Error Estimation in Chemical Process

Equipment
(Training a Model for CSTR Data)

Submitted to:-

Prof. Kailash Singh

Group Members:-

Yogita Kumari(2023uch1286)

Abhinav Thapliyal(2023uch1219)

Yashkanwal Kaur(2023uch1566)

Tanisha Singh(2023uch1487)
Date of Submission: -15November2025
Abstract
This project focuses on error estimation in chemical process equipment using data obtained from a
Continuous Stirred Tank Reactor (CSTR). A machine learning model was trained on experimental
and simulated CSTR datasets to identify, quantify, and predict deviations in measured process
variables. The aim was to evaluate the model's capability to detect systematic and random errors
commonly encountered in chemical engineering operations. The study demonstrates how data-
driven tools can assist chemical engineers in improving equipment reliability, diagnostic capability,
and process performance.

1. Problem Statement
Chemical process equipment such as reactors, heat exchangers, and distillation columns often
operate under uncertainties due to measurement noise, sensor drift, operational fluctuations, or
modeling assumptions. These errors degrade process efficiency and product quality. In this project,
the focus is on estimating and detecting errors in a Continuous Stirred Tank Reactor (CSTR) using
machine learning techniques trained on process data. The goal is to develop a reliable
computational method for identifying error patterns and quantifying deviations from expected
performance.

2. Introduction
Error estimation plays a crucial role in chemical engineering design and process control. Accurate
error quantification helps engineers ensure equipment performs as intended and maintains
product specifications. CSTRs are widely used in chemical industries due to their simplicity and
ability to operate under steady-state conditions. However, process variables such as temperature,
concentration, flow rate, and reaction rate are prone to errors caused by external disturbances or
imperfect sensors. With the rise of AI/ML tools, data-driven methods have become effective for
predicting, analyzing, and minimizing these errors. This report presents an approach to train a
model on CSTR data for error estimation.

3. Objectives
• To understand sources of error in chemical process equipment, specifically CSTRs.

• To collect and preprocess CSTR operational data for modeling.

• To train a machine learning model capable of detecting and quantifying measurement errors.

• To evaluate model performance using statistical metrics.

• To demonstrate how error estimation improves reliability of chemical process equipment.

4. Theoretical Background
A Continuous Stirred Tank Reactor (CSTR) is a well-mixed reactor where reactants are fed
continuously and products are withdrawn at steady state. The dynamic behavior of a CSTR is
governed by material and energy balance equations. Errors in these variables affect reactor
performance and need to be quantified. Error estimation involves statistical and computational
analysis to identify deviations in measurements.

Machine learning models such as Linear Regression, Random Forest, or Neural Networks can be
trained on clean vs. noisy datasets to learn patterns of measurement error. Feature engineering,
normalization, and residual analysis play an important role in understanding error characteristics.

Classification Models for Fault Detection

The fault detection task was addressed using three powerful machine learning models: the Random
Forest (RF) Classifier, Extreme Gradient Boosting (XGBoost) Classifier, and a Deep Neural
Network (DNN). These approaches represent different philosophies of ensemble learning and deep
learning, applied to analyze the complex time-series data from the Continuous Stirred-Tank Reactor
(CSTR).

Random Forest Classifier: The Wisdom of Crowds

The Random Forest model is an ensemble learning method that functions like a panel of independent
experts. Instead of relying on a single, complex decision tree that is prone to noise and overfitting, it
builds a "forest" of many simple decision trees.

 How it Works (Bagging): Each tree in the forest is trained on a random subset of the training
data (a process called bootstrap aggregating or bagging). Crucially, when deciding how to split
the data at any point in a tree, it only considers a random subset of the features.
 Final Prediction: For classification, the final result is determined by majority vote among all
the trees. This aggregation reduces the variance (the sensitivity to small changes in the training
data) and makes the model highly robust and less likely to overfit.
 Configuration in the Study: The model was configured with 300 estimators (300 trees). Its
inherent ability to measure Feature Importance was leveraged to identify which specific process
variables and time steps were most critical for fault diagnosis.

Extreme Gradient Boosting (XGBoost): The Sequential Learner

XGBoost is an advanced and highly optimized form of gradient boosting, which is another type of
ensemble learning. Unlike Random Forest, where trees are built independently, boosting builds trees
sequentially, with each new tree focused entirely on correcting the mistakes of the one before it.

 How it Works (Boosting): It starts with a simple initial prediction. Subsequent decision trees
(called weak learners) are added to the model to predict the residuals (the errors) of the previous
 step. It uses a gradient descent approach to efficiently minimize the error (loss function) as new
trees are added.
 Extreme Optimization: XGBoost is famous for its speed and performance due to optimizations
like parallel processing and built-in regularization (L1 and L2 penalties). These regularization
techniques prevent the model from becoming too complex and over-specialized on the training
data.
 Configuration in the Study: The XGBoost model was trained with 200 estimators and a
learning rate of 0.05. This low learning rate forces the model to learn gradually, ensuring precise
and accurate final predictions.

Deep Neural Network (DNN): Learning Complex Patterns

A Deep Neural Network is an Artificial Neural Network (ANN) characterized by having multiple
hidden layers stacked between the input and output layers, which allows it to learn increasingly complex,
hierarchical representations of the input data.

 Architecture: The implemented DNN used a Sequential model (a simple, layer-by-layer

structure). Data flows forward through dense layers (where every neuron connects to every
neuron in the next layer), each using the ReLU (Rectified Linear Unit) activation function to
introduce non-linearity. The final layer uses Softmax activation to output probabilities for each of
the possible fault classes.
 Regularization (Dropout): To combat overfitting, Dropout layers were strategically placed.
Dropout randomly ignores a fraction of neurons (0.3 or 30%, and 0.2 or 20% in this model)
during each training cycle, preventing co-dependence between neurons and forcing the network
to learn more robust features.
 Training Method: The network was trained using the Adam optimizer to adjust the internal
weights, minimizing the sparse categorical cross-entropy loss function over 25 epochs (full
passes through the training data).

5. Results and Discussion

The trained model demonstrated strong capability in detecting abnormal deviations in the CSTR
dataset. Metrics such as MAE, RMSE, and R² were used to evaluate prediction accuracy. The model
successfully distinguished between normal operational variations and true measurement errors.
Visualization of predicted vs. actual values highlighted regions of high uncertainty, indicating
potential sensor drift or disturbances.

The results show that data-driven approaches can significantly enhance real-time fault detection,
making them valuable tools for chemical engineers.

6. Conclusions
This project demonstrated an effective method for estimating and detecting errors in CSTR
operations using machine learning. The approach provides chemical engineers with a powerful tool
for monitoring process performance, identifying fault patterns, and improving equipment
reliability. Future work may involve integrating the model into a real-time monitoring system and
expanding the dataset with more operational scenarios.

7. References
Fogler, H. S. (2016). Elements of Chemical Reaction Engineering. Prentice Hall.

Seborg, D. E., Edgar, T. F., & Mellichamp, D. A. (2010). Process Dynamics and Control.

Bishop, C. M. (2006). Pattern Recognition and Machine Learning. Springer.

For dataset:- [Link]

domain-adaptation

8. Appendix: Code

Code and output Pdf Attachment-> code [Link]