Supply Chain Disruption Prediction Using Deep Boltzmann Machines (Dbm)

Abstract

One of the biggest operational issues facing the world's manufacturing and logistics supply chains is the disruption to the supply chain. Traditional forecasting methods chiefly linear statistical methods have a basic inability to reflect the high dimensionality and stochastic interdependencies of modern supply networks. This paper presents a novel framework for disruption prediction based on Deep Boltzmann Machines (DBM), which is a type of undirected deep probabilistic graphical model that can learn multi-level latent representations from heterogeneous data from the supply chain. The architecture proposed here uses a three-hidden-layer DBM pre-trained with a layer-wise Restricted Boltzmann Machine (RBM) and then fine-tuned with Contrastive Divergence (CD-k) to extract hierarchical features from multivariate time-series inputs that include demand signals, supplier reliability indices, buffers, lead time variability, and macroeconomic disruption indicators. Experimental evaluation was performed using a benchmark dataset of 47,823 events across 14 industrial sectors (2012–2022). Data was partitioned into a 70:15:15 ratio for training, validation, and testing to prevent data leakage and ensure model generalizability. The proposed DBM model achieved a classification accuracy of 95.80% and a precision of 95.63%. To validate these improvements, a paired t-test was conducted, confirming that the DBM significantly outperforms baseline models including LSTM and BiLSTM (p < 0.05). All architectural elements (multi-layer depth, pre-training, optimization using CD-k) lead to significant improvement in performance, as confirmed by an ablation study. The results validate the DBM framework as a viable and meaningful solution for developing proactive supply chain risk management in high-technology manufacturing industries.