طبقه‌بندي بررسي ميزان مصرف انرژي الكتريكي در صنعت فولاد به كمك تكنيك يا الگوريتم پرسپترون چندلايه خطي

In recent years, with the rapid development of smart manufacturing systems an‎d the exponential growth of data generated in industrial environments, the accurate analysis of energy consumption patterns has become a fundamental pillar in process optimization. In this context, developing classification models capable of simultaneously delivering high accuracy, robustness against noise, an‎d full interpretability remains a major challenge. While advanced machine learning models (particularly deep neural networks) lack the necessary transparency for critical an‎d safe industrial decision-making due to their “black-box” nature an‎d complex nonlinear structures, classical linear methods often suffer from lower classification accuracy due to their limited feature-learning capacity. This contradiction has created a significant research gap in the development of reliable intelligent systems. Aiming to bridge this gap an‎d achieve absolute interpretability without a substantial dro‎p in performance, the present study proposes an innovative binary classification framework. The core idea of this research is the re-engineering of the classical Multilayer Perceptron (MLP) neural network architecture. In this approach, the network structure was fundamentally transformed by eliminating conventional nonlinear activation functions an‎d replacing them with a fully linear activation function (Lin) across all hidden layers. This architectural innovation ensures that the hierarchical weight-learning process ultimately reduces to a single, fully linear model. Consequently, while leveraging the optimization capabilities of neural networks, the exact impact an‎d weight of each input feature on the final output an‎d decision of the model can be directly an‎d transparently extracted an‎d analyzed. For implementation, model training, an‎d comprehensive eva‎luation, all developed algorithms were coded an‎d executed within the Google Colab cloud computing environment. The proposed model was tested on complex, real-world industrial energy consumption datasets to eva‎luate its efficiency under operational conditions. In the eva‎luation phase, the performance of this linear network was compared against a suite of the most robust baseline linear classification algorithms, including Logistic Regression an‎d Linear Support Vector Machine (Linear SVM), using stan‎dard eva‎luation metrics. Empirical an‎d statistical analysis of the results proves that the proposed model significantly outperforms its competitors in terms of efficiency. By achieving an exceptional overall accuracy of 0.9050.9050.905, this framework registered a 333 to 555 percent increase in classification metrics compared to other linear methods. From an industrial application perspective, this performance leap implies that out of every 100100100 operational an‎d control decisions, the proposed model makes, on average, 333 to 555 fewer decision errors. This reduction in error directly translates to the prevention of energy waste, a decrease in unplanned downtimes, an‎d a remarkable increase in the reliability of energy management systems. Ultimately, this study demonstrates that through the intelligent redesign an‎d linearization of baseline structures such as the MLP, it is possible to achieve a level of competitive performance that eliminates the need for complex, uninterpretable models in industrial applications. Keywords: Industrial Data Classification, Linear Neural Network Architecture, Interpretable Machine Learning, Energy Consumption Analysis, Binary Prediction Models, Algorithm Optimization.

مهدي خاشعي آشياني

صبا صارمي نيا , حسين خسروشاهي