Artificial neural network (ANN) applications in bioprocess and fermentation operations deal mostly with estimation, control and optimization. An outline and literature review on ANNs are presented in Section 4.6. ANNs have also been used for classification and FDD, which may be formulated as a classification problem. FDD in chemical processes with ANN was initially proposed by Hoskins and Himmelblau  and extended by various researchers [302, 618, 631]. ANN structures have been proposed to detect multiple simultaneous faults [620, 630],
The usual way to apply ANNs to FDD is to classify process operation states using data representing various states of operation of the process (normal or faulty). As discussed in Section 4.6, ANNs are well-suited to solve complicated classification problems especially in the case of highly nonlinear processes such as fermentations. In the most general case, a set of state or input variables are used as a measurement space (input space) and mapped onto a fault space (output space) where variables reflecting malfunctions reside (Figure 8.6). Data are scaled for faster convergence before training the network. Backpropagation is used in most applications as a training algorithm. Once the network is trained with data (Xi,..., Xn) for particular fault conditions (Fi,..., P5-i) as well as normal operating conditions, new observations can be used to classify process operation as faulty or normal using the trained ANN. Variants to this traditional approach have also been suggested. In one case, a two-stage ANN structure is used where an ANN for discriminating among the possible causes of faults
Input Hidden Output
Layer Layer Layer
Input Hidden Output
Layer Layer Layer
(first stage) is followed by another ANN (second stage) that uses the outputs of the previous one to determine the level of deterioration (severity of the deviations) . Such a network will have a number of outputs that is equal to the number of causes x the number of levels of deterioration, this considerably increases the computational requirement. A cascaded hierarchically layered network is also suggested for simultaneously detecting multiple faults . Recently, an alternative two-stage framework was suggested for use of ANNs in FDD . In this two-stage network, a primary network is trained to determine basic process trends (increasing, decreasing and steady) including the level of change. The secondary network receives the outputs from the primary network and assigns them to particular faults that it is trained for. It is reported that when network receives data for an unknown fault, it assigns the fault to either normal operation or untrained faults class .
Most ANN based FDD architectures assume that input-output pairs are available on-line. But in fermentation processes, very important state variables such as biomass and substrate concentrations are measured off-line in the laboratory while measurements on variables such as dissolved oxygen and carbon dioxide concentrations are available on-line. To develop a reliable ANN-based FDD scheme, values of infrequently measured (or off-line available) variables must be provided as well. This can be done by including some state observers or estimators such as Extended Kalman Filters (EKF) (Section 6.5.4) into the FDD framework. Such cascaded ANN-based fault diagnosis system (Figure 8.7) particularly designed for fermentation processes (glutamic acid fermentation in particular) is proposed by Liu . A typical ANN architecture is used in the classifier that is a multi-layer feed-
forward network although infrequently measured variables are estimated by the EKF. It is reported that once the classifier was trained with on-line measurements and estimates of off-line measurements, it achieved 89% fault diagnosis accuracy and it could be implemented in real-time.
Was this article helpful?