Dimensionality Reduction and Subspace Clustering in Mixed Reality for Condition Monitoring of High-Dimensional Production Data

Abstract

Visual analytics are becoming increasingly important in the light of big data and related scenarios. Along this trend, the field of immersive analytics has been variously furthered as it is able to provide sophisticated visual data analytics on one hand, while preserving user-friendliness on the other. Furthermore, recent hardware developments such as smart glasses, as well as achievements in virtual-reality applications, have fanned immersive analytic solutions. Notably, such solutions can be very effective when they are applied to high-dimensional datasets. Taking this advantage into account, the work at hand applies immersive analytics to a high-dimensional production dataset to improve the digital support of daily work tasks. More specifically, a mixed-reality implementation is presented that will support manufacturers as well as data scientists to comprehensively analyze machine data. As a particular goal, the prototype will simplify the analysis of manufacturing data through the usage of dimensionality reduction effects. Therefore, five aspects are mainly reported in this paper. First, it is shown how dimensionality reduction effects can be represented by clusters. Second, it is presented how the resulting information loss of the reduction is addressed. Third, the graphical interface of the developed prototype is illustrated as it provides (1) a correlation coefficient graph, (2) a plot for the information loss, and (3) a 3D particle system. In addition, an implemented voice recognition feature of the prototype is shown, which was considered to be being promising to select or deselect data variables users are interested in when analyzing the data. Fourth, based on a machine learning library, it is shown how the prototype reduces computational resources using smart glasses. The main idea is based on a recommendation approach as well as the use of subspace clustering. Fifth, results from a practical setting are presented, in which the prototype was shown to domain experts. The latter reported that such a tool is actually helpful to analyze machine data daily. Moreover, it was reported that such a system can be used to educate machine operators more properly. As a general outcome of this work, the presented approach may constitute a helpful solution for the industry as well as other domains such as medicine.

Keywords: covariance graph; dimensionality reduction; immersive analytics; mixed reality; subspace clustering.

Figure 1

The basic idea of this work is to provide visual, algorithm supported insights into large datasets. As a result, these datasets can be visually inspected to enhance or replace current dashboards features.

Figure 2

Reality–Virtuality continuum at a glance. The overlap of virtuality and reality increases towards the right-hand side of the figure, the complexity of the necessary devices, respectively.

Figure 3

The additive manufacturing process includes several phases in which objects are produced from powder in a layer-wise procedure.

Figure 4

Proposed Approach at a Glance. The 3D plot (top left) can be used to display variables and detect clusters. A subspace clustering algorithm (top right) reveals automatically detected clusters. With the use of the PCA plot (bottom left) and additional components (bottom right), dimensionality reduction can be applied and analyzed.

Figure 5

CLIQUE Visual Explanation. (a) The user defines his grid resolution and (b,c) use case dependent thresholds define dense units and cluster definitions.

Figure 6

Overall workflow of the implemented prototype and the CLIQUE algorithm.

Figure 7

(a) Particle-based visualization of additive manufacturing data, where one data point is a print job represented by the variables numberOfLayers, numberOfParts, and numberOfErrors, and (b) sample correlation graph to help the user in selecting variables for the visualization.

Figure 8

The Information Loss Component, explained by (a) the resulting 3D component as a stacked bar, and (b) the composition of the stacked bar, which is generated by the variance of the components.

Figure 9

Backend Strategy for providing the PCA-related features. A RESTful-driven architecture is chosen to offer the possibility of working with the infrastructure as a distributed system.

Figure 10

Calculation times for the conducted PCA. In the worst case, the computation takes around 100 ms to calculate the principal components.

Figure 11

Visualization of two detected clusters (red and yellow) with the predefined subspace clustering grid. Clusters are even detected in overlapping point clouds.