Layer-by-Layer Enhancement Strategy of Favorable Features of the Deep Belief Network for Industrial Process Monitoring

Recently, owing to its powerful feature extraction capability, deep learning has been applied to the field of fault detection, in which fault information is reflected in data fluctuation. Deep learning networks transmit fluctuation to the abstract feature layer, and the running state is determined on the basis of the fluctuation of the feature layer. Adjacent layers are connected by weights, whereas the weight value determines the amplification or reduction of input data fluctuation. Thus, abnormal fluctuation information may vanish during layer-by-layer feature extraction and this condition is detrimental to fault detection. In this study, to avoid such an event, the abnormal fluctuation information extracted by the deep belief network is monitored and enhanced layer by layer. Enhancement degree is determined by the fluctuation degree, and the enhanced features are integrated via support vector data description, moving average filter, and kernel density estimation techniques to visualize the current working status. A simulated network is designed to illustrate the vanishment of abnormal fluctuation information, confirm the necessity of an enhancement strategy, and avoid the increase in false detection rate. The comparison based on a complex numerical process and the Tennessee Eastman benchmark process indicates the high-performance detectability of the layer-by-layer enhancement strategy.