AUTOENCODER NEURAL NETWORKS FOR EARLY WARNING OF QUALITY DECLINE IN NIGERIAN TABLE WATER PRODUCTION: A REAL-TIME MONITORING FRAMEWORK

This study developed an autoencoder neural network model for real-time anomaly detection in Total
Quality Management practice implementation, providing early warning of quality decline before
customer complaints materialize in Nigeria's table water industry. The research addresses the critical
limitation of traditional quality monitoring approaches that detect problems only after they have
manifested in customer complaints, product returns, or regulatory sanctions. Adopting a longitudinal
survey research design, data were obtained from 50 table water firms monitored over 24 months, with
model validation using reconstruction error thresholds and retrospective event analysis. The
autoencoder architecture comprised an encoder with 16 and 8 neurons, a 4-neuron bottleneck layer, and
a decoder with 8 and 16 neurons, trained using the Adam optimizer with a learning rate of 0.001. The
findings revealed that the autoencoder achieved 91% sensitivity and 84% specificity in detecting quality
decline events, with anomaly scores beginning to rise 2 to 3 months prior to complaint surges in 87%
of quality failure events. The optimal detection threshold, represented by reconstruction error exceeding
0.32, successfully flagged anomalies across multiple TQM dimensions, requiring no labeled historical
data for training. CRM declines were the primary anomaly driver in 54% of quality failure events,
followed by lean production at 28%, leadership at 13%, and benchmarking at 5%. The study
recommends that Nigerian table water producers adopt this autoencoder framework as a low-cost,
automated early warning system implementable as a cloud dashboard with SMS alerts, shifting quality
management from reactive complaint handling to proactive prevention.