A Deep Fuzzy Semi-supervised Approach to Clustering and Fault Diagnosis of Partially Labeled Semiconductor Manufacturing Data

Abstract

In the modern age of data collection in manufacturing industry, the sheer volume of measurement data collected may prove difficult for domain experts to create fully labeled training datasets for supervised learning artificial intelligence methods. Semi-supervised learning methods are useful in the realistic scenario where engineers may only be able to annotate limited partial subsets, but existing approaches are limited in scalability for high-dimensional and imbalanced datasets. To address these challenges, a novel framework for semisupervised learning is proposed that hybridizes 1) convolutional autoencoder as a deep unsupervised feature learning technique; 2) fault classification using principal component analysis-based anomaly scoring; and 3) fuzzy c-means clustering. Fuzzy c-means allows for transparency in the degree of membership to each cluster via the fuzzy partition matrix, enabling an adjustable, explainable, and computationally efficient approach for separating normal and faulty clusters in the same space. This also allows industry experts to review borderline cases, creating a support system to curtail costly labeling expenses. The approach is applied to real, high-dimensional data from a modern semiconductor manufacturing application in which fewer than 1000 out of over 59,000 samples are labeled, achieving AUC scores of over 0.94 for classifying the two labeled fault types as well as successful fuzzy clustering. These results show promise for deep, fuzzy semi-supervised applications to improve decision-making in manufacturing operations and other engineering disciplines.

To Trust or Not: Towards Efficient Uncertainty Quantification for Stochastic Shapley Explanations

“…The neural network is built using Flux, a deep learning library supported by the Julia programming language (Innes, 2018). For more information for the developed predictive model and its performance, we refer to past work by Cohen et al (2023). To obtain stochastic Shapley explanations, we utilize ShapML.jl, a relatively efficient Julia implementation of the IME algorithm that has successfully been benchmarked against other state-of-the-art Shapley estimation algorithms such as FastSHAP (Redell, 2020).…”

“…The dataset contains labeled failure mode information pertaining to five rotating components: fan, high-pressure compressor (HPC), low-pressure compressor (LPC), high-pressure turbine (HPT), and low-pressure turbine (LPT). This paper builds on previous XAI work on this dataset, which used Shapley-based explanations to derive meaningful clusters for the context of predicting future faults and the RUL (Cohen et al, 2023).…”

On the Soundness of XAI in Prognostics and Health Management (PHM)

“…They then applied various XAI methods and different baselines to attribute the network predictions to the input. Cohen et al [16] proposed a new clustering framework that uses Shapley values and is compatible with semi-supervised learning problems. This framework relaxes the strict supervision requirement of current XAI techniques.…”

Explainable artificial intelligence-based evidential inferencing on process faults in plasma etching

“…XAI techniques such as feature importance analysis and model-agnostic methods can be used to identify which features of the data are the most important in determining the output of the artificial intelligence (AI) system and to generate explanations that are easily understood by human users [15]. As XAI has gained prominence, its application in interpreting machine learning models within semiconductor processes has increased [16][17][18]. We have employed XAI algorithms, including permutation importance and SHapley Additive exPlanations (SHAP), to analyze essential variables that contribute to predictions in machine learning-based models that are used for diagnosing semiconductor plasma processes in APC applications, such as VM and FDC [13,[19][20][21].…”