Abstract: To address potential unnecessary shutdowns caused by quality-unrelated faults during batch fermentation processes, the paper proposed a noise semi-supervised stacked auto-encoder (NSSAE) to differentiate the quality-relevant and the quality-irrelevant faults. First, mutual information was applied to calculate the contribution from the process variables to quality variables, where artificial noised was introduced to enhance the performance. Second, an NSSAE-based monitoring model was established, wherein indicators for faults and quality variations are separately constructed from the first layer and the last layer of the model. Upon which, kernel density estimation was used to calculate the thresholds for the indicators. Lastly, deep reconstruction-based contribution was used to locate the root cause. Based on the results of numerical simulations and lactic acid bacteria batch fermentation experiments, the NSSAE algorithm proposed in this paper demonstrated the ability to accurately distinguish between quality-related and quality-irrelevant faults. The fault detection rate using the detection index of the first layer of residual space approached 100%. Moreover, the detection index in the final layer of latent space could precisely identify both quality-related and quality-irrelevant faults. Utilizing the DRBC diagnostic method, the specific variable causing the fault can be accurately pinpointed post-fault occurrence. These findings suggest a practical and effective process monitoring method for addressing quality-related and quality-irrelevant fault monitoring issues in the batch fermentation process.