The manufacturing industry widely employs sound monitoring inspired by the ability of operators that can detect problems based on the sounds that machines emit. This monitoring serves as an integral component for predictive maintenance and productivity estimation. To facilitate real-time monitoring, edge devices are employed to manage and collect sound data. A streamlined Convolutional Neural Network (CNN) model was proposed, designed to execute all necessary computations for predictions, taking into consideration the limited computational resources of edge devices. Comparative analysis with renowned CNN models, namely VGG16, VGG19, ResNet-50, and YAMNet, reveals that the proposed CNN model is highly effective in event prediction from sound classification. Remarkably, the proposed model only required 2% of the prediction time as compared to the slowest and most complex model, while preserving an overall prediction accuracy of 98.9%. To balance the minor accuracy trade-off due to the simplicity of the proposed CNN architecture, an algorithm based on the First-In, First-Out (FIFO) queue system was developed. This method led to a reduction in the prediction error rate by up to 25% within a certain interval between the queue elements, in contrast to systems that do not implement this algorithm. The input feature adopted was the normalized Log-Mel spectrum with a duration of one second. A grid search method was utilized for hyperparameter tuning, with the aim of identifying the optimal hyperparameter combination within the constraints of the simplified CNN model architecture. To substantiate the real-time monitoring performance and superiority of the proposed CNN model, the same workflow was applied to the grain leg and plasma cutting machine using sound data collected from each. The results affirmed that the combination of the proposed CNN model and the developed algorithm exhibited exceptional performance under real-world conditions. In conclusion, for real-time monitoring that employs edge devices, the usage of a simplified CNN model and a customized algorithm is advocated to ensure continuous real-time monitoring devoid of errors or network instability.