Phosphate mining companies in Morocco pose many environmental and occupational safety risks, especially through the release of airborne particulates, gas pollutants, and heavy metals. While there is increased implementation of monitoring systems within industrial mining contexts, current methodologies are still predominantly founded on rule-based systems or classical statistics that presume linearity in relationships between an arbitrary set of environmental parameters and the likelihood of an incident. Conversely, mining operations are characterized by intricately dynamic nonlinear combinations of numerous environmental and operational variables. As a result, a potential research opportunity exists for the application of sophisticated machine learning techniques that provide the ability to detect various levels of operational risk within phosphate mining scenarios. This study has three objectives. First, to examine the mining environmental and operational data from the phosphate mining sites to determine the mining operational conditions that present the highest risk. Second, to create a machine learning classification model which utilizes a Feedforward Neural Network (FNN) to identify operational states that are prone to incidents based on multivariate sensor data. Third, to assess the validity and reliability of the model using machine learning validity and reliability evaluation techniques along with statistical validation methods. In this study, an artificial intelligence-based approach for AI-based safety monitoring was proposed by using a Feedforward Neural Network (FNN) on a detailed data set of 1536 hourly measurements, directly recorded onsite at OCP plants in Benguerir and Khouribga. Environmental and industrial parameters (dust concentration, gas emissions, temperature, and toxic metal content) were measured using industrial-grade sensors certified for such a type of application. By means of training the proposed FNN model with adaptive gradient descent and dropout regularization with early stopping, a test mean squared error of 0.057 and over 85% accuracy on incident detection were obtained. Gradient tracking and m-adaptive validation proved the stability and convergence of the model. Emissions and dust were identified as the main risk classifiers in a variable importance analysis. The findings demonstrate that the mining sector may move from reactive to proactive safety management and validate the incorporation of AI into a real-time monitoring infrastructure inside the OCP ecosystem. Practical concerns of industrial data gathering, model interpretability, and the moral application of AI in high-risk settings are also addressed by the study.
Haloui et al. (Tue,) studied this question.