Generation of Interpretable Residuals for Fault Diagnosis Based on Projection Techniques: Leveraging Variable Redundancy
A challenging but common scenario in fault diagnosis of processes concerns both an abundance of normal operation data and expert knowledge available, but no fault data and no predefined set of faults. In that scenario a possible method may consist in generating residuals from the process variables w...
| Autores: | , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Fecha de publicación: | 2026 |
| País: | España |
| Institución: | Universidad de Oviedo (UNIOVI) |
| Repositorio: | RUO. Repositorio Institucional de la Universidad de Oviedo |
| Idioma: | inglés |
| OAI Identifier: | oai:dnet:ruo_________::d5bd2cdcdc234291a801e946ec9b127d |
| Acceso en línea: | https://hdl.handle.net/10651/84121 https://dx.doi.org/10.1109/TCST.2026.3676142 |
| Access Level: | acceso abierto |
| Palabra clave: | Data driven, fault diagnosis, high dimensional spaces, process monitoring, reconstruction method, sparse solutions |
| Sumario: | A challenging but common scenario in fault diagnosis of processes concerns both an abundance of normal operation data and expert knowledge available, but no fault data and no predefined set of faults. In that scenario a possible method may consist in generating residuals from the process variables which are interpretable in the sense that nonzero residuals correspond to variables actually involved in the fault, so the kind of fault occurring can be identified by applying expert knowledge about the process. Under these conditions we lay out the theoretical framework of the generation of residuals with projection techniques with no additional requirements other than to use models that involve low-dimensional latent spaces embedded in high-dimensional ambient spaces. We claim that the estimation of additive faults improves with higher differences between the dimension of the ambient space (measured input data space) and the dimension of the latent space. This improvement can be even greater in the case of sparse faults. In order to take advantage of that fact, we propose possible ways to obtain solutions to a sparse fault reconstruction problem for cases where a linear model is applicable. Some justifications for this idea are provided with simulations and illustrated with a real example using data from a rotating machine. |
|---|