Intelligent Industrial Data Analysis using Neural Networks and ANFIS in MATLAB

Overview

This project presents a comparative study of intelligent data analysis methods for industrial process monitoring data related to low-temperature coke gas separation.

Several machine learning and neuro-fuzzy approaches were implemented and compared in MATLAB, including:

• fuzzy clustering (FCM and Subtractive Clustering),
• feedforward neural networks,
• ANFIS neuro-fuzzy systems,
• Neural Network Fitting App,
• hybrid clustered neural network models.

The goal of the project was to investigate nonlinear approximation of industrial process parameters and compare the effectiveness of different intelligent modeling approaches.

---

Dataset

The dataset contains industrial process monitoring data from a low-temperature coke gas separation system.

Input Features

• Temperature
• Valve opening percentage
• Coke gas flow
• Nitrogen flow

Target Variable

• Ethylene fraction concentration

Two independent datasets were created:

• coke_gas_train.csv
• coke_gas_test.csv

The datasets were generated using non-overlapping sampling strategies to ensure reproducible external testing across all experiments.

---

Implemented Methods

1. Data Preprocessing

• statistical analysis,
• polynomial approximation,
• spline interpolation,
• exploratory data analysis.

2. Fuzzy Clustering

• Fuzzy C-Means (FCM),
• Subtractive Clustering,
• membership matrix analysis,
• cluster center comparison.

3. Feedforward Neural Networks

• multilayer perceptrons,
• Levenberg–Marquardt training,
• train/test evaluation,
• architecture comparison.

4. ANFIS Neuro-Fuzzy Systems

• Sugeno FIS,
• Grid Partitioning,
• Subtractive Clustering initialization,
• neuro-fuzzy approximation.

5. Hybrid Clustered Neural Networks

• FCM-based data segmentation,
• separate neural network per cluster,
• local approximation of operating regimes.

6. MATLAB Neural Network Fitting App

• automatic train/validation/test split,
• regression analysis,
• validation performance analysis.

---

Repository Structure

data/
results/
scripts/
README.md

Main folders

results/
├── plots/
├── screenshots/
├── metrics/
└── models/

scripts/
├── preprocessing/
├── neural_networks/
├── hybrid_models/
└── anfis/

---

Key Results

Best Classical Neural Network

Architecture	Goal	Epochs	Performance
5-40	0.01	12	0.00207

ANFIS Comparison

Method	Test Error
Subtractive Clustering	2.2645
Grid Partitioning	4.6195

Observations

• classical feedforward neural networks achieved the lowest approximation error,
• ANFIS models provided interpretable fuzzy-rule-based approximation,
• clustered neural networks demonstrated regime-based local modeling,
• early overfitting was observed in some automatic fitting experiments.

---

Example Results

Neural Network Test Prediction

Clustered Neural Networks

Neural Network Fitting Performance

---

Technologies

• MATLAB
• Neural Network Toolbox
• Fuzzy Logic Toolbox
• ANFIS
• FCM Clustering
• Subtractive Clustering

---