Industrial datasets can contain a large volume of routine observations alongside only a handful of examples of a particular fault or defect. This imbalance complicates unsupervised analysis: a clustering model may organize the dominant patterns reasonably well while overlooking smaller, operationally important groups. Deep clustering for imbalanced data aims to learn useful representations and cluster assignments together, without relying on class labels that may not be available in advance.
Regularized deep embedded clustering (RDEC) addresses this challenge by combining deep embedded clustering (DEC) with virtual adversarial training (VAT). Its reported application to semiconductor wafer defect maps illustrates why the approach may be relevant to manufacturing analytics, where rare but distinctive patterns can matter.
Why class imbalance challenges clustering
Clustering groups observations according to similarity, but the result depends on how the data is represented and how clusters are formed. In deep clustering, a neural network maps each input into a lower-dimensional latent space, where the model learns cluster assignments. This can make complex inputs easier to organize than clustering directly in the original data space.
DEC uses distances between latent representations and cluster centroids to estimate soft assignments: each observation has a probability of belonging to each cluster. It then builds a target distribution that encourages confident assignments and trains the model to bring its predictions closer to that target.
That process can be sensitive to how centroids are initialized. If a minority group is not well represented by the initial centroids, its observations may be assigned poorly. DEC can also produce problematic assignments for marginal observations—points that are far from all centroids. Its target-distribution calculation can favor a smaller cluster for these uncertain points, even when that does not reflect the underlying data structure. Both effects can become significant when cluster sizes differ substantially.
How RDEC uses virtual adversarial training
RDEC retains DEC’s latent-space clustering objective and adds a regularization term based on VAT. VAT creates a small, deliberately chosen perturbation to an input and encourages the model to produce a similar assignment distribution for the original and perturbed data. The perturbation is constrained so that it represents a nearby variation rather than an entirely different observation.
In practical terms, the added objective encourages local consistency: observations that are close in the original input space should have similar cluster predictions. The model therefore considers not only the relationship between an observation and a centroid, but also how nearby data behave. This can reduce reliance on centroid placement alone and help marginal observations follow the assignments of similar neighbors.
RDEC combines the two objectives during training: DEC’s clustering loss and VAT’s consistency loss. A weighting parameter controls the influence of VAT relative to clustering. The model is first initialized through autoencoder pretraining, then fine-tuned using the clustering and regularization objectives. As with other model-based approaches, the number of clusters is supplied as prior knowledge in the described method.
What the evaluation found
The reported experiments compared RDEC with methods including K-means, autoencoder-based clustering, DEC, and other deep clustering approaches. On the MNIST handwritten-digit benchmark, RDEC achieved 98.41% clustering accuracy. On a highly imbalanced MNIST subset with a minimum retention rate of 0.1, its reported accuracy was 85.45%.
The results also showed a substantial difference on a more severely imbalanced MNIST variant: RDEC achieved 65.40% accuracy, compared with 40.71% for DEC. These figures are specific to the experimental datasets and settings; they should not be treated as a forecast for a plant’s data or a guarantee of performance on other applications.
Application to semiconductor wafer defect maps
The study applied RDEC to wafer maps from semiconductor manufacturing. Each map was converted into a binary vector representing defective and non-defective chips. The evaluation dataset contained approximately 13,000 maps, visually labeled into 11 defect-pattern classes. The class distribution was highly uneven: the largest class represented 68.57% of the dataset, while five classes each represented less than 1%.
With the number of clusters set to 11 for evaluation, RDEC achieved 79.86% clustering accuracy. The reported accuracy for DEC was 42.10%. The authors observed that RDEC more effectively grouped maps from the largest class together and collected maps from small classes into clusters. In a manufacturing context, that is relevant because the analytical task may involve finding recurring defect patterns even when examples of those patterns are scarce.
Considerations for industrial deployment
RDEC is an unsupervised learning method, but applying it still involves decisions and preparation. The described wafer-map experiment used visually reviewed labels to evaluate results, even though clustering itself does not require labels for training. It also set the cluster count to the evaluation’s known class count. In an operational setting, the appropriate cluster count may not be known beforehand, so results require interpretation and validation against process knowledge.
Data representation matters as well. In the wafer-map evaluation, maps were converted to binary vectors, and pretraining used a hand-selected collection of maps identified by on-site engineers. This highlights the role of input preparation and initialization in the workflow; the reported results do not establish that the same representation or settings will suit every industrial dataset.
RDEC also adds computation because VAT requires evaluating perturbed inputs during training. The paper describes the VAT component as increasing the computation needed for network updates, while noting that target-distribution updates account for a major part of computation in both DEC and RDEC. Engineers considering the method should therefore assess model-training requirements alongside clustering quality.
Practical implications
For industrial analytics teams, RDEC offers a way to address two weaknesses of centroid-based deep clustering: sensitivity to initial centroid locations and uncertain assignments near cluster margins. Its combination of latent-space clustering and local consistency regularization produced stronger results than DEC in the reported imbalanced-data experiments, including the wafer defect-map application.
The practical takeaway is to treat RDEC as a candidate for datasets where rare patterns are important and labels are limited—not as a substitute for application-specific validation. Assess the data representation, cluster-count assumptions, training cost, and usefulness of the resulting groups. For manufacturing teams investigating sparse defect signatures, that evaluation can determine whether regularized deep clustering provides a clearer view of patterns that conventional clustering may miss.