Energy efficiency characteristics analysis for process diagnosis under anomaly using self-adaptive-based SHAP guided optimization

Understanding energy efficiency patterns is crucial for developing more effective energy management strategies. However, disruptions from physical characteristics, such as particle accumulation, inhibit the construction of energy efficiency models, pose diagnostic challenges, and require additional fault detection models to isolate this uncertainty. Therefore, this study introduces a self-adaptive, long short-term memory-based energy efficiency model with adaptive moment estimation fine-tuning enhanced by Shapley additive explanation guided optimization. The model adapts its learnable parameters in real-time according to changes in process behavior, which helps in revealing energy inefficiency and particle accumulation through Shapley benchmarking under current operations and energy efficiency characteristics. Validated using a benchmark dataset and applied in a large-scale detergent industry, the model outperforms conventional methods, achieving testing r-squared values of 0.9895 and 0.9859, respectively. Moreover, the proposed model avoided formulating the relationship with faulty variables and demonstrated robust fault detection through energy efficiency patterns without needing fault labels, offering a novel approach to monitoring and optimizing energy efficiency. The adaptive weight analysis emphasized how energy efficiency is influenced by various input variables, leading to an hourly energy saving of 0.0271 GJ/t, equivalent to cost savings of USD 34,408 and a reduction of 115.44 t of carbon emissions.