Neural Nets for Fault Diagnosis Based on Model Errors or Data Reconciliation
The Guide to Fault Detection and Diagnosis provides an overview of Neural Networks. This page more specifically describes the use of neural networks combined with conventional first principles models and also with Data Reconciliation. Neural nets provide a way to “learn” models empirically from data. Data Reconciliation provides a way to use mathematical model for state estimation and fault diagnosis.
Combining neural networks with mathematical models for diagnosis
The following is a description of combining the advantages of known mathematical models with the learning capabilities of neural nets, in particular to reduce training and to improve the ability to extrapolate well outside the range of training data. The application is fault diagnosis in systems modeled with material balance constraints -- a supervisory process control application. Traditional data reconciliation techniques are combined with the pattern recognition benefits of neural networks, improving robustness of the diagnostic application. For other published papers on Data Reconciliation, see Data Reconciliation.
The following technical paper is from an ISA (Instrument Society of America) proceedings:
Neural Networks for Fault Diagnosis Based on Model Errors or Data Reconciliation (pdf)
The abstract of that paper is:
Instrument faults and equipment problems can be detected by pattern analysis tools such as neural networks. While pattern recognition alone may be used to detect problems, accuracy may be improved by "building in" knowledge of the process. When models are known, accuracy, sensitivity, training, and robustness for interpolation and extrapolation should be improved by building in process knowledge. This can be done by analyzing the patterns of model errors, or the patterns of measurement adjustments in a data reconciliation procedure. Using a simulation model, faults are hypothesized, during "training", for later matching at run time. Each fault generates specific model deviations. When measurement standard deviations can be assumed, data reconciliation can be applied, and the measurement adjustments can be analyzed using a neural network. This approach is tested with simulation of flows and pressures in a liquid flow network. A generic, graphically-configured simulator & case-generating mechanism simplified case generation.
The presentation at the ISA conference included additional tutorial information on both neural networks and data reconciliation, as well as screen shots showing how modeling and configuration are accomplished. There are also more extensive graphical depictions of the patterns of equation residuals and data reconciliation adjustments for a simple pipeline example. The presentation overheads are at:
Overheads-Neural Networks for Fault Diagnosis Based on Model Errors or Data Reconciliation (pdf)
The study was based on simulation of the following water grid.