<p>In recent years, the development of neural network-based sensor fault detection, isolation, and accommodation (SFDIA) schemes has made significant progress. However, these schemes may encounter the issue of fault contamination when faced with concurrent failures of multiple sensors. Fault contamination results in inaccurate measurements from virtual sensors, degrading the capability of conventional SFDIA schemes to make correct fault decisions for sensor failures. To solve the fault contamination problem of conventional SFDIA schemes in concurrent failure scenarios, we propose an enhanced neural network-based SFDIA scheme, in which an additional bank of virtual sensors is introduced. These virtual sensors can provide accurate measurements without being contaminated, by decoupling the process variables from the target variables that need to be predicted, detected, and accommodated. The residual signals between the virtual estimates generated by this bank of virtual sensors and the actual sensor measurements enable the SFDIA to detect and isolate faulty sensors, and further ensure that the fault accommodation can be successfully completed. Finally, two experimental cases are designed on the actual industrial process test platform to verify the effectiveness and superiority of the enhanced scheme in dealing with single and concurrent sensor failures.</p>