Fault detection of batch process based on MSICA-OCSVM

When we use the traditional multi-scale independent component analysis method to extract independent component ICA on each scale, and then, using the ICA decomposition on the reconstructed data to construct monitoring statistics, however, data of the reconstruction on the nature was already independent component, it is meaningless to extract them by ICA. Focusing on the shortcoming, this paper proposes a MSICA-OCSVM method that was combined with Multi-scale Independent Component Analysis (MSICA) and One-class Support Vector Machine (OCSVM) to monitor the process. First, we can use the wavelet transform decomposition to monitor data at different scales. And then, the data was processing by threshold denoising, and was monitored on each scale extraction by using ICA independent principal component. Subsequently, we can use the wavelet transform coefficients for each scale would scale back on the reconstruction of the new signal matrix X̂. Finally, new OCSVM model was constructed by the reconstructed matrix X̂. We can make the use of determined hyper-plane to construct a nonlinear statistic, and the appropriate control limits was determined by using kernel density estimation. What is more, this method is applied to penicillin fermentation process simulation platform, the experimental results show that this method can effectively utilize the structure information data compared to traditional MSICA fault monitoring method, the failure rate of false positives, false negative rate was significantly reduced.