Making Factories Smarter Through Machine Learning
because the data memory size required would need to be exceedingly large- impractical and unnecessary.
The other capability provided by the software is the ability to read complex sensors and perform pre-processing in terms of data reduction: For example, vibration is sampled at least two times the vibration frequency. In this case, a fast Fourier transform is performed and only the frequency of interest is stored. This is an area where there is high opportunity for more efficient processing – effectively using machine learning for pre-processing and feature selection.
Next, data acquired from different sources needs to be joined to assure completeness of database and avoid empty data spaces( Not a Number-NaN) 12. This Sensor fusion is performed using different multivariate techniques in real time.
From this point, as shown in Figure 1, data is processed and then served to a superior process, which can be a machine learning algorithm, visualization and / or storage. The data is sent using OPC Unified Architecture( OPC-UA) or other protocols depending on the needs. For example, if data is needed for real time visualization below 45ms, OPC-UA protocol is used.
Figure 1: The elements and the connectivity being utilized to develop and provide updates to the production system. Based on historical data acquired during typical operation, machine learning algorithms, both unsupervised and supervised, utilize this and other real-time operational data to identify and effectively learn system behavior patterns during the machining process. The data is analyzed in
12
R. C. LUO, C.-C. YIH y K. L. SU, « Multisensor fusion and integration: approaches, applications, and future research directions.,» IEEE Sensors journal, vol. 2, n º 2, 2002
- 32- January 2017