A Study On KPI-Related Fault Diagnosis Methods

Due to the efficiency and economics in fault diagnosis, the data-driven fault diagnosis method has achieved the rapid developments in recent twinty years. Compared with the traditional model-based fault diagnosis method, the accurate first-principle mathematical models are no need for the data-driven mthod, which can achieve the fault diagnosis purpose by analyzing the acquired data from the industrial processes. This property of the data-driven fault diagnosis method makes it very suitable for the modern large-scale and complex systems which are difficult to construct their accurate mathematical models based on first principles, e.g. chemical palnt and large-scale electronic circuit.Key Performance Indicators(KPI)-related fault diagnosis method, which was proposed in recent years according to the requirements of the practical industrial applications, is one category of the data-driven methods. In the practical industrial process, there are always several indicators which attract the most attentions, e.g. the thickness and flatness of the strip steel in a steel milling process. These KPI play the important roles for deternmining whether the final products are qualified. The faults occurred in the process are always the main factors to influence these KPI. However, not all of the ocuurred faults have the influences on the KPI. For an industrial process, different measures should be taken to cope with the different kind of faults. How to deternmine whether the ocuurred faults in the industrial process is a KPI related fault is the main research contents of the KPI-related fault diagnosis. This thesis is a conclusion of the author’s research works on KPI-related fault diagnosis research direction. The contents of the thesis mainly include:(1) The research background information of fault diagnosis and its current research status are descriped, ecpecially for the KPI-related fault diagnosis. Several commonly used data-driven fault diagnosis methods are briely introduced. The traditional Partial Least Squares(PLS) exists the problem that it cannot accurately judge whether the occurred fault is KPI related when it is used in KPI-related fault diagnosis. To cope with this problem, an Improved PLS(IPLS) based fault diagnosis method is proposed in this thesis. Based on the PLS, IPLS divides the fault space into two subspaces, i.e. KPI related and unrelated fault subspaces, to achieve the purpose that accurately judging whether the ocuurred fault is KPI related. Compared with the commonly used KPI-related fault diagnosis methods, IPLS has the advantages in simpler algorithm structure, higher fault detection rate and more accurate fault diagnosis results.(2) Most of the existing KPI-related fault diagnosis mehods are modified from PLS algorithm, yet PLS has several shortcomings when it is used in constructing the data models. For example, the number of the latent variables should be deternmined in PLS, and it directly influence whether the constructed model is accurate, and finally influence the fault diagnosis results. However, currently there is no method that is theoretically proved to be the optimal one in deternming the number of the latent variables. In addition, the commonly used KPI-related fault diagnosis methods exist the problem of fault missing detection. To cope with these existing problems in KPI-related fault diagnosis, a Modified Least Squares(MLS) fault diagnosis method is proposed in this thesis. Since MLS is proposed based on the Least Squares(LS), there is no issue as in PLS to deternmine the number of the latent variables. To overcome the overfitting problem of LS when it is applied in modeling, MLS incorporates a data preprocessing technique to extract part of the KPI unrelated variations existed in the process variables, and then it uses the preprocessed process variables and KPI to construct the fault diagnosis model. In addition, to improve its fault detection rate, MLS divides the fault space into five subspaces, and the Q and Hotelling’s 2T statistical indexes are simultaneously applied.(3) Most of the existing KPI-related fault diagnosis method often obtain several fault diagnosis indexes in their models, and finally derive out the complex fault diagnosis logic. To cope with this problem, the fault diagnosis indexes used in KPI-related fault diagnosis are redefined. Compared with the commonly used diagnosis indexes, the redefined ones are simpler and have more clear meanings. Tennessee Eastman(TE) chemical process is a commonly used benchmark for the process control. In order to make the TE process be more convenient to be used for studing the KPI-related fault diagnosis problems, this thesis has made a detailed analysis on the 21 fault types designed in TE process, and according to analytical results, these 21 faults have been finally divided into two categories, i.e. KPI related and unrelated faults.The effectiveness of all proposed KPI-related fault diagnosis methods in this thesis have been verified on the numerical example and the TE benchmark, and the comparisons have been made with other existing KPI-related fault diagnosis methods. Finally, the Conclusion and Future Work part concludes the thesis, and the future works of this thesis is also described in this part.