Source-linked AI summary
AI-based Modeling and Data-driven Evaluation for Smart Manufacturing Processes
Mohammadhossein Ghahramani, Yan Qiao, MengChu Zhou, Adrian OHagan, James Sweeney
TL;DR
Semiconductor manufacturing generates complex, high-dimensional, noisy, and imbalanced data, creating a need for effective feature selection and predictive process analysis. The paper combines preprocessing, a genetic algorithm, and an artificial neural network in a dynamic feature-selection pipeline. The proposed model is reported to achieve accuracy above 90% and to outperform conventional methods.
Problem
Traditional semiconductor fault-detection systems reduce sensory signals to single values, while manufacturing data can be noisy, incomplete, heterogeneous, imbalanced, and high-dimensional.
Method
The paper preprocesses manufacturing data and integrates a binary genetic algorithm with an artificial neural network for dynamic feature selection and classification.
Results
Over 90% accuracy was reported for the proposed model, which was superior to conventional methods in the experimental comparison.
Takeaways & Limitations
The approach provides selected manufacturing features for pattern distinction, predictive modeling, and fault diagnosis in semiconductor processes.
Abstract
from arXiv · showhide
Smart Manufacturing refers to optimization techniques that are implemented in production operations by utilizing advanced analytics approaches. With the widespread increase in deploying Industrial Internet of Things (IIoT) sensors in manufacturing processes, there is a progressive need for optimal and effective approaches to data management. Embracing Machine Learning and Artificial Intelligence to take advantage of manufacturing data can lead to efficient and intelligent automation. In this paper, we conduct a comprehensive analysis based on Evolutionary Computing and Deep Learning algorithms toward making semiconductor manufacturing smart. We propose a dynamic algorithm for gaining useful insights about semiconductor manufacturing processes and to address various challenges. We elaborate on the utilization of a Genetic Algorithm and Neural Network to propose an intelligent feature selection algorithm. Our objective is to provide an advanced solution for controlling manufacturing processes and to gain perspective on various dimensions that enable manufacturers to access effective predictive technologies.
I. INTRODUCTION
Smart manufacturing applies data-driven analytics, machine learning, and artificial intelligence to improve semiconductor production. The paper targets limitations in conventional fault diagnosis by proposing an integrated approach for feature selection, classification, and process insight.
- Smart manufacturing context: Smart Manufacturing uses IoT devices, monitoring sensors, and computational intelligence to transform manufacturing data into real-time insights.These approaches are intended to support process control, productivity improvement, and cost reduction.
- Motivation: Semiconductor production requires identifying influential process factors because failures can produce faulty products and increase manufacturing risks.The approach is motivated by cost-effective, sustainable production and improved product quality.
- Problem: Traditional fault-detection systems reduce sensory signals to single values, making important production features difficult to determine and potentially losing sensory information.Semiconductor data may also contain noise, outliers, missing values, and heterogeneous structures.
- Research objectives: The paper asks whether combining evolutionary algorithms, ANN, ML, and AI can model nonlinearities, support flexible manufacturing, and outperform traditional methods.The proposed direction is a dynamic algorithm for feature extraction and selection in semiconductor manufacturing.
- Reported contribution: The integrated solution selects manufacturing features and uses classification to improve system accuracy and performance, with reported usefulness for fault detection without specialized knowledge.The authors also identify imbalanced data and optimization exploration–exploitation as implementation challenges.
Research Methodology and Contributions
The paper develops a dynamic semiconductor-manufacturing feature-selection and classification pipeline that addresses noisy, imbalanced, and high-dimensional data. It combines preprocessing, GA-based search, and ANN evaluation to identify relevant features and support fault diagnosis.
- Research Methodology and Contributions: Semiconductor manufacturing is complex and costly, so extracting relevant features supports process monitoring and quicker identification of defect root causes.The intended practical outcomes include cost reduction and productivity improvement.
- Research Methodology and Contributions: The pipeline addresses imbalanced defective cases by synthetic minority oversampling, assigns higher misclassification cost to defects, and removes outliers using confidence intervals.These preparation steps transform raw data into features suitable for distinguishing patterns.
- Research Methodology and Contributions: A dynamic feature-selection model integrates a binary GA with an ANN to determine the number and cost of selected features for predictive modeling.The ANN is embedded in the feature-selection procedure and evaluates candidate feature sets.
- Research Methodology and Contributions: GA operations use parent selection, crossover, mutation, and population generation while combining crossover operations to balance exploration and exploitation and reduce premature convergence.The search treats selected features as chromosomes and seeks cost-efficient candidates.
- Related work and gap: Prior manufacturing studies commonly use PCA and KNN, but PCA can obscure original-feature interpretation and nonlinear patterns, while KNN degrades with high-dimensional data.The paper therefore motivates evolutionary global search and ANN-based nonlinear analysis.
III. DATA PREPROCESSING
The preprocessing pipeline represents semiconductor observations with feature vectors and binary production-status labels, then uses Mahalanobis distance to identify outliers.
- Data representation: The SECOM data are represented by an m-feature set F and binary labels L={Failure, Success}.Each observation Xi is an m-tuple containing all features, and yi records its production-status label.
- Data representation: Each observation Xi contains m features, with n observations forming the data matrix.
- Outlier detection: Mahalanobis distance measures each observation’s distance from the mean using the inverse variance-covariance matrix.
- Outlier detection: An observation is treated as an outlier and eliminated when its distance exceeds the selected χ2-distribution quantile, such as the 97.5% quantile.
B. Handling an Imbalanced Data set
The dataset has substantially fewer Failure than Success cases, creating an imbalanced classification problem addressed through density-based synthetic oversampling.
- Class imbalance: 104 Failure cases are relatively rare compared with the Success class, so the two classes do not occupy equal portions of the dataset.
- Oversampling: The study uses density-based SMOTE to generate synthetic Failure instances and make the class distribution more balanced.This approach is selected to address over-generalization concerns associated with nearest-neighbor oversampling methods.
- Oversampling: Oversampling methods create synthetic minority samples based on neighboring observations, but conventional approaches can be negatively affected by over-generalization.
C. Feature Selection
The feature-selection model uses binary genetic search to explore subsets of nearly 600 features, while an ANN evaluates their costs and genetic operators produce successive populations.
- C. Feature Selection: High-dimensional data can cause overfitting, motivating feature selection that chooses a relevant, nonredundant subset rather than mapping features into a new space.
- C. Feature Selection: GA encodes each candidate subset as a chromosome of 0s and 1s indicating whether each feature is excluded or selected.
- C. Feature Selection: The search balances exploration through crossover with exploitation through mutation while retaining lower-cost individuals.
- IV. FEATURE SELECTION MODEL: The model initializes a population of m-dimensional chromosomes and evaluates each individual’s cost with an ANN.
- IV. FEATURE SELECTION MODEL: The feature-selection model is presented as an integrated Artificial Neural Network and Genetic Algorithm approach.
- IV. FEATURE SELECTION MODEL: Parents are selected probabilistically according to costs, then crossover combines their chromosomes to create offspring for the next generation.The reproduction phase includes roulette-wheel sampling and can use single-point, double-point, or uniform crossover.
A. Cost Function and MLP
The cost function evaluates candidate feature subsets by combining prediction error with a penalty for subset size, framing selection as a multi-objective optimization.
- A. Cost Function and MLP: The objective searches for a subset X′ with k<m features that minimizes mean squared error while reducing dimensionality.
- A. Cost Function and MLP: Feature selection simultaneously seeks discriminative features and a small subset, making the task multi-objective.
- A. Cost Function and MLP: The weighted objective function uses Ω as a cost parameter for selecting features, trading off relevance against redundancy.Ω=0 selects all features, whereas a large Ω results in no feature being selected.
- A. Cost Function and MLP: GA receives the Feature-Selection-Cost J, and the ANN calculates the cost values used to minimize the objective function.
V. ANALYSIS PROCEDURES
The analysis procedure selects features dynamically, evaluates subsets with a neural-network cost function, and compares multiple classifiers using accuracy and ROC-based predictive performance.
- Feature selection addresses irrelevant variables that can deteriorate algorithm performance, while conventional approaches include filter, wrapper, embedded, and hybrid techniques.
- The proposed method combines an adaptive dynamic genetic algorithm with a neural network to evaluate feature subsets through an optimized cost function.The initial population reflects data volume, feature count, and sample count; neuron numbers are chosen by trial and error.
- The experiment tested Gaussian SVM, Random Forest, Linear Discriminant, k-NN, and SVM with an RBF kernel using classification accuracy and ROC curves.ROC curves assess predictive performance, while accuracy measures the percentage of correctly predicted classes.
- Cost-function behavior is examined against the number of function evaluations in Fig. 4.
- Gaussian SVM was selected as the classification model, and the proposed feature-selection approach was compared with other methods.The comparison included statistical results such as the percentage of correct predictions.
C. Conventional Methods
The study compares the proposed GA-ANN feature-selection model with conventional dimensionality-reduction and feature-selection methods, reporting higher performance for the proposed approach.
- C. Conventional Methods: The comparison includes PCA, FWE, FDR, SFS, SBS, FFS, CFS, Lasso regression, and ensemble methods for feature extraction.These methods represent conventional approaches used to reduce data dimensionality.
- C. Conventional Methods: The proposed model is reported as superior to the conventional methods in the experimental results.
- C. Conventional Methods: Over 90% accuracy was obtained with the proposed model.An ROC comparison with two traditional techniques is also presented in Fig. 7.
- C. Conventional Methods: Manufacturing Intelligence links business and manufacturing models through sensor data, computing, advanced analytics, and Internet connectivity.
- C. Conventional Methods: The proposed dynamic GA-ANN feature-selection method targets nonlinear relationships and complex patterns that PCA-based approaches do not recognize.The authors compare it with traditional approaches and identify other evolutionary algorithms and multiobjective methods as future comparisons.
APPENDIX A PSEUDO-CODE FOR THE FEATURE SELECTION MODEL
The appendix highlights that genetic-algorithm performance depends on parameter settings and reports their effects in Table III.
- APPENDIX A PSEUDO-CODE FOR THE FEATURE SELECTION MODEL: Genetic-algorithm performance depends on its parameter settings, whose impacts are examined in Table III.
Algorithm 1: Pseudo-code for the feature selection model
The pseudo-code represents feature selection with a binary genetic algorithm that initializes populations, selects parents, applies crossover and mutation, evaluates costs, and returns selected features.
- Algorithm 1: Pseudo-code for the feature selection model: The algorithm takes the feature set and cost function as input and returns selected features represented by a binary vector.
- Algorithm 1: Pseudo-code for the feature selection model: Population initialization precedes parent selection using a Roulette Wheel method based on measured probabilities.
- Algorithm 1: Pseudo-code for the feature selection model: The cost function evaluates offspring produced from parent positions, while mutation is applied during the genetic-algorithm cycle.
- Algorithm 1: Pseudo-code for the feature selection model: The best solution and its cost are taken from the first chromosome before the algorithm returns the selected features.
- Algorithm 1: Pseudo-code for the feature selection model: Crossover randomly uses single-point, multi-point, or uniform crossover according to defined probabilities and returns two offspring.