Soft Sensor For Fermentation Process Based On Extreme Learning Machine

Biological fermentation process is a complicated process with high non-linear,strongtime-varying and uncertainty. Due to the lack of efficient biosensors and the limitation oftraditional offline sampling analysis method, the measurement for key biological parametersis still a problem. With the rapid development of the computer technology, the soft sensortechnology has aroused widespread concern and research. Utilizing soft sensor method tosolve online measurement of biological parameters has become an effective way. In recentyears, extreme learning machine has drawn more and more attentions from researchers allaround the world. So applying the extreme learning machine theory to establish the softsensor for biological parameters of fermentation process is very meaningful.This paper takes the penicillin fermentation process as research background. Mainlydoing some deeply research about soft measurement modeling methods with extreme learningmachine for fermentation process based on giving a detailed analysis of research status of softsensor and extreme learning machine. The main work is as follows:1. Soft sensor modeling method based on standard extreme learning machine isresearched. For the high non-linear character of fermentation process, firstly a secondaryvariables selection strategy based on mutual information was proposed. And then a modelbased on standard extreme learning machine with the selected secondary variables wasestablished. Some simulation results showed its effectiveness and indicated its advantagesthrough the comparison with BP and SVM.2. Soft sensor modeling method based on batch weighted extreme learning machine isresearched. Combine with the actual feature of fermentation process which is that the changetrack of variable is bound up with initial conditions, the paper takes Euclidean distance as thesimilarity between the initial conditions of each batch of training samples and forecast object,achieves batch weighted modeling by designing a new similarity quantitative function to gainpenalty weight for each batch of training samples. Some simulation results showed that thismethod could improve the prediction accuracy greatly.3. Soft sensor modeling method based on Bayesian extreme learning machine isresearched. The ability of probability prediction for soft sensor is always demanded in realsituation, which can make technical staff to evaluate the true situation of fermentation processand model to avoid mistakes and faults in advance. Bayesian extreme learning machine notonly has the ability of probability prediction, but it also can avoid over-fitting problem ofstandard extreme learning machine through introducing prior. Finally some simulation resultsshowed it can improve the generalization ability of model.4. Soft sensor modeling method based on Bayesian extreme learning machine with inputuncertainty is researched. Noisy input is not considered by the modeling method basedBayesian extreme learning machine, this is not consistent with real condition. To address thisproblem, Bayesian extreme learning machine with input uncertainty is proposed. It is aBayesian extreme learning machine framework which allows for input noise and gets themodel parameters which are overall consideration of input and output noise as well as confidence interval of prediction. The effectiveness was verified by the simulation.