The measurements of several sensors of multiple machines are collected in real-time at Alfa Laval’s IIoT platform. It is of great interest to make an accurate forecast for upcoming events, such as machine failure or sensor faults, which are critical enablers for predictive maintenance. However, some machines have little historical data, making forecasting challenging for them. Can we use data from other devices to make a forecast on a newly installed machine or improve the forecasting performance on different machines? This is the central question I try to answer in this research. However, another dimension of complexity makes the problem even more difficult. The devices have some differences. They can vary in version, application, and configuration. In addition, at each instant, the machine can have different modes and user-defined parameters that can be changed anytime by the operator or machine auto-control. All these factors make the forecasting problem extremely challenging, so it does not fit into a general computational framework. I propose a new multi-paradigm ensemble framework, aware of context and domain, that significantly improves the popular ensemble methods and non-ensemble state-of-the-art such as ARIMA and LSTM. The improvement mostly comes from adding a new forecasting paradigm based on multi-domain INN-DTW (one-nearest neighbor with dynamic time warping as a similarity measure) to the ensemble architecture, which creates a surprising harmony with ARIMA. INN-DTW is popular in time series classification, but its application is being investigated for the first time in a heterogeneous ensemble. The other novel component is a meta-selection mechanism for aggregating forecasts, which unexpectedly works better than the widely-used meta-learning approach.

Sensors are being used in many industrial applications for equipment health monitoring and anomaly detection. However, sometimes operation and maintenance of these sensors are costly. Thus companies are interested in reducing the number of required sensors as much as possible. The straightforward solution is to check the prediction power of sensors and eliminate those sensors with limited prediction capabilities. However, this is not an optimal solution because if we discard the identified sensors. As a result, their historical data also will not be utilized anymore. However, typically such historical data can help improve the remaining sensors' signal power, and abolishing them does not seem the right solution. Therefore, we propose the first data-driven approach based on tensor completion for re-utilizing data of removed sensors and the remaining sensors to create virtual sensors. We applied the proposed method on vibration sensors of high-speed separators, operating with five sensors. The producer company was interested in reducing the sensors to two. But with the aid of tensor completion-based virtual sensors, we show that we can safely keep only one sensor and use four virtual sensors that give almost equal detection power when we keep only two physical sensors.

Dimension reduction (DR) methods play an inevitable role in analyzing and visualizing high-dimensional multi-source data. In the recent decades many variants of these methods have been developed in various disciplines and domains. Due to the diversity and an ever-increasing number of developed techniques, choosing the right method for the given problem is a difficult task. In this study we benchmark 87 methods for integrative dimension reduction of mRNA expression and DNA methylation data, which is a common problem in biology and medicine. Our ranking is obtained based on four main factors: quality of dimension reduction (local, global, and local-global neighborhood preservation), clustering quality, speed and sensitivity to input parameters on multiple datasets generated by InterSIM (a semi-realistic multi-source data simulator in the genomics domain). The results are later validated on a real dataset for breast cancer through visual evaluation metrics such as co-ranking matrices, inspection of true cancer sub-types in two-dimensional projections, and LCMC curves. We also demonstrate the relationship between the methods via network inference. The findings in this study can be useful in algorithm selection and planning of experimental design in multi-source data analysis.

Link prediction is the task of social network analysis whose goal is to predict the links that will appear in the network in future instants. Among the link predictors exploiting the time evolution of the networks, we can find the tensor decomposition-based methods. A major limitation of these methods is the lack of appropriate approaches for estimating their parameters and initialization. In this paper, we address this problem by proposing a parameter setting method. Our proposed approach resorts to optimization techniques to drive the search for an adequate parameter and initialization choice.