Graphical Event Models (GEMs) can approximate any smooth multivariate temporal point processes and can be used for capturing the dynamics of events occurring in continuous time for applications with event logs like web logs or gene expression data. In this paper, we propose a multi-task transfer learning algorithm for Timescale GEMs (TGEMs): the aim is to learn the set of k models given k corresponding datasets from k distinct but related tasks. The goal of our algorithm is to find the set of models with the maximal posterior probability. The procedure encourages the learned structures to become similar and simultaneously modifies the structures in order to avoid local minima. Our algorithm is inspired from an universal consistent algorithm for TGEM learning that retrieves both qualitative and quantitative dependencies from event logs. We show on a toy example that our algorithm could help to learn related tasks even with limited data.

The objective of Smart Manufacturing is to improve productivity and competitiveness in industry, based on in-process data. It requires reliable, explainable and understandable models such as Bayesian networks for performing tasks like condition prediction. In this context, a Bayesian network can be classically learned in a supervised, unsupervised way or a semi-supervised way. Here, we are interested in how to perform the learning when the ground truth isn't included in the learning data but is observable indirectly in another related dataset. This paper introduces a fully unsupervised variation of co-training that allows to include this second dataset, with two learning strategies (split and recursive). In our experiments, we propose one simple probabilistic graphical model used for predicting the state of a machine tool from results given by several sensors, and illustrate our unsupervised cotraining strategies first with benchmarks available from the UCI repository, for which 4 out of 5 datasets have best results with the recursive strategy. Finally, the recursive strategy was validated by McNemar's test as being the best strategy on a real industrial dataset.

The objective of Smart Manufacturing is to improve productivity and competitiveness in industry, based on in-process data. Indeed, failures can stop the production for a couple of days and generate costs of non-quality. Failures in industry can either damage the machine or the product being produced. In both cases, the earlier the failure is detected, the lower the impact on production. Thus, monitoring both the process and the machine condition is interesting, due to their potential interactions. Besides, the diagnosis of the nature of the incident is also important, in order to react adequately as fast as possible. It requires reliable, explainable and understandable models such as Bayesian networks for performing tasks like condition prediction. Bayesian networks can be learned with incomplete data and in a supervised or unsupervised way, which is very useful because the collect of labelled data is costly and sometimes impossible, especially in industry where problems are, moreover, very rare. In this paper, we propose a generic architecture based on two Bayesian networks and a collaborative learning strategy that improves the condition monitoring of rotating machines in unsupervised context by using information gathered from process monitoring.