This paper presents a detection and diagnosis fault based on Neural Non Linear Principal Component Analysis (NNLPCA) and a Partial Least Square (PLS). This method is applied on a manufactured system, and the NNLPCA approach is used to estimate the non linear component. This NNLPCA model helps to estimate the prediction error and to define data classes with and without faults. The classes associated to data with faults are isolated by applying a PLS-2. Detecting faults is realized by SPE (square prediction error) statistics method, while locating them is realized by calculating contributions.