Design Performance Analysis of a Self-Organizing Map for Statistical Monitoring of Distribution-free Data Streams

Abstract

In industrial applications, the continuously growing development of multi-sensor approaches, together with the trend of creating data-rich environments, are straining the effectiveness of the traditional Statistical Process Control (SPC) tools. Industrial data streams frequently violate the statistical assumptions on which SPC tools are based, presenting non-normal or even mixture distributions, strong autocorrelation and complex noise patterns.\ud \ud To tackle these challenges, novel nonparametric approaches are required. Machine learning techniques are suitable to deal with distributional assumption violations and to cope with complex data patterns. Recent studies showed that those methods can be used in quality control problems by exploiting only in-control data for training (such a learning paradigm is also known as “one-class-classification”).\ud \ud In recent studies, the use of distribution-free multivariate SPC methods was proposed, based on unsupervised statistical learning tools, pointing out the difficulty of defining suitable control regions for non-normal data. In this paper, a Self-Organizing Map (SOM) based monitoring approach is presented. The SOM is an automatic data-analysis method, widely applied in recent works to clustering and data exploration problems. A very interesting feature of this method consists of its capability of providing a computationally efficient way to estimate a data-adaptive control region, even in the presence of high dimensional problems. Nevertheless, very few authors adopted the SOM in an SPC monitoring strategy. The aim of this work is to exploit the SOM network architecture, and proposing a network design approach that suites the SPC needs. A comparison study is presented, in which the process monitoring performances are compared against literature benchmark methods. The comparison framework is based on both simulated data and real data from a roll grinding application

Continuous detection of concept drift in industrial cyber-physical systems using closed loop incremental machine learning

“…The proposed algorithm consists Fig. 1 Extended closed loop framework for concept drift detection from unlabeled data streams, based on [5] of three primary functions, (1) online learning, (2) incremental and decremental learning and (3) concept drift detection (Fig. 2), each function is discussed below.…”

“…Detection of concept drifts in CPS decreases the negative impact of a compounding error and enables cost-effective predictive maintenance. However, data streams in industrial CPS are composed of unlabeled target variables that do not fit into predefined classes [4,5]. Ensemble learning algorithms that integrate multiple supervised algorithms find it infeasible and impractical to detect concept drifts in this environment.…”