Description
Potash mine stability and structural health monitoring (SHM) are key challenges because these operations face distinct geomechanical risks that progress promptly without early noticeable warning. In contrast to costly and time-intensive traditional destructive methods, acoustic impact evaluation is a non-destructive monitoring approach which enables capturing nonlinear and multi-dimensional in-situ geological complex features. This work focuses on acoustic feature extraction-based artificial intelligence (AI) frameworks for potash mine stability assessment. To capture the intrinsic fractures and geological stress orientations, acoustic impacts-based recordings were collected by striking the mine using an aluminum scaling bar across locations in multiple potash sidewalls and roofs. Each sound recording was evaluated by experienced miners, labeled as tight (safe) or drummy (unsafe or risk zone) for assessing potash mine stability. Numerous feature extraction methods such as principal component analysis, continuous wavelet transform, zero crossing rate and mel-frequency cepstral coefficients (MFCC) were adopted for acoustic feature reduction for reliable data transformation into AI frameworks. Multilayer perceptron neural network, support vector machine and deep residual neural network (ResNet) models were applied to carry out stability predictions and their efficacy assessed using precision, and accuracy metrics, supported by receiver operating characteristic curve analysis. The findings substantiate that ensemble fusion models based on ResNet-MFCC predictions showed superior generalization performance with high precision and classification accuracy on real time acoustic data compared to other AI models. This study builds a scalable and explainable foundation for proactive mine SHM, ground control strategies, supporting improved hazard mitigation, reducing environmental footprint and operational risk in critical mineral projects within subsurface mining settings.