The task of optimizing a complex technological process involving the production of fuel components molded from separated combustible fractions of municipal waste to be applied in the cement industry, thermal power engineering and professional power engineering comprises two main threads:
The selection of technological processes and their mechanical parameters as well as the interrelationship between the processes is defined on the basis of knowledge in the field of morphological composition, the frequency and amount of supplied waste, as well as on the concept involving further handling of the recovered fractions. Quite significant are such physical properties as humidity (fully mechanical sorting plant is preferred for dry waste) and density of waste. Depending on these parameters, the technological process will take different forms. These installations can differ not only in terms of the devices used in them, but also in terms of their number and order of location in the technological line. Only the proper selection of individual elements of the segregation line guarantees the acquisition of the assumed fraction recovery rates from the mixed municipal waste. In this aspect, the knowledge of the processes used to extract individual fractions from waste seems to be of key importance in the design and operation of a waste segregation line.
All components of the system are connected by a control system based on a complex algorithm. In the context of ensuring the correct course of the technological process, we must consider the Molded Fuels Production Plant as a complex, non-linear, multidimensional control object.
A complex, multi-threaded control algorithm of the process, built into the PLC logic controller, freely programmable (supervised by SCADA) performs the task of solving the optimization problem defined as a neural classifier model (network Support Vector Machine—SVM) of non-metallic fraction components (mainly ballistic and optical separation) where it reproduces the patterns of predefined classes-subsets generated by the greedy approximation algorithm (greedy set cover) containing components-fractions, described by a set of features (including physicochemical properties, fuel properties, emission properties, texture, structure, spectroscopic spectra).
Each subset-representing a pattern of a class of fractions-components-making the molded fuel is the effect of optimization, using the greedy approximation algorithm (greedy set cover), maximizing the objective function (calorific value Wd), meeting at the same time the process and technological constraints imposed on the decision variables of the linear programming tasks (humidity, chlorine share, heavy metals). The patterns of classes are described by a set of features (texture, structure, spectroscopic spectrum), which allows the algorithm at a lower level of the control of the sorting process to carry out classification using the network model SVM (MC SVM) on the basis of a set of features assigned to predefined patterns of classes, represented by disjoint subsets of the cover generated by the master algorithm, Greedy Set Cover, and finally to check if the identified component/fraction belongs to the pattern of class by the incident matrix.
3.1.1.1. Algorithmic Subsystem Model for the Optimization of the Light-Caloric Fraction Production Process—Thread A
The optimization of the production process of light fraction (so-called caloric)—consists in the separation (in the optical separator/s) the waste stream components with high calorific value Wd, exceeding 38 MJ/kg in the case of PE (mainly plastics PE, PET, PP) using the advanced MPC predictive control algorithm (based on the modified multi class vector support machine—MC SVM algorithm model) implemented in the optimizing control layer of the SCADA control system.
The control algorithm of the technological process (
Figure 4), built into the PLC logic controller, freely programmable (supervised by the SCADA system) performs the task of solving the optimization problem defined in the form of a neural classifier model (network Support Vector Machine—SVM) of non-metallic fraction components (mainly ballistic and optical separation) where it reproduces the patterns of predefined classes—subsets generated by the greedy approximation algorithm (greedy set cover) containing components-fractions, described by a set of features (including physicochemical properties, fuel properties, emission properties, texture, structure, spectroscopic spectra).
Each subset-representing a pattern of class of fraction-components-forming the molded fuel is the effect of optimization using the greedy approximation algorithm (greedy set cover), maximizing the objective function (calorific value Wd), meeting at the same time the process and technological constraints imposed on the decision variables of the linear programming tasks. The class patterns are described by a set of features (texture, structure, spectroscopic spectrum), which allows the algorithm at a lower level of the control of the sorting process to carry out classification using the network model SVM (MC SVM) on the basis of a set of features assigned to predefined class patterns, represented by disjoint subsets of the cover, generated by the master algorithm, Greedy Set Cover, and finally to check if the identified component/fraction belongs to the class pattern by the incident matrix.
Pneumatic segregation is based on the use of the correlation of the compressed air stream and the density of the segregated wastes, ensuring that at least two fractions differing in density parameter are selected. An undisputed advantage of this process is the ability to define a segregating parameter, which is the resultant of the gravity and momentum of the segregated particles, thanks to which the pneumatic segregation is a very flexible process that can comprise the applicability of many waste fractions belonging to the raw or partially selected municipal waste stream. The ballistic separation has a similar nature, except that the driving force is provided by the impeller that mechanically ejects the waste into the sorting chamber. The process of aerodynamic separation is carried out in aerodynamic separators, and the process of ballistic separation in ballistic separators. They allow the separation of wastes into two main fractions: light and heavy. The light fraction obtained from aerodynamic segregation may include, for example, paper, foil, plastics, fabrics, and then by passing it to the ballistic separator, it is easy to extract geometrically unstable fractions such as foil, paper, fabrics. The heavy fraction remaining from the processes of aerodynamic and ballistic segregation can be passed for further segregation or classed as ballast and taken to landfills.
Optical separation should be considered noteworthy when separating especially the non-metallic fraction. The development of optoelectronics and computer control systems has contributed in recent years to a significant refinement of this process and to the development of optical separators with the constantly increasing separation efficiency of waste fractions, which in consequence led to the change in economic indicators for the processing of wastes using optical separation methods. An optical separator of any fraction extracted from the stream of mixed municipal waste consists of a scanner (detector) with a system of lamps and a compressed air installation equipped with adjustable impulse nozzle bars, powered by a compressor. It is possible to use a collective power supply system for all devices installed in the technological line fed by compressed air and to recover the waste air stream.
In the optical separation the role of scanners is most frequently taken by detectors:
The distribution of mixed wastes is controlled by a conveyor belt with an advanced feed speed control system. All components of the system are connected by a control system based on a complex algorithm, defined for a specific waste fraction. It should be noted that the change of the algorithm in the situation when the process does not require the replacement of hardware can take place online. For the installations with high throughput, in the event of an unstable morphological composition of wastes, it is economically justified to use even many scanners that classify different optical parameters of the sorted waste. The segregation process itself takes place in the separator chamber, which, like for other belt devices used in the sorting plant, must guarantee the minimization of thickness and uniformity of concentration of the waste layer along the entire working width of the belt, in order to eliminate the overlap of individual waste. The working width is determined by the size of the measuring area generated by a scanner or many scanners, and by the width of the impulse nozzle bar. The transported waste stream is passed to the measuring area of the scanner located above the conveyor belt. The scanner identifies properties of the material such as texture, shape, structure, color, density and spectroscopic spectrum required for it to be recognized. The materials with predefined properties are pneumatically separated at the end of the conveyor belt by an impulse nozzle system. Advanced optical separators based on AAS technology are used in the installations for the production of molded fuels because they allow for the separation of e.g., only plastics desired in fuel (PE, PP, PET), and the ballast e.g., PVC containing chlorine, undesirable in combustion processes is passed to landfills.
The output signals of the preprocessor in the form of successive components of Fourier descriptors, after the transformation ensuring the invariance in terms of scaling, rotation and displacement, are becoming input signals for a multilayer neural network that functions as a pattern recognition system, and at the same time performs classification, i.e., assigns the pattern to the appropriate class. The number of input neurons is equal to the number of Fourier descriptors taken into account in the classification. Given that each output neuron represents one class, their number is also constant and equal to the number of classes. The classifier is trained on a set of learning data representing subsequent classes of patterns to be recognized (or detection based on spectroscopic spectra, texture, outline, structure). In the reproduction mode, the classified pattern, after passing through all the phases of the preprocessor, is fed to the network input, stimulating the output neuron that corresponds with a given class. At the recognition stage of patterns, due to noise pollution, the output signals of network neurons can assume continuous values in the range [0,1], instead of the expected binary values, with
one corresponding to the recognized class. In the classification problem, we used the MC SVM (Multi Class Vector Support Machine) classifier based on a unidirectional neural network implementing various types of activation functions, including, polynomial, radial and sigmoid functions. The task of the classification is to maximize the margin of separation between two different classes described by the set of pairs (
xi,
di), where
xi is the input vector and
di is preset value (for two classes it reaches 1 for class 1 or −1 for class 2). Assuming a linear separability of both classes, the equation of hyperplane separating both classes can be written in the following form [
19]
where
,
.
In this equation, assuming
N inputs, the weight vector
w is N-dimensional. The weight
b is the polarization. The decision equations that determine class membership take the following form:
or after transformations:
If a pair of points (xi, di) satisfies the above equation with an equal sign, then the vector xi = xsv forms the so-called support vector. Supporting vectors are those data points that are closest to the optimal hyperplane.
The problem of training of SVM linear networks, i.e., the selection of synaptic connection weights for linearly separable training data, narrows down to maximizing the separation margin (
Figure 5). It is a problem of quadrant programming with linear weight constraints, which is solved by the method of Lagrange multipliers by minimizing the so-called Lagrange function.
Taking into consideration the fact that the task of training narrows down to maximizing Lagrange function relative to Lagrange multipliers, the primal problem transforms into a dual one, which is formulated as follows [
19]:
with constraints:
By solving the above optimization problem involving Lagrange multipliers we can determine the equation of optimal hyperplane, determined by the weight vector x and the parameter of polarization b.
When solving the problem of classifying non-linearly separable patterns, the problem is narrowed down to determining the optimal hyperplane that minimizes the likelihood of classification error on the training set with the widest possible separation margin.
As in the case of linearly separable patterns, the primal problem is reduced to a dual problem, which is formulated as follows:
with constraints:
For
i = 1, 2, …,
p and for the constant value C adopted by the user. Thus, from the solution of the dual problem we obtain the expression on the vector of weights of the optimal hyperplane in the form:
The summation applies only to training components for which Lagrange multipliers are different from zero. They are the so-called support vectors, whose number is Nsv.
The optimal hyperplane equation depends only on the support vectors. Other vectors from the training data set do not affect the solution result.
The equation for the output signal
y(
x) of the SVM linear network is expressed by the following relationship:
It is a linear equation relative to the input variables described by the vector x and weight dependent on non-zero Lagrange multipliers and corresponding to them Nsv support vectors xi and the setpoints di.
Linear non-separability of patterns does not mean a lack of their separability at all.
Generally, if x is an input vector describing the pattern, then after projecting it into the K-dimensional space, it is represented by a set of features φj (x) for j = 1,2,..., K.
As a result of this transformation, the equation of the hyperplane in the linear space is defined by the formula:
where
wj denotes the weights from
φj (
x) to the output neuron. The vector
W is K-dim and
b is the weight of the polarization. The features of the process described by the functions
φj (
x) have taken over the role of the individual variables
xj.
The value of the output signal is defined by the formula:
The primal problem is solved by the transformation into a dual problem, identically as for networks with linear pattern separability by minimizing Lagrange functions:
The solution to this optimization problem assumes in the first stage that the partial derivatives of Lagrange’s functions should be compared to zero with respect to, w, b and ϵ.
The primal problem is transformed into a dual problem defined relative to Lagrange multipliers
αi in the following form:
with constraints:
The function K (
xixj) present in the formulated dual task is a scalar product of the vector function F
:
This product defines the so-called kernel function.
Ultimately, we obtain the expression defining the SVM nonlinear output signal:
which depends on the kernel function K (
x,
xi) and not on the activation function
φ (x).
Due to the classification problem necessitating the split of data into a larger number of classes, multiple classifications are required using the method “one against all” and “one against one”. In the “one against all” method with M classes, we define M SVM networks that recognize exactly one class. The method requires the training of M SVM networks, each of which is trained on a different data set.
After training all M networks, there is a reconstruction step in which the same vector x is fed to each SVM network, and the output signals (M decision functions) of all trained SVM networks are determined.
The SVM network was selected for the task of classifying patterns due to generalization capabilities. The SVM network is only slightly sensitive to the selected training hyperparameters determining the number of neurons in the hidden layer.
The algorithm for classifying waste fractions in the optical separator, based on the classic version of SVM network, allows to separate fractions-components on the basis of the predefined class patterns described by the vector of parameters characterizing spectral bands (spectroscopic spectra) of textures, structure, shape.
Due to the need to increase the quality indicator of the control of a technological process realized in the optical separator, in terms of the acquisition of products with precisely defined physicochemical, combustion and emission properties, the SVM algorithm was modified by expanding the library of base class patterns (defined on the basis of spectroscopic spectra, textures and geometric features) with a library of predefined class patterns described by physicochemical and combustion properties (calorific value, humidity, volatile content).
The master algorithm, i.e., the greedy approximation algorithm that solves the problem of set cover, is responsible for generating class patterns corresponding to the optimized, disjoint subsets of the set, whose elements are the components of the waste fraction directed to the optical separator(s).
The class patterns are described by a set of features (texture, structure, spectroscopic spectrum), which allows the algorithm at the lower level of the control of the sorting process to carry out a classification using the SVM network model (MC SVM) on the basis of the set of features assigned to the predefined class patterns, represented by disjoint subsets of the cover generated by the master algorithm Greedy Set Cover.
The modification of the algorithm involving the solution of the optimization problem consists in implementing a method that allows to check whether the identified component (based on the analysis of images) belongs to the class pattern defined on the basis of fuel-related and physicochemical properties and in meeting the criterion of the formulated objective function [
20], i.e.,
where:
is the vector of the dimension
(the number of fraction components participating in the molding process of a fuel component) with the components
denoting the calorific values of the molded fuel components. Therefore, the optimization process is narrowed down to a linear programming task (linear objective function and linear constraints) with the constraints imposed on decision variables, determined by the vector
, with the components
that in this way determine a set of permissible solutions.
= of the dimension such that: —denotes the share of the j-th fuel of the fraction.
The objective function given in the general form (1) can be written as follows:
while the system of limit equations in the form:
is a one-dimensional matrix (vector of the left sides of the limit equations) of the dimension
such that the components denote the minimum contents of harmful substances (chlorine, sulfur, PCB, metal, mercury, cadmium) and:
denotes a one-dimensional matrix (vector of the right sides of the limiting equations) of the dimension
such that the components denote the maximum contents of harmful substances. The meaning of particular components of the vector is analogous to
.
The method solving the optimization problem of the coverage of Greedy Set Cover.
The parameter of the method is the pair consisting of a finite set X (the set of waste fraction components) and family F of subsets X (corresponding to the predefined class patterns defined on the basis of fuel-related and physicochemical properties and meeting the criterion of the formulated objective function, i.e., , such that the elements of the set X belongs to at least one subset of the family F: In this case, the subset covers its elements. The solution of the method is the subfamily whose elements cover the entire set X: .
GreedySetCover(X,
) [
19]
- 1
- 2
- 3
while
- 4
to select , which is maximizing
- 5
- 6
- 7
return
The presented listing is the implementation (in pseudo-code) of the GREEDY-SET-COVER algorithm, the operation of which is as follows. In each phase, U is the set of elements not yet covered. The set contains the constructed cover. Line 4 is the step in which the greedy decision is made, i.e., the subset S is selected, covering as many as possible of the uncovered elements (components of the fraction). After the selection of S, its elements are removed from U, and S is added to . When the algorithm stops work, the set is the subfamily F, covering X.
The mathematical model for the optimization of the production process of light (the so-called caloric) fraction-consisting in the separation (in the optical separator/s) of non-metallic components with high calorific value Wd from the oversize heavy fraction of waste stream, is represented by the matrix (column vector).
Z = (zj) of the components with stored indices of class patterns corresponding to the generated by the optimization algorithm Greedy Set Cover, subsets of the coverage of the set of the components pij containing the i-th component contained in the waste stream directed to the optical separator, in the j-th class pattern.
The number of the predefined class patterns is determined by the number of optimal subsets
satisfying the criterion below, found by the Greedy Set Cover algorithms, i.e., greedy approximation algorithm of set coverage:
The master algorithm relative to SVM checks in each iteration whether the identified (based on the analysis of spectroscopic spectra) component/fraction belongs to the pattern of the class defined on the basis of fuel-related and physicochemical properties and meets the criterion of the formulated objective function.
To sum up, the algorithm implemented in the control optimizing layer of the SCADA supervisory system must reproduce the patterns of the predefined class patterns-subsets generated by greedy set cover containing components-fractions described by a set of features, including physicochemical properties, fuel-related properties, Wd, texture, spectroscopic spectra, structure.
Thus, each subset-representing the class of fractions-components-forming the molded fuel is the effect of the optimization with the use of greedy set cover algorithm in line with the criteria on Wd and with the constraints on the decision variables of the optimization task.
In order to resolve the problem of making sequences of control decisions (generating optimal control decision trajectories for the actuators of optical separators), a polymorphic component model of the class representing the MC SVD algorithm was developed, taking into account changes in its definition, with logical representation of knowledge about the object for which the training process consists in successive validation and updating of knowledge and the use of the results of this update for the validation and reconstruction of data structures.
The Multi-Class-Support-Vector-Machine class type contains field definitions with logical representation of knowledge about the control of a technological process and methods which implement the algorithms that realize the task of specializing and adapting the class component in response to the changes in the operating status of devices, process parameters and technological parameters. The specialization follows the training process in the space of states and events through validation, reconstruction of data structures of the object and reconfiguration of the control algorithm.
The component-oriented model of the sorting process of the heavy oversize fraction in the optical separator is a computer implementation of mathematical models (data sets and algebraic expressions) describing individual components (objects, processes, relations). The essence of the model definition involves the use of special class types, whose design allows mapping the specific properties of object components and operations performed on data structures in the form of the so-called fields and methods. Object fields have another important advantage, enabling (through appropriate assignments) to refer to the components which are created in other programs, e.g., Matlab, launched in different address spaces and which make available a number of component categories stored in executable libraries (DLL or EXE), which can be used to develop their own programs by directly using the predefined interfaces or by defining child objects. The purpose of these references is to use the properties and methods (of object components of the predefined libraries).
Definition of the “one against all” MCSVM algorithm.
Algorithm 1 MCSVM |
1: Input: Category N, input for training samples; testing sample T. |
2: Output: Categories of T. |
3: Algorithm: |
4: // training section |
5: for n = 1 to N |
6: Positive Sample ← , Negative Sample ← other samples except |
7: Store the data of classifier |
8: end for |
9: // testing section |
10: for n = 1 to N |
11: Use classifier to calculate the value of |
12: end for |
13: Compare all , output the n corresponding to the maximum of |