Data-driven Approaches to Fault Detection and Diagnosis under Multiple Faults

Abstract

Fault detection and diagnosis (FDD) has been an important issue in chemical industry for optimal operation and process safety. FDD has three different approaches which are model-based approach, knowledge-based approach and data-driven approach. Recent advances in data acquisition and storage techniques have enabled high-frequency sampling and processing of sensor signals. Therefore, the data-driven methods can handle the limitations of the traditional FDD method.

To improve the FDD performance, three advanced FDD schemes were proposed. The first proposed method was the combination of model-based and data-driven approaches. If the unknown parameters of the process model is inaccurate, the result of FDD with model-based approach can be poor. In addition, since some processes, such as pharmaceutical process, are hard to collect measurement data, the robust parameter estimation with limited data is necessary. In this reason, Bayesian inference was introduced to estimate the unknown parameters of physiologically based pharmacokinetic (PBPK) model with a small number of data. With the proposed estimation scheme, the estimation result was more robust than the least squares method. In addition, the model mismatch was reduced by introducing the drug dissolution model (DDM) into the PBPK model. With these results, FDD performance of model-based approach can be improved.

When the abundant data collection is possible, faulty state data can be classified by the differences between the normal data sets and fault data sets. To describe the data differences, Support vector machine (SVM) which is one of the machine learning technique was introduce to help the transient analysis of water pipe network to diagnose the partial blockage. The time domain transient data were convert to the frequency domain data to find the differences between the normal pipe and blocked pipe. With test experiences with various sizes of the blockage, normal, small blockage, medium blockage and harsh blockage transient data were collected. SVM structures of four cases of blockages were constructed with converted transient data. Finally, SVM structures can classify the blocked pipe and its blockage size automatically with the transient analysis data.

The data-based model is accurate when the learning data describes the characteristic of the process perfectly. Usually, it is impossible to collect perfect learning data from the operating process. Therefore, knowledge-based model can help to reduce model mismatch of the data-based model with prior information of the process and intuition of the expert engineer. Bayesian belief network (BBN) is data-based model which describes the causality between the measurements of the process. To construct BBN structure with imperfect data, weight matrix from the signed digraph (SDG), which is one of the knowledge-based model, was proposed and applied to the structure learning algorithm. In addition, the root cause of the pre-defined fault scenario also introduced into the BBN with prior information of the process. Three case studies was conducted to verify the FDD performance of BBN-based fault diagnosis method with single fault scenarios and multiple fault scenarios. The BBN-based method was effective for all case studies compared with the traditional PCA-based method. Moreover, the fault diagnosis rate of the BBN-based method was better than the PCA-based method for not only single fault cases but multiple faults cases. Consequently, the BBN-based fault diagnosis method, which is the combination of knowledge-based and data-driven approaches, can improve the FDD performance compared with the traditional data-based approaches.

With the three proposed ways to improve the traditional FDD approaches, accurate and real-time process monitoring is possible. Therefore, the proposed methods can help to maintain the process when the failures occur and remain the process with optimal operation condition.