherein said analysis is carried out based on the detection of different types of defects, the permitted speed of a predetermined vehicle and a defect score assigned for each road segment, in order to use devices that are not integrated into the vehicles or equipment traveling on the roads.The present invention also relates to a method of road maintenance that includes measuring the rolling resistance coefficient of a plurality of road segments of a road network with the measurement system and method of the present invention.

[0009] In document CN105139585B, it refers to an intelligent early warning and forecasting method for soil slope hazard, which includes the following steps: 1) Establish a rainfall monitoring station on-site, bury a soil infiltration meter and a displacement meter; 2) Real-time collection of rainfall conditions, soil permeability and slope displacement data and transmission to remote clients; 3) Data format processing; 4) Calculate the weight of rainfall conditions, soil permeability, slope displacement and other data that affect the safety factor of the slope; 5) Determine the safety factor of each time period of the slope;6) Calculate the relationship among rainfall conditions, soil permeability coefficient of the slope and deformation, displacement and safety factor in each period, and predict the occurrence time of soil slope failure; 7) Obtain the critical time of slope failure and instability of potential damage, and predict the occurrence time of disaster; 8) Combine steps 6) and 7) results to provide early warning and prediction of slope hazards. Compared with the prior art, the present invention is more reasonable and quantitative, and more realistically carries out early warning and prediction of soil slope hazards.

[0010] In document CN111504268B, entitled "Intelligent early warning and forecasting method for dangerous soil slope cases", which describes an intelligent early warning and forecasting method for dangerous soil slope cases, belonging to the technical field of geology, aims to provide an intelligent early warning and forecasting method for dangerous soil slope cases, which accurately forecasts such cases.The key points of the technical outline are as follows: S1: Select a monitored side slope area and divide it into multiple monitoring areas based on a finite element method; S2: Establish an information acquisition system and a hazard prediction system, and acquire the height, slope, and area of ​​a monitored area through the information acquisition system; S3: The information acquisition system wirelessly sends the monitoring data to the hazard prediction system.The soil permeability meter and displacement meter acquire the soil permeability and soil displacement of a monitored area in real time, the information acquisition system acquires the height, slope, and area of ​​the monitored area in real time, a three-dimensional model is constructed through the hazard prediction system, the three-dimensional model acquires various parameters in real time and dynamically predicts the side slope, and the possibility of landslides on the side slope is predicted more accurately.

[0011] Other documents describe technologies oriented to similar processes, such as document US10821992B2, which describes a method for predicting the risk of a work machine for transporting loads overturning. The method includes: obtaining topographical data of the terrain of a geographical area near the work machine using a terrain topographical detection system; extracting the terrain slope from the topographical data; obtaining information on the weight of the transported load using an integrated load weighting system or receiving load information from the device that loaded it; determining the maximum permissible slope for the work machine based on the weight information; and predicting the risk of overturning if it approaches a geographical area with a slope greater than or close to the maximum permissible.

[0012] For example, patent application US20190187684A1 describes methods and systems for data collection in mining environments with haptic feedback and alarms with continuous monitoring. A system may include data acquisition circuitry for interpreting various sensing values; data analysis circuitry for analyzing those values ​​and determining at least one of the following: sensor state, process state, or component state, and an alarm value; analysis response circuitry for executing an action in response to the sensor state, process state, or component state, and for continuously monitoring the alarm value; and a haptic user device.

[0013] Likewise, document US6424919B1 describes a method for selecting a design parameter for a drill bit. This method consists of inputting the value of at least one property of the terrestrial formation to be drilled into a trained neural network. This network is trained by selecting data from drilled wells. This data comprises values ​​of the formation properties in the formations through which the wells have been drilled. The values ​​of the formation properties correspond to the values ​​of at least one drilling operating parameter, the drill bit design parameter, and the values ​​of the penetration rate and wear rate of the drill bit used in each of the formations. The well data are input into the neural network for training, and the design parameter is selected based on the results of the trained neural network.

[0014] However, the technical problem has not been adequately resolved, since the various existing solutions do not yet constitute a definitive or comprehensive solution to the problem of predicting the risk of earth or material displacement in mining operations, especially reliable and safe, especially in small-scale mining, where some existing solutions cannot be applied for various reasons, including viability, economy or long-term effectiveness.To ensure a “quality” geomechanical prediction, beyond the experience of the technical team involved or the level of recognition of the pit slopes, it is critical to be able to identify, characterize and correlate the variables conditioning the initial stability of the massif and the variables triggering changes in geomechanical behavior, in a reliable, safe and low-cost manner, which allows its application and usefulness in a wide range of mining extraction processes, as well as in mining companies across the board.

[0015] To make a rigorous and accurate prediction, it is essential to model how each of these variables affects the stability of the pit slope, and thus be able to determine their impact when updating predictions and interpretations of expected behavior.

[0016] Summary of the invention

[0017] The proposed invention considers a predictive system that comprises the use of artificial intelligence capable of supporting human work through the automation of pattern detection that can anticipate a massive rock mass rupture in mining operations in order to provide more time to the mine operation in order to guarantee the safety of workers and the operation, where said system autonomously analyzes the information captured by monitoring systems and devices, and based on advanced computational learning, as well as different artificial intelligence and machine learning techniques, delivers an estimate of the occurrence of rock mass rupture updated in real time, agnostic to the user in charge of the system.

[0018] Detailed Description of the Invention

[0019] The proposed invention corresponds to a system for predicting geotechnical instabilities in a mining pit that allows determining the conditioning and triggering variables as input for a predictive model of geotechnical instabilities.

[0020] Monitoring and controlling slope stability in open-pit mining operations is critical to ensuring worker safety, maintaining operational continuity, and minimizing environmental impact. Mine slopes are excavations that are in constant motion, and any excessive displacement or deformation can lead to rock mass instability and other catastrophic events that can result in loss of life and significant property damage.

[0021] In this context, geotechnical radars have emerged as an invaluable tool for real-time monitoring of slope movements. These highly accurate radar systems can detect and record millimeter-scale displacements on the slope surface, providing a high-resolution data source that can be analyzed to identify patterns and trends that indicate potential instability.

[0022] The volume and complexity of data generated by geotechnical radars pose a significant challenge to their analysis and interpretation. Data sets can contain thousands or millions of data points, each representing the displacement of a specific point on the slope surface over time. These data must be efficiently processed, filtered, and analyzed to extract useful information and detect early signs of abnormal or dangerous movement.

[0023] This is where machine learning techniques, particularly recurrent neural networks such as LSTMs (Long Short-Term Memory), offer a promising approach for slope stability analysis and prediction. LSTMs are capable of capturing and modeling complex patterns in data sequences, making them ideal for analyzing time series such as radar-recorded displacements.

[0024] By using LSTM networks trained with historical displacement data and other relevant variables, it is possible to develop predictive models capable of anticipating future slope movements. These models can provide early warnings of potential instabilities, enabling the timely implementation of preventive and mitigation measures.

[0025] The main attributes of this predictive model are:

[0026] 1. Identification of the conditioning and triggering variables of the geomechanical and geotechnical model as input to the predictive model.

[0027] 2. Identification, analysis and dynamic prediction of geotechnical instability events in the mining pit.

[0028] 3. “State dependent” (ML) prediction.

[0029] 4. Generation of event-specific alerts and recommendations.

[0030] This solution consolidates data from multiple existing systems, including numeric, text, and logical data.

[0031] This data can be structured or unstructured and comes from various instrumented data sources currently operating in large-scale mining, without the use of invasive hardware. These data are processed using an artificial intelligence (AI / ML) model.

[0032] With this system, the Geotechnics, Safety, and Long-Term Planning and Development departments will be able to increase their ability to detect instabilities at the bank level, substantially improving their ability to anticipate low-velocity events and "sudden events" in order to take timely action.

[0033] The proposed invention includes a database comprising detailed information from technical publications relating to various current approaches to monitoring and predicting geotechnical instabilities.

[0034] In turn, the invention also includes the identification and assessment of a series of related variables, considered both conditioning and / or triggering a landslide event, given the geomechanical behavior of the massif, which will depend not only on the resistance of each of the materials that compose it, but also and to a greater extent, on the interaction between them, which today will allow for improved assertiveness in the prediction of instabilities.

[0035] The proposed invention comprises a specific methodology that includes a series of steps, variable identification, variable weighting and multivariate evaluation, where all related processes are known as “KEPLER Methodology” (FIGURE 1).

[0036] The proposed invention encompasses the technical feasibility of identifying conditioning variables in landslide events, and the technical feasibility of the different data sources (prisms, radars) is validated, confirming that the structure allows for the integration of variables and joint analyses with data from the aforementioned sources. A preliminary analysis is performed using real data contained in the database, which allows for the identification of commonalities between the behaviors of the different variables, providing an initial approximation for the integration and multivariate analysis of the compiled data.

[0037] The main stages of information processing correspond to:

[0038] 1. Efficiently collect and store raw data generated by geotechnical radars in a Data Lake environment, using the SQLite Database for its portability and ease of implementation.

[0039] 2. Conduct a comprehensive geographic analysis of the data collected by the radars to understand their spatial distribution and identify significant geospatial patterns that may influence project outcomes.

[0040] 3. Implement data regularization and cleansing processes to eliminate inconsistencies, outliers, and missing data, ensuring the quality of the data used in the analysis. Develop an optimized data warehouse to store processed data in a structured manner, facilitating its access and subsequent analysis by end users.

[0041] 4. Obtain a predictive model of risky displacement of mining slopes by training a LSTM (Long Short-Term Memory) network, taking advantage of the capacity of this recurrent neural network to model temporal sequences and predict future events based on historical data.

[0042] Data collection and storage:

[0043] Data processing is a crucial step in developing accurate and reliable machine learning models. Before data can be used to train and evaluate models, it must undergo a series of transformations and preprocessing to ensure its quality, consistency, and relevance.

[0044] In the context of this project, the data comes from various geotechnical monitoring sources used to monitor movements and displacements on mine slopes. This raw data may contain noise, outliers, inconsistencies, and complex patterns that make direct analysis difficult for machine learning algorithms.

[0045] Data processing for machine learning involves the following stages: Data collection and storage: Raw data is collected and stored in a Data Lake environment, a centralized repository that enables fast and flexible access to data in its original format.

[0046] Data Cleaning: Data cleaning techniques are applied to identify and address outliers, missing data, inconsistencies, and errors in the data. This may involve removing, interpolation, or imputation of missing values, as well as applying filters and anomaly detection algorithms.

[0047] Data transformation: Data is transformed to suit the requirements of machine learning algorithms. This can include feature normalization or scaling, categorical variable encoding, complex feature decomposition, and the application of transformations such as Fourier Transforms or Wavelet Transforms.

[0048] Feature selection: The most relevant and predictive features of the data are identified and selected, discarding those that do not provide significant information to the model. This is achieved through techniques such as correlation analysis, supervised or unsupervised feature selection methods, and dimensionality reduction.

[0049] Data Warehouse Generation: Processed data is stored in a data warehouse optimized for efficient analysis and query. The data warehouse organizes data into a relational schema, making it easier to access and use by machine learning models and analytics tools.

[0050] Data processing for machine learning is critical to ensuring that machine learning algorithms receive high-quality, relevant data, which in turn improves the accuracy, interpretability, and scalability of the resulting models.

[0051] Geographic Analysis of Data

[0052] Geographic analysis of data collected by geotechnical radars is a crucial step in understanding the spatial distribution of the data and identifying significant geospatial patterns that may influence project outcomes. This analysis involves integrating the data with geographic information systems (GIS) and applying spatial visualization and analysis techniques (FIGURE 2).

[0053] First, radar data must be georeferenced, associating each data point with its corresponding geographic coordinates. Once the data is georeferenced, GIS tools can be used to visualize it on interactive maps and perform advanced spatial analysis. Recorded data can be measured by prism-type sensors, spatially distributed over the volume to be monitored, radars, located at strategic points to monitor a considerable area of ​​slope, or other types of sensors such as geophones, piezometers, etc., where the spatial location of the sensors is important to characterize the location of displacement trends (Figure 2). Some of the techniques and analyses that can be applied include:

[0054] • Density Mapping: Using this technique, areas with a high concentration of data points can be identified, which may indicate areas of increased movement or activity on the slopes.

[0055] • Spatial Pattern Analysis: Statistical methods and spatial pattern analysis techniques are applied to detect clusters, hotspots, and spatial trends in the data.

[0056] • Overlay Analysis: Radar data is overlaid with additional layers of information, such as geological maps, slope models, and mining infrastructure layers, to identify potential correlations and factors influencing slope behavior.

[0057] • Line of Sight Analysis: The line of sight between the radars and the monitored slopes is evaluated, which can help identify areas with visual obstructions or limitations in radar coverage.

[0058] • Temporal Trend Analysis: Changes in the spatial patterns of the data are analyzed over time, which can reveal trends or changes in slope behavior.

[0059] • The results of geographic analysis can provide valuable information about the spatial and temporal characteristics of the data, identify priority areas of interest, and help better understand the factors influencing slope stability. This information can be used to optimize radar placement and configuration, as well as guide the development of more accurate and contextualized predictive models.

[0060] Data Regularization

[0061] Data regularization is a crucial step in processing and preparing data sets for training predictive models. Raw data collected by geotechnical radars may contain inconsistencies, outliers, and missing data that must be addressed to ensure the quality and reliability of the resulting models. In this particular case, the goal is to regularize the data and reduce it to a frequency of 1 minute (FIGURE 3).

[0062] The data regularization process involves the following tasks:

[0063] • Identification and Treatment of Outliers:

[0064] Outliers are observations that deviate significantly from the general trend of the data and can be caused by measurement error, noise, or exceptional events.

[0065] • Statistical techniques and outlier detection methods, such as standard deviation tests, quartile analysis, or clustering algorithms, are applied to identify these outliers.

[0066] Depending on the nature of the outliers and the project context, they can be eliminated, replaced by interpolated values, or treated as valid but infrequent values. • Handling Outlier Data:

[0067] Missing data can occur due to sensor failures, interruptions in data transmission, or recording errors.

[0068] Different techniques can be applied to handle missing data, such as interpolation (linear, polynomial, spline), use of average values, or model-based imputation.

[0069] The choice of method depends on the amount and pattern of missing data, as well as the underlying characteristics and assumptions of the data.

[0070] • Data Normalization and Scaling:

[0071] Data normalization and scaling are important steps to ensure that different data features have a similar scale and range.

[0072] Techniques such as min-max normalization, z-score normalization, or decimal normalization can be used to scale data to a specific range.

[0073] Data scaling can improve the performance and convergence of machine learning algorithms, especially when working with features of different magnitudes.

[0074] • Noise Management and Data Smoothing:

[0075] Radar data may contain noise due to environmental factors, electromagnetic interference, or sensor limitations.

[0076] Data smoothing techniques, such as moving average filters, Savitzky-Golay filters, or curve fitting, can be applied to reduce noise and highlight underlying trends and patterns in the data.

[0077] • Detection and Correction of Inconsistencies:

[0078] Quality controls should be implemented to detect and correct data inconsistencies, such as duplicate values, incorrect formats, or integrity constraint violations.

[0079] This may involve applying validation rules, consistency checks, and data cleansing procedures.

[0080] Predictive Model Development Based on Pattern Recognition

[0081] Identifying patterns in data is a crucial aspect of analyzing displacements recorded by geotechnical radars on mine slopes. These patterns can reveal valuable information about the behavior and dynamics of movements, which is essential for developing accurate predictive models and making informed decisions (FIGURE 4).

[0082] In this project, advanced techniques are implemented for identifying patterns in data, including: • Representative curves: A specialized algorithm is developed to identify representative curves in radar data. These curves correspond to significant behavioral patterns frequently observed in the data and can be used for subsequent analysis and prediction.

[0083] • Speed ​​and Cumulative Reverse Speed ​​Calculation: Important metrics such as speed and cumulative reverse speed are calculated from displacement data. These metrics provide valuable information about the movement and dynamics of objects detected by the radar, which can help identify relevant patterns and trends.

[0084] • Identification of zones of interest: An algorithm is developed to identify zones of interest within the representative curves. These zones correspond to areas of high activity or significant events that require more detailed analysis and may be indicators of potential instabilities or abnormal movements.

[0085] • Extraction and clipping of regions of interest: The identified regions of interest are extracted and clipped from the original representative curves. This allows for a more specific and detailed analysis of these areas, eliminating any irrelevant information or information unrelated to the event or pattern of interest.

[0086] • Fukuzono and Fukuzono Regression Algorithms: Fukuzono curve generation and Fukuzono regression algorithms are implemented for data analysis and prediction. These algorithms are widely used in the field of geotechnics and allow curve fitting to displacement data, capturing patterns and changes in movement behavior.

[0087] Identifying patterns in data is a critical step in understanding the underlying phenomena driving movement and displacement on mine slopes (FIGURE 4). By extracting and analyzing these patterns, the present invention can provide valuable insights that inform the development of robust predictive models and informed decision-making for risk management and safety in mining operations (FIGURE 5).

[0088] Development of predictive models with LSTM networks

[0089] The development of accurate and reliable predictive models is a key objective of this project. Specifically, the Long Short-Term Memory (LSTM) recurrent neural network technique is used to model and predict the behavior of mine slopes from displacement data collected by geotechnical radars. LSTM networks are a specialized type of recurrent neural network designed to process and model data sequences, such as time series, text, audio, and video. Unlike traditional feedforward neural networks, LSTM networks have the ability to retain and utilize information from the past to influence current predictions or outputs. In the context of this project, LSTM networks are used to capture the long-term patterns and dependencies present in mine slope displacement data.Their internal architecture, which includes state cells and input, forget, and output gates, allows them to retain and use relevant information from the past while processing data sequences.

[0090] LSTM Neural Network Architecture: The main components of an LSTM neural network are a sequence input layer and an LSTM layer. A sequence input layer feeds sequential or time-series data into the neural network. An LSTM layer learns long-term dependencies between time units in the sequential data. The following diagram illustrates the architecture of a simple LSTM neural network for classification (See Figure 6). The neural network starts with a sequence input layer followed by an LSTM layer. To predict class labels, the neural network ends with a fully connected layer, a softmax layer, and an output classification layer.

[0091] Developing predictive models with LSTM networks involves the following steps:

[0092] • Data preparation: The displacement data is preprocessed and divided into training, validation, and test sets. Normalization or scaling techniques are also applied to improve training performance.

[0093] • Network architecture definition: The structure of the LSTM network is defined, including the number of LSTM layers, the number of LSTM cells in each layer, additional dense layers, and the output layer. Network hyperparameters, such as learning rates, batch sizes, and smoothing factors, are also configured.

[0094] • Network training: The LSTM network is trained using training data, iteratively passing data sequences through the network and updating the weights and biases to minimize a predefined loss function.

[0095] • Validation and hyperparameter tuning: The performance of the LSTM network on the validation set is evaluated, and hyperparameters are tuned as needed to improve performance, using techniques such as cross-validation or hyperparameter search.

[0096] • Model evaluation: Once the LSTM network has been trained and tuned, its performance is evaluated on a separate test dataset. Relevant evaluation metrics, such as prediction error, accuracy, or domain-specific metrics, are calculated.

[0097] • Integration with other algorithms: Integration of trained LSTM networks with other algorithms and techniques, such as Fukuzono regression, is explored to further improve accuracy and understanding of underlying patterns in the data.

[0098] Predictive models based on LSTM networks offer a promising approach for analyzing and predicting slope stability in open-pit mining. Their ability to capture and model complex patterns in data sequences, combined with additional techniques such as Fukuzono regression, can provide valuable information for the early identification of potential instabilities and the timely implementation of preventive and mitigating measures.

[0099] The structure of the LSTM network used in this project is composed of several interconnected layers, each with different functions and configurations. The main layers are described below:

[0100] • Input Layer: This layer receives the input data, which in this case would be the sequences of mining slope movements recorded by geotechnical radars. The input data must be preprocessed and normalized before entering the network.

[0101] • LSTM Layers: The core of the network consists of one or more stacked LSTM layers. Each LSTM layer consists of multiple interconnected LSTM cells, each with its own state cell and input, forget, and output gates. The number of LSTM cells in each layer determines the network's ability to capture patterns and dependencies in the data.

[0102] • Dense (Fully Connected) Layers: After the LSTM layers, one or more dense (fully connected) layers can be added to perform additional transformations on the outputs of the LSTM layers. These dense layers can help combine and process the features extracted by the LSTM layers.

[0103] • Output Layer: The output layer produces the network's final predictions, which in this case may be the expected future displacements of the mine slopes. The output layer configuration depends on the type of prediction task, such as regression (continuous outputs) or classification (categorical outputs). In addition to these core layers, the LSTM network may include additional layers, such as regularization layers (dropout) to prevent overfitting, normalization layers (batch normalization), and activation layers (such as ReLU or Softmax) as needed.

[0104] For this project, the variables displacement, velocity, and inverse velocity are used as fundamental characteristics in the architecture of the LSTM network for predicting an object's motion. These variables provide crucial information about the dynamics and behavior of the system under study.

[0105] Displacement, measured in millimeters (mm), represents the change in an object's position over time. It is a direct measure of the motion experienced by the object and is usually calculated as the difference between the object's final position and its initial position. Velocity, on the other hand, refers to the rate of change of displacement with respect to time and is measured in millimeters per minute (mm / min). Velocity captures how fast the object is moving at a given time and is calculated as the first derivative of displacement. The inverse velocity, simply the inverse of speed, is used in some models to represent the behavior of the system in an inverse manner. Although its use can vary depending on the context, in this case it is used to capture inverse patterns in the object's motion.Acceleration, although not explicitly mentioned in the problem, is also a relevant variable as it represents the rate of change of velocity with respect to time. In the context of the problem, these variables are used as input features in the LSTM network architecture. The LSTM network uses historical data sequences of displacement, velocity, and inverse velocity to predict displacement and inverse velocity at the next time step. This allows the network to capture and model complex temporal relationships in the data and make accurate predictions about the object's future motion.

[0106] The specific structure of the LSTM network, including the number of layers, the number of LSTM cells in each layer, the sizes of the dense layers, and the hyperparameters (learning rates, batch sizes, etc.), will be determined through a tuning and validation process using the training and validation data. To explain in detail how the LSTM network architecture uses the variables displacement, velocity, and inverse velocity to predict displacement and inverse velocity, we must first understand what these variables represent and how they relate to each other.

[0107] • Displacement: Displacement represents the change in an object's position over time. In this context, displacement is measured in millimeters (mm) and is the variable we want to predict.

[0108] • Velocity: Velocity is the rate of change of displacement with respect to time. It is calculated as the first derivative of displacement. In this case, velocity is measured in millimeters per minute (mm / min).

[0109] • Inverse Velocity: Inverse velocity is simply the inverse of speed. It is used in some models to represent the behavior of the system inversely. It is an interesting variable in the context of prediction because it can provide useful information about how it relates to displacement.

[0110] • Acceleration: Acceleration is the rate of change of velocity with respect to time. It is calculated as the second derivative of displacement. However, acceleration is not explicitly mentioned in the description, but it is important to mention it as it is a variable that can influence the prediction of velocity and inverse velocity.

[0111] Now, regarding how these variables are used in the LSTM network architecture to predict displacement and inverse velocity, sequences of past states are used to feed the LSTM network. This means that the network takes a sequence of historical data (e.g., displacement, velocity, and inverse velocity at previous timesteps) as input and produces a prediction for the next time step. Since it was mentioned that three-hour sequences are used for prediction, the LSTM network is likely configured to take displacement, velocity, and inverse velocity data from the past three hours as input and predict the displacement and inverse velocity for the next time step.

[0112] The layers of the LSTM network are designed to learn and model the complex relationships between input variables over time. The LSTM network has the ability to remember previous states and learn temporal patterns in the data, making it well-suited for modeling time series such as this. After training the LSTM network on historical data, it is expected to be able to generalize and make accurate predictions about displacement and inverse velocity for new input data.

[0113] Training LSTM networks is a complex process that requires careful data preparation, proper configuration of the network architecture, and precise hyperparameter tuning. However, when done correctly, LSTM networks can capture complex patterns and long-term dependencies in the data, making them a powerful tool for modeling and predicting time series, such as the mine slope displacements in this project.

[0114] In another embodiment of the invention, the KEPLER system and / or methodology further comprises a system for monitoring displacements and deformations in mining slopes, based on computer vision and predictive modeling techniques. This complementary system combines optical flow calculation using the Farneback optical algorithm developed in Python with a data prediction model based on the XGBoost algorithm, providing advanced tools for evaluating and anticipating critical dynamics in ground structures. The main objective is to strengthen the monitoring and prediction capacity, improving the robustness of the analysis against variable environmental conditions.

[0115] In this embodiment, the algorithm estimates dense optical flow in monocular image sequences by calculating displacements between consecutive image pairs. An approach based on a fixed analysis mesh, superimposed directly on the processed images, is adopted. This mesh eliminates the need for semantic segmentation or feature-dependent landmarks, reducing errors associated with variations in illumination, texture, and contrast in the obtained images.

[0116] Additionally, in addition to calculating displacements, the system generates derived metrics such as velocity, acceleration, and inverse velocity, which are essential for describing terrain dynamics. These metrics are visually represented in 2D and 3D using advanced graphics libraries, such as Plotly, integrated into the workflow. To extend the monitoring capabilities, a predictive model based on XGBoost is incorporated, trained with time series of accumulated displacements. This model uses delays (lags) as predictor variables, generating estimates of future displacements. Predictions are evaluated using metrics such as the mean square error (MSE), whose interpretation is dynamically displayed in the graphical interface. This model is capable of identifying trends and fluctuations in the data, facilitating proactive decision-making in managing risks associated with slope stability.

[0117] Finally, the system integrates an interface developed in Streamlit, which allows users to upload images, view results in real time, and analyze the generated projections. This allows for modular and automated applicability, not only improving the efficiency of data analysis but also offering a tool adaptable to diverse mining operating conditions. Overall, this methodology represents a significant contribution to slope stability monitoring, ensuring greater reliability and accuracy in the identification and prediction of critical instabilities. In another embodiment of the invention, the technology consists of an automated system for calculating and analyzing displacement based on sequential images, using advanced computer vision techniques.The core of the system is based on the algorithm, implemented in Python, which allows for dense optical flow calculations, generating vector fields that describe the apparent motion of each pixel between consecutive frames. This approach guarantees a detailed and accurate representation of displacement; the system includes several key features:

[0118] 1. Automatic Calculation of Average Displacement: From the magnitudes of the displacement vectors, an average value representative of the global movement between each pair of images is obtained.

[0119] 2. Interactive Visualization: A graphical interface was developed using Streamlit, allowing users to upload images, select analysis ranges, and visualize results in 2D and 3D formats. This includes interactive graphs of accumulated displacement, velocity, and inverse velocity, as well as visual representations of optical flow with overlaid vectors.

[0120] 3. Dynamic and Metric Analysis: The system allows for the evaluation of displacement trends and patterns over time, identifying inflection points, accelerations, and decelerations. It also incorporates a predictive module based on XGBoost, which extrapolates future displacement and evaluates the quality of predictions using metrics such as the Mean Square Error (MSE).

[0121] 4. Scalability and Flexibility: The developed architecture allows for processing image sequences of different resolutions and formats, dynamically adjusting to the analysis needs. The interactive tools provided facilitate customization of the process according to the user's specific objectives.

[0122] This complementary system represents a comprehensive solution for automated image-based displacement analysis, combining advanced optical flow techniques, interactive visualization, and time series prediction. Its robust and scalable design makes it suitable for applications in diverse fields such as biomechanics, robotics, infrastructure monitoring, and other environments where motion analysis is critical.

[0123] This breakthrough lays a solid foundation for future improvements, such as the integration of deep learning algorithms or optimization for real-time processing.

[0124] Examples

[0125] Example 1. Example of the Methodology of the proposed invention

[0126] The proposed invention will utilize a working methodology that allows for complementary interaction between scientific and technological research and development, in order to achieve the stated objectives, such as the development of artificial intelligence algorithms that enable highly reliable prediction of the risk behavior of mining slopes. The methodology comprises an integrated set of techniques and methods that allows for a homogeneous and open approach to each of the project's lifecycle activities.

[0127] For the development of the system, two sequential stages are considered: (1) MCCR - Risk Behavior Classification Model: this first stage analyzes each of the prisms in real time, to define those that present risky behavior, and if defined positively, that is, "risky", they move on to the analysis stage with the MPOD collapse occurrence prediction model, considering their own behavior and that of nearby prisms.

[0128] (2) MPOD - Landslide Prediction Model: this second stage analyzes each of the prisms classified as risky and the data of nearby prisms to make a prediction of collapse over time.

[0129] The two-stage sequence model can be reviewed below. The stages for the joint development of the models are as follows: i) Definition of specifications: Research is conducted to define and document the different requirements of the hardware and software prototype for the system proposed in this project. The computational intelligence techniques and / or tools that can be used to generate the predictive models are also investigated.

[0130] i) High-level design: prior research is carried out to obtain a design of the general architecture of the proposed system, identifying its essential hardware and software modules that allow the fulfillment of the project objectives, mainly the prediction of failures in main fans in underground mining, together with the issuance of different types of alarms. iii) Detailed design: through a cyclical system of research and development, the detailed engineering of the different modules proposed in the high-level stage is obtained. iv) Prototype implementation: the detailed software design is materialized, incorporating the artificial intelligence modules that carry out the prediction. v) Testing stage: the correct operation of each software module will be verified independently, checking its proper functioning. vi) Integration: the different modules that make up the system are integrated.On the one hand, the correct functioning of the entire integrated system is verified, meeting the previously established requirements, and on the other, the fulfillment of the main objective, which is to achieve a high degree of reliability in prediction. vii) Operational testing of the system: Tests are performed in scenarios similar to the real environment.

[0131] It is necessary to highlight the need to address an independent methodology for generating the software module with the prediction algorithms, which is developed simultaneously with the one established in general mode for the project integration. Therefore, the first and second modules are, in turn, subdivided into two main stages: a) Pattern Identification b) Obtaining AI Models The above can be observed in the diagram in Figure 11.

[0132] For the development of predictive algorithms, the classic methodology for data mining projects will be used in principle, which is the discovery of knowledge in databases or KDD ("Knowledge Discovery in Databases"), the process of which can be observed in the figure below:

[0133] Subsequently, with the knowledge gained from the KDD process, the adoption of a classic methodology for developing data science projects, such as CRISP-DM 2 ("Cross Industry Standard Process for Data Mining"), is considered to extract knowledge from experimental data. CRISP-DM is based on standard processes and is an open-source solution that describes common activities used by analytics and data science experts. CRISP-DM defines a lifecycle focused on data exploration and analysis, consisting of six phases:

[0134] 1) Understanding project needs: Focused on determining project objectives and requirements. The Data Science solution to be used to achieve the project objectives will be evaluated to define a preliminary plan designed to achieve them.

[0135] 2) Data Understanding: is responsible for collecting data, understanding its nature and meaning, to recognize the validity of initial exploratory analysis of data sets and the formulation of the first hypotheses about them.

[0136] 3) Data Preparation: activities necessary to build the final dataset that will be trained through machine learning tools.

[0137] 4) Evaluation: of the models from the previous phase to determine if they are very useful to the defined needs and if they meet the project objectives.

[0138] 5) The review or improvement of the processes that can be carried out to implement the new solution and finally the approval of the continuity of the implementation of the solution.

[0139] 6) Implementation that involves the exploitation of the modules within a production environment.

[0140] The project aims to produce a software prototype, validated through testing that extends the capabilities developed in this project.

[0141] The scientific hypothesis proposed is the ability to predict beyond mere trends, as the current system achieves, and with less need for exogenous variables, thus increasing the model's truly predictive capacity. The scientific method is used as a general framework to test these hypotheses.

[0142] 2.2 Pattern Identification

[0143] The term pattern recognition refers to a type of information processing that has great practical importance and provides solutions to a wide range of problems. Some of these problems are solved by humans without much effort. However, in many cases, solving these problems using computational tools becomes extremely difficult.

[0144] Reducing these difficulties by building methods and algorithms is the main objective of first identifying and then using patterns in this Systems Engineering project.

[0145] In this regard, it is concluded that efforts should be concentrated on finding patterns in the following behaviors:

[0146] • Accumulated Reverse Velocity (ICV)

[0147] • Initial Distance Gradient (SD)

[0148] Example 2. Example of embodiment and implementation of the invention

[0149] For the development of the proposed invention, the following stages were considered:

[0150] E1: Understanding Project Needs: Focused on determining project objectives and requirements. The Data Science solution to be used to achieve the project objectives will be evaluated to define a preliminary plan designed to achieve them.

[0151] E2: Data Understanding: is responsible for data collection, understanding its nature and meaning, to recognize the validity of initial exploratory analysis of data sets and the formulation of the first hypotheses about them.

[0152] E3: Data Preparation: Activities required to build the final dataset that will be trained using machine learning tools.

[0153] E4: Evaluation of experimental models: using the data prepared in the previous stage, to determine whether they are very useful to the defined needs and whether they meet the project objectives.

[0154] E5: Review and / or improvement of the process: for the implementation of a prototype prediction model.

[0155] E6: Implementation of the initial prototype: to allow the evaluation of the technical feasibility of establishing a system with computational intelligence that generates early alerts for decision-making.

[0156] Example 3. Implementation of a model based on LSTM architecture

[0157] The present invention comprises promising results in the field of displacement and reverse velocity prediction on mine slopes. Through the implementation of a Long Short-Term Memory (LSTM) neural network architecture, both displacement and reverse velocity were successfully predicted, projecting both phenomena toward a 2-hour future horizon. This achievement is the result of exhaustive data analysis and processing, as well as careful configuration of the LSTM network parameters. The prediction process was carried out using a 3-hour data sequence as input for the LSTM network. This choice of time interval is critical to capturing relevant temporal patterns and relationships in the data, allowing the neural network to learn and generalize effectively. The prediction is clearly evident in Figure 8 of the study, which presents the results obtained.These figures provide a clear and concise visualization of the accuracy of the predictions made by the LSTM network. Figure 15A shows the prediction results for displacement, while Figure 15B presents the corresponding results for reverse velocity. In both figures, a close correspondence can be observed between the predictions made by the LSTM network and the actual values ​​observed in the test data. This finding suggests that the neural network has effectively learned the underlying patterns in the training data and has been able to accurately generalize to predict future displacement and reverse velocity behavior on mine slopes.

[0158] It is important to note that the prediction is not only limited to the visual agreement between the predictions and the real data, but is also supported by quantitative evaluation metrics. During the training and validation process of the LSTM network, various evaluation metrics have been used, such as the mean square error (MSE) and the coefficient of determination (R A 2) to quantify the accuracy of the predictions. The results obtained from these metrics confirm the effectiveness of the LSTM network in predicting displacement and inverse velocity on mining slopes. The MSE values ​​obtained are low, indicating a good ability of the network to accurately predict future values. Furthermore, the coefficient of determination R A 2 is close to 1 , suggesting that most of the variability in the test data is explained by the neural network predictions.

[0159] Example 4. Implementation of KEPLER methodology.

[0160] The summary of the KEPLER methodology, as indicated by the example algorithm described in Figure 16.

[0161] Example 5. Example of application of the invention including optical flow information, (KEPLER-Fasat Gamma)

[0162] The developed code's main purpose is to analyze optical flow and displacement within a sequence of images using the Farneback algorithm. Additionally, it incorporates functionalities to predict displacement trends using machine learning models such as XGBoost. The code's operation, input and output data, and the technical and theoretical justification for each component are described in detail below.

[0163] 5.1 The User Interface

[0164] Component The system's user interface, developed with Streamlit, offers an interactive and accessible experience for analyzing optical displacements and flows. It is designed to guide the user through the various stages of image processing and the interpretation of the generated data. Key features of the interface include:

[0165] A. Image Upload and Preprocessing i. File Upload: Users can upload sequential images in common formats such as JPG, JPEG, or PNG using a multi-upload component. The interface automatically validates uploaded images, ensuring they are valid and, if necessary, adjusting their size to ensure uniformity between image dimensions.

[0166] i. Dynamic Notifications: If images do not meet requirements, such as incorrect format or insufficient number of images, the interface displays clear and informative error messages. Once successfully uploaded, the user is notified of the number of images processed.

[0167] B. Analysis and Visualization Options i. Image Selection: Users can select a start and end image to define the analysis range. The interface ensures that the selection is valid (the start image must precede the end image) and alerts about incorrect settings.

[0168] i. Optical Flow Visualization: Several visualization options are offered: a. Original Image: Displays the selected image for reference. b. Farneback Optical Flow: Presents a visual representation of the optical flow using a color map and vectors that indicate the direction and magnitude of motion. c. 2D Optical Flow: Uses interactive heat plots to visualize flow magnitudes on a continuous scale. d. 3D Optical Flow: Represents flow magnitudes as a three-dimensional surface, highlighting variations in displacement in different areas of the image.

[0169] C. Interactive Graphics: i. Full Displacement: Displays a line graph representing the average displacement between pairs of images.

[0170] i. Cumulative Sum: Plots the progressive accumulation of displacement over time. iii. Velocity: Plots the discrete derivative of the accumulated displacement, identifying changes in motion intensity. iv. Inverse Velocity: Provides a visualization of the reciprocal of velocity, highlighting regions of low motion intensity.

[0171] D. Predictive Module with XGBoost i. Customizable Input: The user can define the number of future steps to predict average displacements. The interface takes this value and automatically adjusts the predictive model.

[0172] i. Real-Time Results: The predictive module generates predictions and plots them alongside historical data in a combined chart. It also calculates metrics such as the Mean Squared Error (MSE), displaying its value and a visual label (green, yellow, red) indicating the quality of the predictions. iii. Metric Exploration: Interactive charts allow the user to explore trends, such as cumulative sum, velocity, and inverse velocity, with the included predictions. This helps assess the impact of future data on system evolution.

[0173] E. Interactivity and Ease of Use i. Interactive Charts: All charts include zoom, scroll, and highlighting tools, allowing the user to explore specific details of the data.

[0174] i. Dynamic Messages: The interface uses real-time messages to inform the user about the progress of the analysis, issues encountered, or highlighted results. iii. Customization: Colors, annotations, and visualization options are designed to maximize data clarity and interpretability.

[0175] 5.2 Data Upload

[0176] The system works with images in JPG, JPEG, or PNG formats, which must be sequential to adequately capture movement or displacement. If the image dimensions are inconsistent, the code automatically resizes them, ensuring the uniformity necessary for analysis. Images are uploaded through an interactive form implemented in Streamlit and processed using the loadjmages function, which converts them to grayscale. The validity of each file is also verified, ensuring that at least two images are uploaded to begin the analysis. The type of images used as input data and the input configuration via Streamlit are then observed.

[0177] 5.3 Optical Flow

[0178] Optical flow computation is performed using the Farneback algorithm, available in OpenCV. This algorithm generates a flow field that describes the horizontal and vertical displacement vectors between consecutive images. Configurable parameters include: pyramid scale (pyr_scale), pyramid levels (levels), window size (winsize), iterations per level (iterations), and polynomial parameters (p°ly_n and poly_sigma). These values ​​ensure a balance between accuracy and computational performance. For optical flow visualization, an HSV color-based approach is used, where hue represents the direction of motion, and saturation and value indicate its magnitude. Additionally, 2D and 3D visualizations are generated using libraries such as Plotly, allowing for interactive exploration of the results.These representations include flow vectors overlaid on the images, 2D heat maps, and 3D surfaces that highlight the magnitude of displacement.

[0179] It is noteworthy that the user can see the change between the uploaded images considering the first image as a reference as an initial state as shown below with respect to image 1 and image 13, below, the variation in deformation between these two images is observed, Faneback optical flow, deformation in 2D and 3D. 5.4 Data Capture for Graphics

[0180] To capture displacement data, we begin by calculating the optical flow between consecutive pairs of images. We then use the Farneback algorithm, an advanced method that generates a vector field describing the apparent motion of each pixel between the images. This vector field is essentially a dense representation of displacement vectors, where each vector has horizontal and vertical components. From these vectors, we calculate the magnitude of the displacement at each pixel. We then average these magnitudes across the entire image, thus obtaining a value representative of the overall displacement between two consecutive frames.

[0181] Once the average displacements have been calculated for the entire image sequence, I proceed to represent this data using specific graphs. The first graph I generate is the cumulative displacement. In this graph, the average displacement values ​​obtained for each image pair are iteratively added together. This graph allows us to visualize the overall progress of the movement over time, identifying global trends.

[0182] The next step is to calculate the velocity, defined as the discrete derivative of the accumulated displacement. This is obtained by subtracting consecutive values ​​of the accumulated displacement. The calculated velocity is a metric for identifying changes in the intensity of motion and provides information about accelerations or decelerations in the analyzed system.

[0183] Also, the inverse velocity, which corresponds to the reciprocal of velocity. This analysis is particularly valuable in contexts where slower motion has specific implications or greater relevance. I carefully handle cases where velocity tends to zero to avoid indeterminate divisions, using thresholds and regularization techniques in the calculation.

[0184] For visualizations, I use the Plotly library, which allows me to create interactive and highly customizable graphs. This includes line graphs that clearly depict trends in cumulative displacement, velocity, and reverse velocity. I also incorporated annotations and markers that highlight key events, such as velocity spikes or turning points in motion. The results are shown below with 20 uploaded images.

[0185] 5.5 Data Forecast by XGBoost

[0186] The system includes an XGBoost-based prediction module optimized for modeling time series dynamics in the average shift. The model uses as input lags generated from historical values ​​of the accumulated shift, configuring a predictive feature framework that captures short-term temporal relationships. In this case, up to 5 consecutive lags are generated as independent variables. The dataset is divided into training and testing partitions using 80 / 20 cross-validation with adjusted proportions. The model tuning employs optimized hyperparameters such as n_estimators=200, learning_rate=0.05 and a maximum depth of 4 (max_depth=4), ensuring a balance between complexity and generalization capacity.

[0187] Predictions are evaluated using the Mean Square Error (MSE), the interpretation of which is reinforced by an on-screen color scheme: green (<10) for highly accurate predictions, yellow (10-20) for moderate accuracy, and red (>20) for less reliable results. This approach allows the user to quickly diagnose model quality.

[0188] The interface presents interactive graphics that show:

[0189] 1. Extended Prediction: Representation of the observed displacement together with the predicted values ​​in the future range defined by the user.

[0190] 2. Cumulative Sum: An extension of the cumulative sum of historical displacement including predictions.

[0191] 3. Velocity (Derivative): Calculation and visualization of displacement rates of change, augmented by projected values.

[0192] 4. Inverse Velocity: Analysis of fluctuations in the reciprocity of the predicted velocity.

[0193] These visualizations allow for a comprehensive assessment of trends and include a trend line based on a polynomial fit to the predictions, providing additional context on the overall direction of future displacement.

[0194] The output data includes processed images with visual representations of optical flow, interactive graphs showing accumulated displacement, velocity, and inverse velocity, as well as predictions of future displacement with estimated trends. These results enable in-depth and comprehensive analysis of the motion detected in the images.

[0195] Optical flow analysis has significant applications in diverse fields such as robotics, environmental monitoring, and computer vision. This project combines advanced computations, interactive visualizations, and machine learning-based predictions, providing a comprehensive tool for research and industrial applications. The clear presentation of results ensures easy interpretation, while the predictive capabilities add value for informed decision-making. Brief description of the figures

[0196] Figure 1. Summary of the technology's operation, which includes a central methodology (KEPLER) that enables the creation of a predictive model and the delivery of predictions based on data collection and weighting of individual variables, as well as multi-variable analysis.

[0197] Figure 2: Example of geo-referencing of radars (top) and prisms (bottom) for the application of the instability detection code.

[0198] Figure 3: Descriptive image of the data regularization, with a frequency of 1 minute.

[0199] Figure 4: Image showing the representative curve and analysis pattern.

[0200] Figure 5: Flowchart of the algorithm for identifying representative curves.

[0201] Figure 6. Example of deformation behavior descriptors, which determine the starting points of a potentially dangerous deformation event, based on A. Relative Velocity, and B. Cumulative Reverse Velocity.

[0202] Figure 7: Use of different variables (relative velocity and inverse of velocity and displacement) used to feed pattern recognition algorithm and estimation of future instability.

[0203] Figure 8: Architecture of a simple LSTM neural network for classification.

[0204] Figure 9: Flowchart of the process for defining the structure of the LSTM network.

[0205] Figure 10: General diagram of the predictive system, divided into two main stages.

[0206] Figure 11: Stages for the development of the MCCR and MPOD module.

[0207] Figure 12: Main steps within the interactive and iterative KDD process.

[0208] Figure 13: Pattern extraction stages.

[0209] Figure 14: Example of linear projection to predict “failures” by means of regression of the behavior of the Reverse Velocity.

[0210] Figure 15: Displacement prediction with a horizon of 2 hours into the future, based on A. Displacement, and B. Cumulative Reverse Velocity.

[0211] Figure 16. KEPLER System Methodology. Details of the KEPLER methodology used in the predictive model for surface rupture detection in mining operations, showing some of the critical steps in the methodology, cutoff values, and results determination.

[0212] Figure 17. Example of optical flow application on mining slopes. Figure 18. Displacement graphs and displacement derivatives, with predictions.

[0213] Figure 19. Testing and application images of the complementary predictive optical flow system (Fasat Gamma) of the invention.

Claims

Claims 1. A system for predicting geotechnical instabilities in a mining pit that allows determining the conditioning and triggering variables as input for a predictive model of geotechnical instabilities in real time, CHARACTERIZED in that it comprises developing a preliminary analysis using real data contained in the database, which allows identifying the common points between the behaviors of the different variables, generating the initial approximation for the integration and multivariate analysis of the compiled data, which in turn comprises: a. An initial set of input data, b. At least one device for capturing information in real time, where said device can be selected from the list comprising: a radar device, a prism, an optical device for image capture, an optical device for capturing consecutive images and / or videos, and any combination of these. c.A remote access platform, comprising: a set of instructions, methods and data analysis protocols, used for the development of the predictive model, d. A database, comprising information relating to the system of the invention, the platform of the invention, the initial set of data, real-time data received from at least one real-time information device and a library of AI models available to be used, trained, modified and applied in the development of predictive models, in the proposed invention. e. Generation of a geotechnical alert, indicating a potential landslide or material failure, directly to the user and / or to a vehicle control system at a mining site.

2. A method for predicting geotechnical instabilities in a mining pit that allows determining the conditioning and triggering variables as input for a predictive model of geotechnical instabilities in real time, CHARACTERIZED because the main stages of information processing correspond to: a. Efficiently collecting and storing the raw data generated by at least one information capture device in the Database, b. Performing an exhaustive geographic analysis of the collected data, in order to understand its spatial distribution and identify significant geospatial patterns that may influence the project results, c.Implement data regularization and cleansing processes to eliminate inconsistencies, outliers, and missing data, ensuring the quality of the data used in the analysis by developing an optimized data warehouse to store the processed data in a structured manner, facilitating its access and subsequent analysis by end users. d. Obtain a predictive model of risky displacement of mining slopes by training a LSTM (Long Short-Term Memory) network, leveraging the capacity of this recurrent neural network to... model time sequences and predict future events based on captured data and historical data present in the database, and where in addition, the sequential stages of the process comprise in turn the following steps: e. Gathering information from the place where the implementation of the proposed invention will take place, or site of activity, f. Capturing information from the site of activity, g. Installation of at least one device for capturing information in real time at the site of activity, h. Capturing information from the device for capturing information in real time, and entering said information into the database of the invention, i. An information evaluation process, which comprises the following stages: i. Identification of patterns from the data obtained, i. Obtaining specialized LA models for the corresponding scenario and model, iii. Preparation and development of a Risk Behavior Classification Model (RBCM); iv. Selection of the most appropriate LA model available in the database based on the RBCM. v. Development of a Landslide Prediction Model (MPOD): vi. Delivering the recommendation or alert in real time based on the prediction models and data obtained; vii. Iterating the process each time new data is entered or a real-time change event occurs during system operation at the mining site.

3. The method of the invention according to claim 2, CHARACTERIZED in that the pattern identification comprises the identification of at least the following behaviors: i) Inverse Accumulated Velocity (ICV), and i) Initial Distance Gradient (SD), and wherein in addition, the pattern identification comprises the following stages: a. E1: Understanding the project needs: focused on determining the objectives and requirements of the project. The Data Science solution to be used to achieve the project objectives will be evaluated, in order to define a preliminary plan designed to achieve the objectives. b. E2: Understanding the data: is responsible for collecting data, understanding its nature and meaning, to recognize the validity of the initial exploratory analysis of the data sets and the formulation of the first hypotheses about them. c. E3: Data preparation: Activities required to build the final dataset that will be trained using machine learning tools. d. E4: Evaluation of experimental models: Using the data prepared in the previous stage, to determine if they are useful for the defined needs and if they meet the project objectives. e. E5: Review and / or improvement of the process: For the implementation of a prototype prediction model. f. E6: Implementation of the initial prototype: To assess the technical feasibility of establishing a system with computational intelligence that generates early alerts for decision-making.

4. The method of the invention according to claim 2, CHARACTERIZED in that the preparation and elaboration of the MCCR and MPOD models comprises the following steps: a. Data preparation, wherein the displacement data is preprocessed and divided into training, validation and test sets, and wherein in addition, normalization or scaling techniques are applied to improve training performance. b. Definition of the network architecture, wherein the structure of the LSTM network is defined, including the number of LSTM layers, the number of LSTM cells in each layer, the additional dense layers and the output layer, and wherein it also comprises a stage of configuring the network hyperparameters, which can be selected from the list comprising: learning rates, batch sizes and regularization factors. c.Network training, where the LSTM network is trained using the training data by iteratively passing data sequences through the network and updating the weights and biases to minimize a predefined loss function. d. Validation and hyperparameter tuning, where the performance of the LSTM network is evaluated on the validation set and hyperparameters are tuned as needed to improve performance, using techniques that can be selected from the list comprising: cross-validation and hyperparameter search. e. Model evaluation, where once the LSTM network has been trained and tuned, its performance is evaluated on a separate test dataset; and where relevant evaluation metrics such as prediction error, accuracy, or domain-specific metrics are also calculated. f.Integration with other algorithms present in the database, where the integration of trained LSTM networks with other algorithms and techniques is explored to further improve accuracy and understanding of underlying patterns in the data.

5. The method of claim 4, characterized in that the main layers of the LSTM network comprise the following: a. Input Layer, wherein said layer receives the input data, and wherein the input data must be preprocessed and normalized before entering the network; b. LSTM Layers; c. Dense Layers; d. Output Layer; e. Additional Layers: wherein the method of the invention may comprise, in addition to these main layers, at least one additional layer, which may be selected from the list comprising: a regularization layer to avoid overfitting, a normalization layer, an activation layer, and any combination thereof, as necessary to adjust and improve the predictive model.and where in addition, the method of the invention comprises as fundamental characteristics in the architecture of the LSTM network for the prediction of the movement of an object, a set of variables and these are: i) displacement, i) speed, iii) inverse speed, and iv) acceleration, where: i) the displacement, measured in millimeters (mm), represents the change in the position of the object over time, and corresponds to a direct measurement of the movement experienced by the object. i) Velocity, which corresponds to the rate of change of displacement with respect to time and is measured in millimeters per minute (mm / min), and where velocity also captures the speed at which the object is moving at a given time and is calculated as the first derivative of displacement, iii) Inverse velocity, which corresponds to the variable that is used in some models to represent the behavior of the system in an inverse manner, and where it is also used to capture inverse patterns in the object's motion, and iv) Acceleration, which represents the rate of change of velocity with respect to time, 6. A method for the operation and functioning of the system described in claim 1, CHARACTERIZED in that it corresponds to the methodology called KEPLER, and comprises the following steps: a. Data entry from the instrumentation system; b. Binary type progress evaluation, where if the minimum data required to indicate the process is reached, the process continues to the next step; and where if the minimum data is not reached, more information is requested to be entered and a return to the previous step is made; c. Instability evaluation for each sensor and / or information capture device, where: i. If the recorded data does not indicate a tendency towards instability at the analyzed point, then a green light is displayed, indicating that the mining operation can be continued safely. ¡i. If the recorded data effectively indicate a tendency towards instability of the analyzed point, then at least a predictive model and reverse velocity are developed for the sensor and / or information capture device, where based on the slope of the reverse velocity forecast, at least three states or conditions are calculated for the sensor and / or information capture device, as follows: a. If the slope is greater than 0.4 m, a RED indicator is shown, which represents a “high risk” condition for operation at the mining site, b. If the slope is in a range less than 0.4 m and greater than 0.2 m, an ORANGE indicator is shown, which represents a “moderate risk” condition for operation at the mining site, c. If the slope is less than 0.2 m, a YELLOW indicator is shown, which represents a “lower risk” condition for operation at the mining site.

7. The system of claim 1, CHARACTERIZED in that it further comprises a complementary system, wherein the complementary system consists of an automated system for calculating and analyzing displacement based on sequential images, using advanced computer vision techniques, wherein the core of the complementary system is based on a specialized algorithm implemented in Python, which allows calculating optical flow in a dense manner, generating vector fields that describe the apparent movement of each pixel between consecutive frames; and wherein said complementary system allows guaranteeing a detailed and precise representation of the displacement in the system of claim 1; and wherein the system in turn comprises the following key functionalities: i.Automatic Calculation of the Average Displacement, obtained from the magnitudes of the displacement vectors, where an average value representative of the global movement between each pair of images is obtained. i. Interactive Visualization, comprising a graphical interface that allows users to upload images, select analysis ranges, and view results in 2D and 3D formats, where said visualization includes interactive graphs of accumulated displacement, velocity, and inverse velocity, as well as visual representations of optical flow with overlaid vectors. iii. Dynamic and Metric Analysis, where this complementary system functionality incorporates a predictive module that extrapolates future displacement and evaluates the quality of the predictions using metrics such as the Mean Square Error (MSE), which allows evaluating displacement trends and patterns over time, identifying inflection points, accelerations, and decelerations. iv. Scalability and Flexibility, where the developed architecture allows processing image sequences of different resolutions and formats, dynamically adjusting to the needs of the analysis, facilitating the customization of the process according to the user's specific objectives, through the interactive tools provided.

Images