Research On Data-based Fault Detection For The Typical Abnormal Operating Conditions Of Fuel Cell

As an environmentally friendly energy technology, PEM fuel cells have received wide attention in the last two decades. Generally it is necessary to detect the faults online to ensure the healthy and stable operation of the equipment, and prevent from damage(even paralysis) caused by long-standing faults, it is necessary to have real-time online test for typical abnormal operating condition of system. PEM fuel cell system structure is complex, with dynamic non-linear characteristics, which makes the system difficult to carry out precise mechanism modeling, thus limiting the apllication of fault detection and other analytical model based on a priori knowledge. The rapid development of sensor technology enables us to get a lot of process data of state information when system is running, which makes data-driven approach to PEM fuel cell systems for fault detection highly feasible.In this paper, we study a data-based approach for nonlinear PEM fuel cell system while non-Gaussian distribution of industrial process data is used. The paper also propose an appropriate fault detection algorithm, and realizes fault detection for PEM fuel cell under typical abnormal operating condition. The main research topics are:(1) For the non-linear characteristics of PEM fuel cell system, we combine the locally weighted projection estimation algorithm and PLS fault detection algorithm, and derive the new detection methods and obtain the corresponding fault detection strategies.(2) The total projection to latent strectures theory is used to further improve the algorithm, which overcomes the shortage of incomplete orthogonality of space projection. On this basis, the four sub-space statistics are obtained, which realizes the real-time monitoring of the main space of atoms that are totally correlated to output varaibles, the sub-space that are orthogonal to output, the sub-space that has a large dsicrepency and the residual space. By combining the above statistics analysis,information about the current state of the system is acquired.(3) For non-Gaussian distribution of the process data, Parzen window method is propsed to calculate statistical thresholds, which has enhanced the accuracy of the algorithm. Combined with improved algorithm, we perform the simulation verification for nonlinear numerical examples.(4) Based on the data of real PEM fuel cell system, we use the designed diagnostic strategies to implement the fault detection of two typical abnormal working conditions,they are electrode water flooding and proton membrane dehydration.