PSI - Issue 5

Alexander Serov et al. / Procedia Structural Integrity 5 (2017) 1160–1167 Alexander Serov / Structural Integrity Procedia 00 (2017) 000 – 000

1164

5

The model of Cognitive Sensor includes two aspects of perception of the world. Static aspect is expressed in the model as a set of states S . Dynamic aspect of the model is expressed by a plurality of transitions between states. Accounting the dynamics of states is the key to one important problem of AI, the problem of constructing a hierarchical system of representations. Combining these two aspects into one model makes it possible to build the perception of cognitive system adaptively to the set of processes, which artificial intelligent system is able to observe in the outside world. Construction of monitoring systems ordinary is based on the knowledge about normal and abnormal states of monitored object. This way assumes the presence of a set of interrelated scientific and technological structures responsible for developing, researching and supporting the means of certain monitoring systems. Personnel involved in this work should have expertise in the field of control methods. These personnel should also thoroughly understand the design and technology of controlled equipment. Gradual accumulation of knowledge allows to improve both methods of control, and the design of the object of control. This approach, however, has a number of shortcomings. These include increasing demands for training and an increasingly narrow specialization of monitoring systems. Increasing degree of automation of monitoring data analysis can help in solving these problems. We suppose that in the not too distant future, the technology of engineering used by mankind will undergo a significant transformation: the technology of evolving machines should appear. The emergence of machines that develop and produce other machines will allow human society to completely change its technological foundation. One of the main prerequisites for the emergence of this technology is the implementation of cognitive functions in technical systems. In this section, we would like to illustrate how cognitive functions could be implemented in Structural Health Monitoring systems. We suppose to use ideas of Artificial Subjective Reality, and implement these ideas in cognitive architecture of Dynamic Artificial Neural Networks. Architecture of DANN for using in the field of Structural Health Monitoring is based on the following. We suppose that SHM system is equipped by the set of sensors: Sr = ( Sr 1 , …, Sr M ), where M is a total number of sensors. All sensors make periodical measurements synchronously. Each sensor can process measured data according the logic of Cognitive Sensor model. Use of ASR model supposes the absence of pre-determined information about world observed by SHM system. So the main purpose of actions performed by CS-DANN architecture is to construct representations of observable world on the basis of data gathered by sensor elements. Interaction of separate Cognitive Sensors inside this architecture may be explained by principle of self-organization. Evolution of cognitive CS-DANN architecture may be realized by different scenarios. According scenario represented here this evolution has three main stages. First stage may be characterized as autonomous evolution of different CS. On this stage each Cognitive Sensor makes formation of separate neural network. Neural networks built by different CS doesn’t interact each o ther. CS-DANN structure at this stage may be described as a set of separated subnets. Second stage of evolution is associated with beginning of mutual processing of data by subnets of different Cognitive Sensors. Interaction of networks constructed by different sensors results in originating of network elements that belong to several subnets. This stage of evolution of CS-DANN structure is completing with arising of united neural network. Third stage of the change of neural network may be described as an evolution of the totality of united subnets produced by different Cognitive Sensors. Transition between different stages of evolutionary growth of network occurs as a step-wise process. In numerical experiments with CS-DANN architecture we use emulation of data streams gathered by M = 8 tilt sensors. These sensors were used for condition monitoring of real technical object. Test object was exposed to several different types of impact. Our purpose during numerical simulation was to identify how different types of impact affect the evolution of CS-DANN structure. At the beginning of experiment each CS starts processing data streams; first stage of data processing is associated with intensive originating of perceptive neurons (PN). Spikes generated by originated PN are coming into domain of signal preprocessing. These spikes in turn initiate originating of neurons of preprocessing layer. Each neuron of preprocessing layer starts emitting spikes when it is activated by the set of input spikes from perceptive neurons. Here we will mean that the state s of separate Cognitive Sensor is described by activation of neuron from its preprocessing layer. Transition between states is described by the sequence of activation or reactivation of neurons. On this stage of 4. CS-DANN Architecture for Structural Health Monitoring Systems