Technology of Cognitive Condition Monitoring Systems

Serov, Alexander; Research Group of Automatic Intelligent Data Acquisition (RG AIDA); Russian Federation

Serov, A.; Research Group of Automatic Intelligent Data Acquisition (RG AIDA); Russia

ID: ECNDT-0031-2018
Session: Structural Health Monitoring 4
Room: H1
Date: 2018-06-13
Time: 11:50 - 12:10

We present new intelligent technology which may be applied in the field of NDT for automation of analytical processing of data gathered during condition monitoring of technical systems. Proposed numerical technology is based on a model of Cognitive Sensor (CS). Cognitive Sensor is the main element of architecture of subsystem which is used in Condition Monitoring System (CMS) for the processing of data streams. Described condition monitoring system is realized as Artificial Neural Network which has time-dependent structure – Dynamic Artificial Neural Network (DANN). For the development of this type of neural networks we use methodology of Artificial Subjective Reality (ASR).
Model of proposed neural network in general case has several layers of neurons: input layer, output layer and several hidden layers. Structure of network depends upon experience of processing data gathered by sensors. Advantages of proposed model include ability to learn both linear and non-linear patterns on the basis of processing data streams and ability to implement Life-Long Machine Learning principles. Harmony Theory formulated by Smolensky for dynamical systems may be applied for learning of Cognitive Sensors. Learning of proposed cognitive architecture may be realized as a combination of supervised and unsupervised Machine Learning techniques.
In this paper we represent the model of cognitive CMS which is based on principles of self-organization. Dynamics of structure of neural network in this case includes several different stages: separated evolution of DANNs associated with each CS, interaction of different DANNs aiming to construct single neural network, and evolution of cooperative DANN. We represent results of numerical investigation of proposed models on real-life dataset. In final part of the paper we review the set of most hard problems associated with implementation of ASR-CS based systems and discuss future ways of development of intelligent systems on the basis of CS-DANN architecture.