Predictive modeling based on neural networks for the early detection of anomalies in heat exchangers

Delayed detection of anomalies in industrial heat exchangers compromises their efficiency and operational reliability. Objective: To develop and evaluate predictive models based on neural networks to identify anomalies in shell-and-tube heat exchangers. Methods: A quantitative, applied study was conducted using a non-experimental, longitudinal, retrospective, and comparative design. Historical time series of temperature, pressure, and flow rate from an industrial plant located in El Palito, Venezuela, resampled every 15 minutes, were analyzed. Autoencoder, Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRU) architectures were compared alongside linear reference models. The evaluation included 10,354 test windows distributed across two scenarios: NIVEL, with a tendency toward blockage, and FLUCTUATION, without such a tendency. Results: In the LEVEL scenario, peak regression achieved the highest predictive performance (R² = 0.9276), followed by GRU (0.8980), persistence (0.8632), LSTM (0.8414), and Autoencoder (0.7560). In the FLUCTUATION category, persistence explained only 9.1% of the variance, while no model exceeded the 51.8% achieved by ridge regression. The discrimination of instrumental failure markers yielded AUC values ranging from 0.58 to 0.76. Compared to the physical clogging index, the Autoencoder identified advanced clogging conditions (AUC = 0.84), outperforming recurrent architectures and linear benchmarks. Conclusions: The GRU performed best among the neural networks, although it was outperformed by ridge regression. The comparison across scenarios highlighted the influence of the obstruction trend on predictive capability. The autoencoder demonstrated potential for discriminating advanced fouling conditions. The study establishes a reproducible procedure involving chronological partitioning, controlled preprocessing, and comparative evaluation; its application for early detection requires further validation using independent maintenance records and operational data