Fault diagnosis and prognosis approach using data-driven structurally generated residuals
(English) In this thesis, we will propose methods to combine structural analysis methods and data-driven techniques, extending the applicability of conventional model-based diagnosis methods and proposing an extension to the prognosis. Some researchers have already explored similar ideas: for instan...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Estado: | Versión publicada |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | CBUC, CESCA |
| Repositorio: | TDR. Tesis Doctorales en Red |
| OAI Identifier: | oai:www.tdx.cat:10803/694830 |
| Acceso en línea: | http://hdl.handle.net/10803/694830 https://dx.doi.org/10.5821/dissertation-2117-433727 |
| Access Level: | acceso embargado |
| Palabra clave: | Fault diagnosis Fault prognosis Data-Driven methods Strutural analysis LPV Set membership methods Àrees temàtiques de la UPC::Informàtica 004 - Informàtica 68 - Indústries oficis i comerç d'articles acabats. Tecnologia cibernètica i automàtica |
| Sumario: | (English) In this thesis, we will propose methods to combine structural analysis methods and data-driven techniques, extending the applicability of conventional model-based diagnosis methods and proposing an extension to the prognosis. Some researchers have already explored similar ideas: for instance, there are some researchers have tried to use techniques like Grey-box recurrent neural networks to generate residuals in order to develop hybrid fault diagnosis methods. Another outstanding research is about combining state-space neural networks and model-decomposition methods for fault diagnosis. However, these studies focus more on the diagnosis part without considering the extension to the prognosis. Furthermore, most of the existing prognostic approaches are based on application-dependent methods that extract features of the measured variables. In this dissertation, we will apply the idea of extending residuals or analytical redundancy relations (ARRs) analysis from diagnosis part to prognosis part, including several enhancements as avoiding the need to have the mathematical model of the system and considering modelling uncertainty using interval methods. In this thesis, some data-driven structurally generated residuals will be used with the analysis purpose. Given agraphic (or textual) system description and the available input/output measurements, the structure of ARRs between some inputs and outputs can be determined with the aid of the Structural Analysis (SA) of the system. Then, using a machine learning data-driven approach applied to historical non-faulty data, analytical relations between inputs and outputs can be obtained. Thereby, instead of finding ARRs from a physical-mathematical model, ARRs are obtained by combining SA and data-driven approaches. For the linear system, functions calibrated by means of the System Identification Tool-box of MATLAB® will be used for identification. And to deal with general non-linear systems, the adaptive network fuzzy inference system (ANFIS) data-driven approach is used to implement the diagnosis system. Once the ANFIS model has been identified, it is reformulated in linear parameter varying (LPV) form. Then, a fault detection scheme based on a zonotopic LPV Kalman filter and pole placement method is developed. Finally, a fault isolation scheme based on an improved Demptser-Shafer reasoning approach is proposed. Residuals are used for fault detection purposes activating fault signals when residual values reach anomalous values. In addition, it is possible to predict future faults by means of the detection of anomalous residual deviations. Once an anomalous change in the residual trend has been detected, it is proceeded to estimate when this residual deviation will result in a fault detection and therefore which will be the Remaining Useful Life (RUL) time of the system. For this purpose, the future residual evolution is estimated by means of a regressor function. Nominal and interval parameters of regressor function are estimated with available residual data providing nominal and interval values of the RUL of the system. Along this dissertation, two study cases: a brushless direct current (BLDC) motor and a well-known case study based on a four-tanks system have been used to illustrate the effectiveness of the proposed methods and algorithms. |
|---|