It is o er information technology of identi cation and forecasting of a complex technological system based on structural and parametric modeling in combination with neural network and agent technologies. The function of the neural network module or intelligent agent is to re ne the initially speci ed coupling coe cients between the monitored state and target parameters and to recognize abnormal situations in the system in order to make optimal decisions. The task of recognizing situations consisted in classifying them based on real-time presentation of the current states of the system by belonging to the areas of decision-making. It is o er variant of the architecture of the Hamming neural network with a multilayer recurrent structure, as a specialized heteroassociative memory device with pairs of interconnected input and output vectors. Proposed information technology used in problems of identifying the anomalous state of technological systems of food production and making optimal decisions in the management of the quality of products of agroprocessing enterprises.
Information technologies for the structurally-parametric and situational analysis of complex chemical-technological and biotechnological systems based on the Copyright c by the paper's authors. Copying permitted for private and academic purposes.
In: S. Belim et al. (eds.): OPTA-SCL 2018, Omsk, Russia, published at http://ceur-ws.org
processing of statistical data on a managed object, in conditions of uncertainty
and risk, require the development and inclusion of intelligent modules for
recognizing complex situations for computer support for the adoption of optimal
adequate solutions [
The existing direction of the structural-parametric and situational analysis of
the state of the technological system [
The structural-parametric model (SPM) of the system is represented in the form of a cellular matrix with blocks of indicators placed along the main diagonal and o -diagonal blocks, communication operators between the parameters and their functional groups. The absence of connections is described by the zerooperators k k
0 , which determine the non-working domain of interaction.
Initially, the characteristics of the links are determined expertly path and re ned in the presence of statistical data with the determination of the correlation coe cients and linear multiple regression Pij of the current deviations x1; :::; xn of the system state variables xi from the given norms xj0, depending on the deviation of the factors of the controlled set xj ; j = 1; n j 6= i xi = mi X Pij xj ; i = 1; n; j = 1; n; j 6= i j=1 2
Situational Model of Technological System On the basis of SPM, a situational matrix model of the system Cij xj ; i; j = 1; n by multiplying kCij kn by the diagonal matrix of the normalized deviation vector x1; ::: ; xn : 1 c12 : : : c1n c21 c22 : : : c2n : : : : : : : : : : : : cn1 cn2 : : : 1 x1
x2 : : : : : : : : : : : :
= xn
x1 c21 x1 : : : c12 x2 : : : c1n xn x2 : : : c2n xn : : : : : : : : : cn1 x1 cn2 x2 : : : xn Cij = Pij xxj00 ; i; j = 1; n; j 6= i
i where xi0; xj0 - are the admissible deviations of the variables.
The development of information technology for identifying and forecasting the state of a complex technological system in real time is associated with a rational combination of applied mathematical statistics with the analysis of fuzzy data and self-learning based on methods of arti cial intelligence, neural network and agent technologies.
Then follows the transition to the cognitive matrix kCij kn of relative comparable characteristics of the relationships between the di erent physical quantities xi and xj according to the formula: (1) (2) (3) where xi = jxi xxi0i0j ; i = 1; n - normalized deviations of the state parameters from the range of permissible deviations xi0.
As a result, the elements of the main diagonal of the situational matrix re ect the current normalized deviations of the xi controlled factors from the given values xi0, and the o -diagonal Cij xj ; i; j = 1; n; (i 6= j) - the contributions of the deviations xj ; j = 1; n, to the deviation xi; i = 1; n in accordance with the system of equations xi =
N X Cij j xj ; i; j = 1; n; i 6= j (4) with ordering by rows of all a priori known causes of the deviation of xi, and by columns - of the possible investigative e ects of the deviation of xi on other parameters.
In the general case, the situational matrix kcij xj kn with a multitude of functional elements fx1:::xng and the links between them kcij kn describes a structurally complex situation of cause-e ect interaction of elements in the current state of the system, by combining an a priori knowledge base on the structure of links to current information x.
A formalized algorithm for identifying an abnormal situation in a technological system is as follows.
In the line of the maximum, diagonal element corresponding to the maximum deviation from the norm xi0 in the observed set of state parameters, the maximum nondiagonal element corresponding to the main cause that caused this deviation. Then, on the found column, need to go to the new element of the main diagonal, after which in the new line founds the main cause of the anomaly on this cause-and-e ect step. The search continues until a diagonal deviation are founds, in the line of which all nondiagonal elements will be zero, which means nding the original cause of the anomalous situation.
The registration of current situations in real time complements the original database with the subsequent recalculation of regression coe cients.
However, SPM in the mode of passive observation and accumulation does not always ensure the necessary speed and accuracy of decision making in problems of identi cation and forecasting due to the inadequacy of statistics and the inadequacy of regression bounds. The methods, uses for the passive accumulation of data and active experiment in real time for an operating technological system are practically not realizable, because require a long period of observation or an active experiment with a su cient number of repetitions and veri cation of reproducibility.
In this case, the IT technology of situational modeling of technological systems in real time under conditions of uncertainty and fuzzy data requires further intellectualization based on neural network and agent technologies.
The intellectual function of a neural network module or self-learning agent is in
re ne and correct the originally de ned coupling coe cients between the states
and target parameters, as well as in recognize and classify anomaly situations in
the system as belonging to decision-making classes based on present real-time
conditions [
For situation analysis in conditions of fuzzy and inadequate information, a variant of architecture of the arti cial neural network (ANN) Hamming with a multilayer recurrent structure is proposed as a specialized heteroassociative memory device with a prede ned training sample of reference situations and associated pairs of input and output layer vectors ( gure 1).
The inputs of the network receive values, n components of the current situation vector x1; :::; xn and the network problem is to nd the minimum hemming distance between the input vector and the reference vectors training samples, coded in the network structure.
First ANN layer (neurons 1-3) with unidirect propagation of its output signals to the neurons of the output layer (11-13) has xed values of weights corresponding to the components of the vector of the observed situation (image) so that wi(j1) = x(ji) for i = 1; p (p is the number of neurons of the rst layer).
Similarly, the weights of the output layer (neurons 11-13) correspond to the next vectors of reference situations y(i), related to x(i): wi(j2) = yj(i) (5)
The hidden layer, MAXNET, consists of neurons with feedbacks on the principle of \everyone with each". In this case, with a proper output, the neuron is connected by a positive (exciting) feedback with a weight equal to +1, and with other neurons - negative (overwhelming) feedback with a weight inversely proportional to the number of neurons p.
Neurons of the MAXNET layer (1-3) function in WTA (Winner Takes All) mode so that the network weights should amplify the neuron's own signal and weaken the others. To achieve this e ect - wimi = 1; and p 1 1 < wi(jm) < 0 for i 6= j.
To ensure absolute convergence of the weight algorithm wimi should di er from each other: wi(jm) =
1 p 1 + where - random variable with a su ciently small amplitude.
Neurons of di erent layers of ANN are function di erently. Neurons of the rst layer calculate the Hamming distances between fed on input N - dimensional vectors x and the vectors of the weights w(i) = x(i) individual neurons of this layer (i = 1; 2; :::; p), applied to the input of the network. The values of the output signals of these neurons are determined by the formula: ybi = 1 dh(xi; x)
N (6) (7) (8) (9) where dH (xi; x) denotes the Hamming distance between the input vectors x and x(i), i.e. the number of bits by which these two vectors di er.
The output signals yi of the neurons of the rst layer become the initial states b of the MAXNET layer neurons in the second phase of the network functioning. The task of the neurons of this layer is to determine the winner, i.e. a neuron, whose excitation level is closest to 1 by the recurrence formula: yi(k) = f (yi(k 1) +
X wi(n)yj (k j6=i 1)) at the initial value yj (0) = ybi.
Such a neuron points to an image vector with a minimum Hamming distance to the input vector x.
The activation function f (y) of neurons in the MAXNET layer is given by expression: f (y) = y for y 0 0 for y < 0
The iterative process terminates at a time when the state of the neurons is stabilized and the activity continues to manifest only one neuron, while the rest are in the zero state. The active neuron becomes the winner and, through the weights wi(j2) of the linear neurons of the output layer, represents the vector y(i), which corresponds to the vector x(i), recognized by the MAXNET layer as nearest to the input vector x.
In the process of network operation, we can distinguished three phases. In the rst of them an N -element vector x is fed to its input. After the presentation of this vector, the signals that de ne the initial states of the neurons of the second layer are generates at the outputs of the neurons of the hidden layer, i.e. MAXNET.
In the second phase, the MAXNET-initiated signals are deletes, and the iterative process (8) within this layer are starts from the initial state formed by them. The iterative process terminates at a time when all the neurons, except for the winner with an output signal equal to 1, goes to zero state. A neuronwinner with a non-zero output signal becomes a representative of the data class to which the input vector belongs.
In the third phase, the same neuron, by means of weights connecting it with the neurons of the output layer, forms a response at the output of the network in the form of a vector y, corresponding to the exciting vector x.
The data, receives from the real-time monitoring system and fed to the input of the neural network should be normalized by delta coding, with a pixelated calculation of the di erence in the values of the object's state parameter in the current and previous control cycles, which can signi cantly reduce the dynamic range of the data.
For n parameters, the di erence frame are represents as a column vector x dimensionality n:
X = (x1; x2; :::; xj ; :::; xn)T
Y = (y1; y2; :::; yj ; :::; yn)T at which on the output of the ANN will form a column-vector Y , are formed as:
Trained network before the start of functioning with working technological parameters need to check for the quality of training and the ability to generalize the acquired knowledge. In this case, are found out, whether the results that the network gives out at the outputs are within the permissible error when the sample is fed into the network with predetermined values of the output parameters, but di erent from the training sample.
In the test operation of the network in the food industry of the agro-industrial complex (AIC), the minimum error in training was 1:04%, which corresponds to an allowable error of 1:5%, agreed with the technologists responsible for the quality of food products.
In a particular test implementation, the INS learning algorithm was reduced to the following sequence of actions:
1. Formation of a matrix of reference samples X of the size k Hamming network (Table 1): n of the where T = 0:5p so that the outputs of the neural network can take values within [0; T ].
3. Entering the synapse values of feedbacks of neurons of the hidden layer in the form of elements of a square matrix of size p p: where " 2 [0; p1 ], or in the matrix form:
E = 1 f or diagonal elements
" exclude diagonal elements
Synapses of feedbacks of a Hamming neural network with negative weights are inhibitory.
4. Setting an allowable di erence of output vectors on two consecutive iterations, Emax = 0:1, for estimating the stabilization of the solution found.
The neural network algorithm for classifying situations in the observable system to support the making of managerial decisions in real time under conditions of certainty reduces to the following.
An unknown binary vector are feds on the network inputs !x signals of the current state of the system parameters: xij = 1 if the parameter is within the norm 1 if case of deviation f rom the norm n x1n x1n ::: xjn ::: xpn (10) (11)
In the case of deviation of the state and output values of the neurons of the rst layer are calculated by the formula:
The activation linear-threshold function (10) uses to calculate the outputs of the neurons of the rst layer - !y1. The outputs of the neurons of the second layer are assigned the values of the outputs of the neurons of the rst layer, obtained at the previous step: !y(20) = !y1, after which the rst layer of neurons is practically not involved.
For each q-th iteration in hidden layer calculates new values of states and outputs of neurons by recurrent ratio: s(2qj+1) = y2(qj)
n " X i=1;i6=j y2(qi)
The new output values !y(2j+1) are determines using the linear threshold activation function for processing the corresponding states of the neurons - !s (2j+1). This cycle repeats until the output vector stabilizes in accordance with the condition: k !yj(q+1) !yj(q) k
Emax
In the ideal case, after stabilization, there is an output vector with a single positive element with the remaining zero elements, the index of which indicates the class of the unknown input image of the situation.
If the input image data is very noisy or there is no suitable standard in the training sample, several positive outputs corresponding to the accuracy condition (14) can be obtained as a result. In this case, it is concluded that the input image can not be assigned to a certain class, but the positive output indices indicate the most similar standards.
On an example of classi cation according to three reference situations (Table
2) in the technological system of confectionery production [
In accordance with the learning stage algorithm, a matrix (3 9) is formed to con gure the Hamming neural network for 3 reference images with 9 inputs (Table 2).
1 2 3 1 2 3 If use T = p2 , we determine the threshold of the activation function T = 1.5.
With restriction are " 2 (0; 13 ), the absolute value of the weight of each inhibitory synapse is " = 0:3 and Emax = 0:1 and the matrix of the inverse synapse weighting coe cients (11).
"jp = 1 for j = i
0:3 for j 6= i
The test vector feds to the network inputs: !XT = [1; 1; 1; 1; 1 1; 1; 1; 1] and the condition (14) determines the column vector of the states of the neurons of the rst layer, and at the output of the activation function of the state (10) is the vector-column of the output values of the neurons of the rst layer: 8:00 !s1 = 2:00 3:00
4:50 !y1 = 2:00 3:00
The ANN outputs assigned the corresponding output values of neurons of the rst layer. Then, using the ratio (13), a series of output vectors calculates iteratively until the stabilization condition is satis ed. ANN signals obtained in iteration cycle q when the test situation feds to its inputs, represents in Table 4.
As we can see from the table, the criterion for stopping the feedback loop of the signal after feedback mades after the 4-th iteration. The positive output value of the i-st neuron indicates that the input vector should be assigned to the i-st class. 4
Agent-based Situational Modeling of Systems
The presented neural network technology for recognizing and classifying
situations in real time are suggest to use in describing the dynamics of agent
behavior in complex situations. An intelligent agent is understood as [
Self-learning, purposeful agents are able to accumulate knowledge based on large amount of data and ontology of events in the process of interaction with other agents and the environment, adapt to the situation, choose a strategy for achieving the chosen goal and assess the degree of its achievement.
The general algorithm of the behavior of the intellectual agent [
The parametric description of an agent includes a set of goals and a knowledge
base in a speci c area, a vector of characteristics of its state; bank of models
and strategies of behavior, description of external relations with agents and
the environment. Practical implementation of agent technologies is associated
with the development of simulation systems that provide an experimentation
environment, an agent-oriented language for describing models and software for
organizing experiments [
1. Parametric description of the external environment of the agent's activity with the formalization of a set of factors of in uence on the functional state and objective function of the agent in situational decision-making conditions.
2. Parametric description of functional blocks of the technological system in the form of a set of vectors of input and output factors, state parameters and objective function.
3. Description of the autonomous intelligent agent with a set of state variables, input and sensory variables that communicate with other agents and the environment, as well as dynamics of agent behavior with procedures for learning and identifying current situations and making decisions in the form of discreteevent descriptions and decision-making strategies in conditions of su cient, incomplete and fuzzy information.
4. Creation of agent-oriented model of real-time management of the technological system, which includes, in accordance with the functional scheme of the system: - components describing the state and dynamics of agent behavior; - organizational components that de ne the structure of interrelations between agents and functional blocks of the system;
- mobile components - to describe messages transmitted through a communication channel between agents and moving objects.
5. Software description of the components of the model of the system under
study in a high-level algorithmic language or agent-oriented modeling language
in a universal simulation system [
Agent technologies with neural network algorithms of behavior of learning agents with recognition of current situations open up new possibilities of virtual research of the in uence of various technological factors on the abnormal states of the system and the adoption of optimal solutions in the control system. 5
The proposed direction of intellectualization of situational modeling of systems
is the basis for constructing intellectual expert systems (IES) for making optimal
decisions and operative management of the quality of food products at all stages
of its production at processing enterprises of the agro-industrial complex [
The outlined approach to the development of IT technologies for the identi cation of multi-factor and weakly formalized technological systems based on arti cial intelligence and agent modeling opens new possibilities for computer support for making optimal decisions in conditions of fuzzy information, uncertainty and risk.