SECOM Semiconductor Manufacturing process dataset with an initial 591 features were reduced to 63 and

processed by PCA which led to the Support Vector Classifier (SVC) producing the most accurate results at

98.6% while maintaining robust calibration. The visualization includes both a correlation heatmap showing

related features and pie charts showing class distribution before and after data balancing techniques are

applied. This research presents implications for predictive maintenance within semiconductor fabs together

with future work recommendations.

Oversampling, Feature Reduction, SVC

The predictive maintenance approach in semiconductor manufacturing proves essential for decreasing

correcting the calibration reissues and misclassification risks found in imbalanced datasets. The article

provides an all-encompassing approach which optimizes data cleaning and feature reduction and oversampling

and hyper parameters adaptation through specialized techniques for the SECOM semiconductor manufacturing

dataset with its imbalanced characteristics.

In recent times machine learning prediction applications has attracted in a variety of industrial products. In the

field of semiconductor manufacturing, early work by Susto et al. (2015) introduced multiple classifiers for

managing high dimensional data within semiconductor manufacturing. This ground breaking study did not

explicitly address the challenges caused by imbalanced target classes distribution, common problem in real

world fabrication environments. Other researchers have directed its investigations towards discovering sensor

do not address the effect of oversampling on model calibration. Chawla et al. (2002) and He et al. (2008) [4]

sample distributions yet results in incorrect probability estimation according to van den Goorbergh et al.

(2022). The research on anomaly detection and inter-sensor transfer learning conducted by Yan et al. (2024)

offers useful information regarding model performance in industrial settings while primarily examining

anomaly detection features instead of complete predictive maintenance solutions. The work presented by

Wang et al. (2022) together with Lee et al. (2014) showed how statistical filtering and principal component

analysis (PCA) succeed in reducing manufacturing data dimensions while maintaining their vital variation.

Research is lacking to determine how algorithms that reduce features work together with oversampling

techniques to improve both model performance and calibration when used in semiconductor manufacturing

operations. Furthermore, several studies have explored how machine learning model calibration performs

following the implementation of oversampling techniques. Support vector machines operate well on

imbalanced data records according to Farrag et al. (2024), yet probabilistic output calibration needs precise

by reducing production interruptions and increasing operational efficiency according to the research findings.

Farrag et al. (2024) conducted a study which introduces a predictive framework for minority classes within

semiconductor manufacturing data that manages noise together with class imbalance concerns. The developed

model exhibits positive results through its 0.95 AUC value and its 0.66 precision and 0.96 recall quality

metrics which provide details about future maintenance operations and product quality levels. The study by El

Mourabit et al. (2020) offers a machine learning predictive system for semiconductor failures which includes

data preprocessing and feature selection to enhance prediction accuracy. Guo et al. (2024) examines different

detection methods alongside classification methods and location methods in transmission lines and distribution

systems while explaining machine learning applications for fault diagnosis. Salem (2018) investigates fault

diagnosis systems in semiconductor production which face challenges due to misbalanced and partial data. The

study of Random Forest algorithms to provide insights about their usage both for predictive maintenance and

imbalanced data applications in semiconductor manufacturing. while previous work has provided valuable

insights into individual components, oversampling, feature reduction, and machine learning algorithms, few

studies have offered an end-to-end framework that simultaneously addresses the imbalance, dimensionality,

and calibration challenges in semiconductor predictive maintenance. This research seeks to fill that gap by

combining robust preprocessing with advanced oversampling and model tuning, thereby achieving superior

predictive performance and calibration.

The SECOM (Semiconductor Manufacturing Process) dataset contains process control data from a semi-

conductor enterprise into a high-dimensional time-series format. The dataset contains 591 numeric features

along with a timestamp field while its target variable represents failed products with label -1 and passed

along with streamlining computational complexities in later modeling stages. Next, missing data treatment was

conducted. Features with more than 20% missing values were eliminated, as excessive imputation could

introduce noise and bias. For the remaining features with occasional missing entries, we performed mean

imputation based on the missing at random (MAR) assumption. This step preserves data volume while

ensuring statistical integrity. The reduced set of 63 features remained dimensionally high and potentially

displayed multicollinearity. Next Principal Component Analysis (PCA) underwent implementation to achieve

a new feature space transformation. PCA achieves two objectives: first it simplifies multidimensional data by

maintaining most data variations and secondly it addresses multicollinearity issues that decrease model

interpretability and performance. The analysis of principal components resulted in 32 features which

maintained greater than 90% of all original data variations. The data transformations through this method

7% of the data which leads to standard classification models strongly favoring the pass class predictions.

Standard classification models can perform well in terms of accuracy when they predict every instance to be

non-faulty but they miss all real failures that reduces predictive maintenance effectiveness. For balancing the

imbalanced data we used the common oversampling techniques SMOTE (Synthetic Minority Over-sampling

Technique) and ADASYN (Adaptive Synthetic Sampling). The SMOTE approach creates synthetic samples

through the connection points of minority class examples with their nearest neighbors on the data set

dimension. The methodology considers synthetic data more beneficial than basic duplicate data production

because standard duplication leads to overfitting. By applying ADASYN on SMOTE technology more

emphasis is placed on the classification-intensive points near the decision border to improve training accuracy.

The training process included class weighting as a means to decrease bias that results from class imbalance.

Once the data was appropriately preprocessed and rebalanced, we conducted an extensive comparative analysis

of several supervised learning algorithms. Which includes both traditional classification models and more

advanced ensemble-based methods. Specifically, we trained and evaluated Logistic Regression, K-Nearest

Neighbors (KNN), Gaussian Naive Bayes, Decision Trees, Bagging Classifier, AdaBoost, Gradient Boosting,

Random Forest, and Support Vector Classifier (SVC). We evaluated these models performance in regard to

working with high-dimensional data streams and imbalanced distributions while performing predictive

maintenance operations. Each model required optimization for its performance which was achieved through

the use of Bayesian Optimization for hyperparameter tuning. Bayesian Optimization proved to be a superior

choice compared to traditional grid search and random search methods because it excels in finding optimal

solutions efficiently for complex search domains. Bayesian Optimization estimates probabilistic behavior of

the objective function thus carefully determining which set of hyperparameters requires testing next. The

The heatmap (Figure 1) visualization depicts the interrelations of selected features. A solid positive linear

relationship exists between attributes '180' and '316' and a profound negative relationship appears between

'122' and '130' while '59', '103', '210', '348' strongly affect the target variable. The visual presentation identifies

information from the dataset. The effective balancing of the dataset requires the use of oversampling methods

SMOTE and ADASYN because target distribution changes can be seen in the produced pie charts.

In this article, we have developed a framework for semiconductor manufacturing predictive maintenance

which addresses problems caused by unbalanced data distribution. Our method establishes significant model

performance improvement thorough data cleaning procedures with advanced feature reduction techniques and

time sensor implementation tasks for maintenance scheduling dynamic industrial operational settings.