Titulo Estágio
Transfer learning approaches for linear and non-linear soft sensors based on Process Analytical Technology
Local do Estágio
CERES-DEQ
Enquadramento
With the development of measurement systems and metrology, data currently being acquired in the industry present diverse modalities. Products are no longer characterized by just scalar measurements, i.e., single quantities that address specific aspects of their quality, but by data arrays, usually arising from automatic measurement systems. Common examples include data collected from Process Analytical Chemistry (PAT) techniques, such as near-infrared spectroscopy (NIR), Raman spectroscopy, nuclear magnetic resonance (NMR), fluorescence spectroscopy, among others. These spectra are examples of functional data, i.e. data with an intrinsic continuous structure and logical ordering. Soft sensors based on such spectroscopic data (henceforth called PAT soft sensors [1, 2]) are currently being used in the scope of predictive frameworks to infer key properties of industrial products. Their utility stems from the simple, fast, accurate, and non-destructive nature of spectroscopic measurements as well as their inherent ability to convey useful information about several properties at once.
The practical implementation of PAT soft sensors faces, however, several challenges. One problem is that, after some time, the spectroscopy equipment experiments drift, and the trained model no longer holds as valid, forcing another laborious round of calibration. Furthermore, a model calibrated at a given time for a device cannot be directly applied to other devices, even from the same purpose and from the same vendor [3]. These situations either increase the costs of operating soft sensors, or render them useless after some time.
Here, we will explore transfer learning methods to update the model with as little effort as possible. The transfer may occur in time (for the drift case) or across devices (for the second case). Classical linear methods, such as PLS, will be compared with non-linear methods, such as CNN.
References
1. de Souza, D.C.M., et al., PAT soft sensors for wide range prediction of key properties of diesel fuels and blending components for the oil industry. Computers & Chemical Engineering, 2021. 153: p. 107449.
2. Rato, T.J. and M.S. Reis, SS-DAC: A systematic framework for selecting the best modeling approach and pre-processing for spectroscopic data. Computers & Chemical Engineering, 2019. 128: p. 437-449.
3. Reis, M.S., et al., A Federated Classification Approach of Waste Lubricant Oils in Geographically Distributed Laboratories. Industrial & Engineering Chemistry Research, 2022. 61(48): p. 17544-17556.
Objetivo
Explore and assess the potential of using modern transfer learning methods to improve the management of soft sensors regarding drift phenomena and multiple devices.
Plano de Trabalhos - Semestre 1
1. State-of-the-art review
a. Classical methods for spectral data
b. Workflow for building PAT soft sensors
c. Linear and nonlinear PAT soft sensors
d. Transfer learning techniques
Plano de Trabalhos - Semestre 2
2. Familiarization with the current techniques for constructing PAT soft sensors
3. Test of non-linear methods for building PAT soft sensors
4. Detection of model mismatch (drift case)
5. Analysis of data from multiple devices
6. Definition of transfer learning principles to apply and compare in both cases
7. Tests and Analysis of results
8. Dissertation Writing
Condições
Access to multicore workstations and GPU-enabled mainframe for running the models.
Orientador
Marco Paulo Seabra dos Reis
marco@eq.uc.pt 📩