PSI - Issue 5

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

1166

7

Fig. 3. Values of prediction RMSE measured for subnet of CS 5 in the range between 260000 and 270000 measurements of tilt.

Each quantity has its own sensitivity with respect to external influences on the monitored object. Perceptive neurons are generated when cognitive system meets new values, i.e. something that still was not percept before in its experience. Appearance of new perceptive neurons means enlargement of the range of possible input signals. Generation of new perceptive neurons on late stages of history of data processing by CD-DANN architecture is an alarming symptom. This may mean sufficient impact on the object of control or sufficient change of its health. Total number of states ( S ) of Cognitive Sensor is sensitive only for creep and static types of impact. This may be explained by implementation of definite preprocessing technique in our experiments. During part of experiment represented in Fig. 1 total number of perceptive neurons was not changing. Appearance of new states in this case may be explained by the significant intensity of the load that was concentrated in time. Total number of transitions between states ( R ) of Cognitive Sensor (Fig. 2) is sensitive to all types of impact. High sensitivity of this quantity is the reflection of its physical nature. Each time when CS-DANN architecture generates new neurons the number of transitions is growing. And as well it is growing when cognitive system meets new sequence of activation of neurons. RMSE of prediction (Fig.3) is a most sensitive quantity which was found in our experiments on analysis of health of technical object. In our experiments we used one of most simple methods for prediction: maximum likelihood prediction which was applied for non-hierarchical representations. High sensitivity of RMSE with respect to change of statistics of neurons firing makes possible to identify weak change of dynamics of outer world. But otherwise this super-sensitivity may blur important changes in the system (compare, for example, Fig.2 and Fig.3 for dynamic and repair impacts). In normal case prediction RMSE tends to decrease to some typical values; increasing of prediction error mean that cognitive system meets unordinary dynamics of the world. On the basis of analysis of numerical experiments with CS-DANN architecture we can conclude that both the change of structure of artificial network and change of dynamics of firing spikes by neurons may be used for detection and identification of events related to structural health. In paper, Serov (2016), we mention that realization of high cognitive functions is two-fold problem connected with calculation of time-and-space model by cognitive system. First part of this problem relates to architecture of perceptive subsystem. Separate Cognitive Sensor is able to build a model of the world, which is based on the concept of sequence. In this type of model the concept of time is natural. It is introduced by cognitive system on the basis of empirical principles as a value which characterizes the dynamics of changes of state. But it is fundamentally impossible introduce the concept of space in this model. Empirically space appears if architecture of cognitive system includes multiple sensors designed to measure the same physical quantity. Second part of mentioned problem relates to way of evolution of neural architecture in attempts to improve time-and-space model. During first stage of evolution of CS-DANN architecture each Cognitive Sensor calculates its own model. Results of these activities are stored in architecture as a