Evolving environments challenge researchers with non stationary data flows where the concepts-or states-being tracked can change over time. This requires tracking algorithms suited to represent concept evolution and in some cases, e.g. real industrial environments, also suited to represent time dependent features. This paper proposes a unified approach to track evolving environments that uses a two-stages distance-based and density-based clustering algorithm. In this approach data samples are fed as input to the distance based clustering stage in an incremental, online fashion, and they are then clustered to form µ-clusters. The density-based algorithm analyses the micro-clusters to provide the final clusters: thanks to a forgetting process, clusters may emerge, drift, merge, split or disappear, hence following the evolution of the environment. This algorithm has proved to be able to detect high overlapping clusters even in multi-density distributions, making no assumption about cluster convexity. It shows fast response to data streams and good outlier rejection properties.

No abstract

A real challenge for manufacturing industry is to be able to control not only the manufacturing process but also the production quality. Products that are suspected to be faulty are deviated from their nominal path in the production line and inspected more closely. The fact that some products deviate from the nominal path and others fail at some check operations can be used as an indicator of poor product quality. Based on this idea, this paper proposes a method to compute a product quality index or more exactly a penalty index taking into account both product path and production batches. The method relies on categorizing the products according to how they follow the production path and process mining techniques. The originality of the proposed index is to be built from advanced data analysis techniques enhanced by expert know-how. The quality index highlights risk of customer return, which is highly relevant information for the after sales service. The significance of the method is illustrated on a printed circuit board production line using surface mount technology at Vitesco Technologies. Data is collected from the real manufacturing execution system. The results obtained over more than 10000 single electronic boards show that 91.7% of the products are in good compliance with respect to the requirements. For the other products, the method identifies the root causes of poor quality that may call for maintenance or reconfiguration actions.

Testing equipments are a crucial part of production quality control in the automotive industry. Their health needs to be controlled carefully to avoid quality issues and false alarms that reduce production efficiency, potentially leading to huge losses. The main challenge for this control is the large number of features leaning for automated reasoning. A data-based Health Monitoring System could be a solution. In manufacturing industries, a widely accepted index for evaluating process performance is the capability. It combines statistical measures for normal distributions in order to verify the ability of a process to produce an output within the specification limits. In this article we propose  a capability-based prognosis and diagnosis method based on test data. Capability is calculated and compared to a known threshold. If the index value exceeds the threshold, then a diagnosis phase is initiated to find out which parts of the equipment are faulty. Data temporality is also taken into account. Data trends are used for prognosis.Test data are splited into periods. To respect the normality assumption of the capability, it is proposed to use a Gaussian Mixture Model (GMM) classification to extract all normal distributions found in one data period. Two approaches are discussed for selecting the number of clusters used for the classification. The first approach is based on the well-known Bayesian Information Criterion (BIC). The second approach uses a multi-criteria aggregation function learned by using machine learning on a synthetically gene-rated dataset. Some of the criteria used in the aggregation are inference based. Others are classical statistics extracted from the classes obtained by the GMM.For each of these classes the capability index is calculated and used for diagnosis and prognosis purposes. This method is applied on real data from In-Circuit Testing (ICT)  machines for electronic components at a Vitesco factory in France.