A crucial part of large mining projects is the transport of the extracted raw materials and the maintenance of the corresponding infrastructure. This is associated with enormous costs, as it mainly involves the use of heavy machinery with correspondingly high energy consumption and expensive repair costs.
Particularly in view of the fact that mainly in-mine materials are used for the construction of haulage routes and that these are subject to heavy loads, haulage routes are extremely susceptible to various types of damage, for example puddles, ruts and potholes, and to severe weather conditions such as floods and storms. On the one hand, this leads to damage to trucks, increased fuel consumption and injuries to workers; on the other hand, the destruction of conveyor lines can disrupt the entire conveying process, resulting in high operational downtime and lost revenue.
All these factors increase the costs and reduce the efficiency of mining operations, which is why haulage lines require regular maintenance and servicing. Usually, maintenance is coordinated by separate teams, which means a huge investment of people and time, especially for large mines.
Automated monitoring of conveyor lines can be achieved by real-time analysis of satellite and/or UAV data with a machine learning (ML) model. An ML algorithm is able to recognize patterns in labeled image training data and apply them to new data. Thus, a model can be developed that detects and classifies objects, such as potholes, ruts, stones, puddles, etc., in the image data from the conveyance routes and determines the geometry of the roadway, such as slope and width.
For satellite data, one can use open source data sources, such as ESA's Sentinel data. However, training an ML model requires a sufficiently large data set with high-resolution imagery to identify smaller objects and detect differences between the roadway and its surroundings. Open source satellite data usually does not provide the desired resolution in the visible spectral range for such purposes, while the use of unmanned aerial vehicles is costly.
In addition, a data source that provides new data at short time intervals is needed for regular monitoring. Depending on the data source, one has update frequencies of 3-15 days for satellite data. For UAV data, the time intervals can be chosen more flexibly and individually.
To detect and classify objects in images, image segmentation algorithms are used, such as U-Net or Mask R-CNN architectures. These classify each pixel of an image into the predefined categories, allowing different objects in an image to be identified and distinguished.
To solve the problem of low resolution and smaller update frequency of satellite data, different data sources can be used on the one hand, for example the use of Sentinel-1 data in addition to Sentinel-2 data, which provide different spectral information. On the other hand, additional time series data can be analyzed based on image data sequences and/or georeferenced vibration data measured by sensors in the trucks or at the conveyor belt using IoT sensors.
Receive news about Machine Learning and news around dida.
Successfully signed up.
Valid email address required.
Email already signed up.
Something went wrong. Please try again.