-2-FIELD OF THE INVENTION
This invention relates to a process of controlling Blast Furnace Operation to optimize its productivity by generating a hybrid distribution model.
More particularly this invention at a particular instance can suggest the required change in burden distribution in the Blast Furnace so as to achieve favourable set of operating conditions to enhance Blast Furnace performance in terms of productivity and coke rate.
BACKGROUND OF THE INVENTION
The Blast furnace can be considered to be both a chemical reactor as well as counter current heat exchanger. The burden and coke are charged from the top and they descend due to continuous creation of voids by combustion of coke at the tuyeres and removal of iron and slag to the hearth bottom. The hot reducing gaseous product of combustion ascends through the coke grid wetted with molten slag and iron above the raceway and travels upwards through a column of dry lumpy burden.

-3-
Burden distribution (B.O.) is a generic term used to denote the radial material distribution as well as the particle size distrfcutton in a blast furnace. The iron-bearing materials (usually sinter and/or ore) and coke are charged at the top in separate layers in blast furnace. The distribution of burden materials at the top plays a very important role in the performance of the blast furnace, particularly the productivity and the coke rate. It determines the rate of gas flow, solid - gas contact, location and configuration of cohesive zone and the characteristics of dead man zone, thereby affecting the productivity, energy consumption and quality of hot metal.
In the existing art, the blast furnace operators mostly depend on the measurements taken from the inputs and outputs in the control of the process. Because of the hostile environment in the blast furnace (high temperature and pressure, dust, risk of explosion, etc), it is difficult, and expensive, to carry out direct measurements of its internal state. Therefore, the state is usually estimated on the basis of information obtained at the process boundaries such as top gas temperature, hot metal temperature, probe measurements and wall temperatures.

-4-
The experience of blast furnace operator is normally taken into account to interpret the information available for vertical and radial distribution of variables, such as heat load on the boundary and the distribution of the gas across the section of the furnace. Inside the furnace, gas distribution is largely dependent on the distribution of the burden and its estimation is a challenge to the blast furnace, though the measurement from the above burden probes gives some insight to the radial gas distribution in the shaft.
To solve the above constraint methods for automatic classification of Above Burden Probe (ABP) profile is thus warranted. The classification of ABP profile will demonstrate the merits of rapid, consistent and automatic interpretation as well as the risk of misclassification of complex patterns of ABP profile when human experts are involved.
If at any instance, the ABP profile, for the current operating conditions, is unfavourable one, then the operator needs to change the burden distribution so as to achieve target ABP profile. At present, Burden distribution is changed on

-5-
tbe basis of the experience of the operator as welJ as process experts. As there are large number of complex ADP profiles, they need to be classified into limited number of classes so that it may be easy to suggest the required change in Burden distribution so as to move from current to target class of ABP profile.
Importance of the process parameters to be taken into consideration for suggesting the change in Burden Distribution:
Stock Temperatures (ST):
The stack temperature at some particular layers can go down and may increase depending on the other process parameters such as Coke rate, PCI (pulverised coal injection), side working etc. When the side working is reduced it is observed that the cohesive zone drops down. So the stack temperatures play an important role in changing the burden distribution.

-6-
He«t Flux Area (HFA):
Increase or decrease in Heat Flux for some effective areas indicate possible change of burden. This change in heat flux can be due to scaffolding or scaffold removal. So, in order to counter react to these changes, burden needs to be changed.
Pulverised Ceal Injecttoii (PCI):
Pulverised Coal Injection plays an important role in deckling the thermal condition of the furnace. The amount of PCI injected effects the burden and simultaneously calls for a change in the burden distribution. Increase in PCI calls for reduction in coke rate in the burden and vice versa.
Peripheral Weridng Index (PWI)i
Peripheral working index gives an insight of the amount of coke charged on to the walls of the furnace. The greater the amount, the higher is the index. Side working is done HI order to maintain the stack temperatures, which calls for a change in burden distribution.

-7-
Central Working Index (CWI):
Central working index gives an over all picture of the amount of coke charged in the centre of the furnace. In order to shift from peripheral working to central working the distribution is changed which is reflected in the ADP profile.
Penm K (PERMEABILITY):
The stable operation of a blast furnace depends on the even rise of the gases and the chemical and thermal efficiency of the furnace depend on the extent of gas-solid contact. The pressure drop across the blast furnace is expressed as function of permeability factor of the furnace and gas velocity. A large number of factors affect the permeability resistance index viz., properties of the burden and physical properties of the gas, especially the density of the gas. Thus in order to maintain permeability of the furnace blast flow and top gas pressure is monitored which result in fine-tuning of the burden.

-8-
EUCO:
One of the essential requirements for a high performing blast furnace with regard to productivity and coke rate, is consistency in operation with regard to coke burning rate, specific wind rate, efficiency of gas utilisation, total heat demand, top gas temperature and top gas composition. An increase in the efficiency of gas utilisation significantly decreases the coke rate. This change is incorporated by changing the burden.
Selution Loss Carbon (SLC):
Solution loss carbon indicates the thermal condition of the furnace. High value of solution loss carbon gives indication of high fuel rate or fuel consumption of the furnace. In order to counter react this situation PCI is increased or decreased.

-9-
Uppcr K, Middle K, Lower K:
The permeability resistance for the upper, middle and lower stack regions are known as upper k, middles k and lower k. These are monitored with the help of stack pressure probes. Deviations in the individual permeability may result in the change of burden distribution, b tow ing and oxygen enrichment.
Top gas temperature (TOT):
An optimum top gas temperature is a must for smooth operation of the furnace. Moreover it is a good indicator of furnace efficiency. Low top gas temperature can result in problems in the gas cleaning system. The top gas temperature should be high enough to remove any moisture from the burden. Methods of increasing the top gas temperature are:
- lowering the stock line. This may result in change of the distribution pattern;

-10-
- decreasing the RAFT (Raceway Adiabatk Flame Temperature) to change the amount of direct to indirect reduction resulting a higher top temperature.
EUH2:
Amount of water ingress is always followed by appreciable dip in thermal condition and may result in chilling of the hearth. Top gas analysis and monitoring hydrogen percentage in top gas is one of the took, which may indicate the water ingress. In order to counter react to the above situation burden can be changed in order to thermally stabilise the furnace.
Abhay Butsari and Henrik Saxen in Steel Research 66,1995, No.6, 2 31-235 reported a neural network based classification of temperature measurements from an ABP In their work neural networks were used because they are known to be fault-tolerant in certain respects. Also, it can be expected that the relations between inputs and classification provided by human expert are both non-linear and complex.

-11-lt is reported that a large set of temperature profiles into six different stereotype patterns was first classified by an expert from Blast Furnace supervision and roughly 40% of the data collected was used for training the feedforward neural networks, while the remaining were used for the evaluation of the networks' performance in order to determine an appropriate network size. In general, it was found that the network could classify the profiles consistently and rapidly. The configuration (6,10,6) i.e. a network with 10 hidden neurons was found to be optimal. The classifier has been developed as a part of supervision and control system.
As supervised learning method has been used, so one needs to know beforehand the number of classes.
In the current investigation it has been chosen to use unsupervised learning method, which can also be used for classifying unsymmetrical profiles and also, no prior knowledge of the number of classes is required.

-12-
In H. Saxeh, L. Lassus, M Seppa'nen, and T. Karjalahti in iron making and steel making, 2000, vol.27. Pg 207-211 reported a Self Organising Map (SOM) based model for classification, visualisation, and interpretation of blast furnace wall temperature distribution is presented. The model is based on an unsupervised learning method and depicts the results on a two-dimensional feature map, which is used as an operation diagram when the evolution of the wall temperatures is studied. The classifier has been implemented in the automation system of two Finnish blast furnaces and has proved to be a useful tool for operator guidance in daily practice. The model has been further extended by core la ting the wall temperature classes with important performance indices of the furnace, which provides an interpretation of the temperature patterns in terms of process variables that characterize the process and are better understood. The theory of the model is described in the paper and some examples are presented to illustrate its features and use.

-13-
It is important to note that variations of the wall temperature patterns are affected not only by the thickness of accretions but also by the gas flow distribution . Even though these two phenomena are strongly interconnected, it is advisable to consider other supporting measurements before definitive conclusions are drawn concerning changes in the internal state of the furnace. In tools for operator guidance, the way in which to balance information from different models seems to be one of the most challenging tasks.
In the current investigation, an effort has been made to develop model which can take into account all the important process parameters before drawing conclusion concerning changes in the internal state of the furnace.
DESCRIPTION OF THE INVENTION
The main objective of the invention is to develop an offline hybrid burden distribution model which can suggest the required change in burden distribution in the blast furnace so as to achieve favourable set of operating conditions leading to enhanced blast furnace performance in terms of productivity and coke rate.

-14-
Another objective of the invention is to develop an ABP classifier to be integral part of this hybrid model.
A further objective of the invention is to consider a certain set of raw material properties (like composition and granulornetry of coke, sinter and coal respectively; also, Coke Strength Ratio (C.S.R.) and Coke Reduction Index (C.R.I.)), the deviation in certain process parameters like perm K, upper K, tower K, mid K, Heat flux, stack temperature, Pulverized Coal Injection (PCI) and top gas temperature calls for a change in burden distribution so as to move the parameters like ABP, Central Working Index (CWI), Peripheral Working Index (PWI), Solution Loss Carbon (SLC) and top gas composition, in favourable direction.
A still further objective of the invention is to generate the model through sequential steps of data collection, data preparation followed by the SOM (Self Organising Map) based grouping of all the process parameters and ABP profiles to help m finding the correlation between the process parameters and ABP classes. ABP classes will then be correlated to burden distribution classes and to the blast furnace performance.

-15-
Thus, at any instance, if the operator obtains undesirable ABP profile, he can get a guidance from the hybrid model as to how to change the burden distribution so as to achieve target ABP profile and favourable group of process parameters.
According to yet another objective of the invention profile of the burden distribution suggested by the model is simulated by the models that are already existing in the level 2 system of Blast Furnace. For current investigation according to the invention it is chosen to take up the case study at *F Blast Furnace at TATA STEEL, JAMSHEDPUR, INDIA.
According to yet further objective of the invention it is proposed to use an approach based on artificial intelligence, so that model may be tuned for new conditions. The self-organising map (SOM) is an excellent ANN tool used in exploratory phase of data mining. It projects input space on prototypes of a low-dimensional regular grid that can be effectively utilised to visualise and explore properties of the data.

-16-
It has been preferred to use SOM because it uses unsupervised learning method. When the number of SOM units is large, to facilitate quantitative analysis of the map and the data, similar units need to be grouped, i.e., clustered.
Still a further objective of the invention is to use two-stage procedure - first using SOM to produce the prototypes that are then clustered in the second stage - is found to perform well when compared with direct clustering of the data and to reduce the computation time.
In the present invention a classifier for classifying ABP has been developed by two-stage procedure i.e. SOM followed by k-means. K-means is a statistical partive clustering method.
According to the present invention there is provided a process of controlling blast furnace operations to optimize its productivity by generating a hybrid distribution model through two stage investigating procedure i.e. Self Organizing Map (SOM) and followed by K-means to develop a classifier

-17-
for classification of above burden probe (ABP), stack temperature pattern and heat flux, the said classifier being formed by preparing a matrix and cluster of matrix in which for a certain set of raw materials properties like composition and granulometry of coke, sinter and coal, coal strength ratio (CSR) and coke reduction index (C.R.I.), certain process parameters like perm k, upper k, lower k, mid k, heat flux (HF), stack temperature (ST), pulverized coal injection (PCI) and top gas temperature (TGT) being correlated with ABP profile and hence with burden distribution by algorithm comprising the steps of Initialization of weights associated to neurons and other model parameters in different combinations and net work topology arranged In a grid of SOM; normalizing the inputs in input vector (X) by Yninimax' method; developing SOM and updating weights; capturing weights (Wf) of the said SOM based classifier, developing final vector Wfi into further group containing smaller number of clusters through K-means; using shift average inputs formed by using layerwise sectorial average and sectorial area averages for training and development of SOM for classification/grouping of ABP profile, stack temperature pattern and heat flux;

-18-
grouping all the process parameters to develop SOM_1 for a input vector P4, the deviation in which calls for a change in Burden Distribution; developing S0M_2 for identifying similar group of data for an input vector P5; wherein SOM_1 is being formed by communicating input vectors of a set of significant variable parameters such as r (raw material properties), p_k (perm k), u_k (upper k), l_k (lower k), m_k (mid k), pci (pulverized coal injection), tg_T (top gas temperature), cwi (central working index), pwl (peripheral working index), sic (solution loss carbon), tg_c (top gas composition), using som.abp (above burden probe classifier), SOM.stck (stack temperature classifier) and SOM.hf (heat flux classifier) generated from the input vector to identify and descrfoe the class of abp profile [abp (k)], stack temperature pattern [stck (k)] and heat flux group [hf (k)] through transmitting them to SOM_1 by activating node number of the said three SOMs respectively; wherein SOM_1 helps to identify the target ABP, with the corresponding group of process parameters, which is achieved so as to move the current group of significant process parameters in a favourable direction and SOM_2 helping to identify the Burden Distribution that needs to be

-19-
carried out to achieve target ABP and performance indices that are to be achieved; the hybrid burden distribution model thus formed controls the steps of blast furnace operation to optimize productivity by indicating the amount of coke, sinter and ore to be charged at chute angle to achieve target ABP corresponding to a duster generated through the model.
DETAILED DESCRIPTION OF THE INVENTION
This invention discloses a set of variables which discriminates between the set of burden classifications although other variables might be included in the list. The present invention will be better understood with description in relation to the accompanying drawings in which
Figure 1 represents a schemetic view of Blast Furnace process.

-20-
Figure 2 represents the horizontal probe with its eleven measurement points
Figure 3 (a) represents the layers 9,10,11 from which temperature measurements are used for the development of stack temperature classifier.
Figure 3 (b) represents the six sectors in the cross sectional view of the furnace at each stack layer.
Figure 4 represents outline of the Hybrid Burden Distribution Model.
Figure 5 represents ABP classes obtained as practical data at F Blast Furnace for different conditions and process parameters.
In Figure 1 in the Blast furnace, the consecutive layers of ores and coke are shown by charging from the Bell hopper arrangement. Highest temperature zone is maintained in the bosh zone by air blast through tuyers. Below the bosh layer

-21-
Itqutd slag and liquid metal is tapped through tap hole. Gas outlet is shown schematically for the top gas exit. Distribution of burden as well as maintenance of particle size distribution at various levels of the schematically represented * blast furnace is the most important criteria for performance of the Blast Furnace and its productivity. As discussed earlier, an automatic classification of ABP profile taking into consideration all the process parameters to suggest any change of burden distribution is investigated to prepare a hybrid burden distribution model to improve Blast Furnace operations and enhancing productivity is the subject matter of the present invention.
In the present invention a two stage tovestigating procedure i.e. SOM followed by k-means has been adopted to develop a classifier for classification of ABP.
The classifier is developed by projection of input space of process parameters and ABP, on prototypes (vectors representing each class/cluster of the cluster map produced by SOM) wherein each observation being represented as a matrix and effectively utilised to visualise and explore properties of the data through SOM followed by K-means.

-22-
In Figure 2 represents horizontal probes above the burden level with 11 measuring points.
Figure 3 illustrates layers 9,10,11 for temperature measurements which are used for the development of stack temperature classifier.
The mode of operations of Self Organizing Map (SOM) and subsequent k-means are illustrated as follows:
THE SELF-ORGANIZING MAP (SOM)
SOMs constitute an unguided learning method in which an n-dimensional input space is grouped into a regular two-dimensional line of knots, where each knot includes a various amount of samples that are very similar to each other. This similarity is measured by Euclidean distance. In other words, the SOM is a projection of the multidimensional density function into the two-dimensional

-23-
space. Clusters are finally formed by classifying the knots into larger groups. In a way, the resulting map is very general, considering that beforehand assumptions about the shapes and the number of the clusters do not need to be made. The horizontal and vertical axes of the map should not, however, be interpreted generally, because the SOM may move the adjustment of the samples in a nonlinear way. In other words, given the tendency of the SOM to preserve the local structures in a dataset, the interpretation should also be made locally. In addition to the cluster map, the SOM produces pictures of component planes that show the distributions of the component values corresponding to the map. This operation makes the visual interpretation of the clustering both easier and faster.
Cluster analysis organizes data into groups according to similarities among them. In metric spaces, similarity is defined by means of distance based upon the length from a data vector to some prototypical object of the cluster. The prototypes are usually not known beforehand, and are sought by the clustering algorithm simultaneously with the partitioning of the data. Each observation

-24-
conststs of / measured variables, grouped into an /-dimensional column vector Xk
= [Xk,1 , Xk2 _.., Xk,1]7. A set of N observations is denoted by X and
represented as a matrix X = [x1 t X2 ,..„ ., XN]. In pattern recognition
terminology, the columns of X are called patterns or objects, the rows are called the features or attributes and X is called the pattern matrix. The objective of clustering is to divide the data set X into c clusters.
The SOM algorithm is a kind of clustering algorithm which performs a topology preserving mapping from high dimensional space onto map units so that relative distances between data points are preserved. The map units, or neurons, form
usually a two dimensional regular lattice. Each neuron, i= 1, , c, of the
SOM is presented by an l-dimensional weight, or model vector mk = [mk,1.
mk2 , mk1]T. These weight vectors of the SOM form a codebook, and is
considered as cluster prototypes. The response of a SOM to an input xk is
determined by the reference vector (weight) mjo which produces the best match
of the input iok = argmin II m1-XR II (1)

-25-
Where iok represents the index of the Best Matching Unit (BMU) of the k-th input. The number of the neurons determines the granularity of the mapping, which affects the accuracy and the generalisation capability of the SOM. The neurons of the map are connected to adjacent neurons by a neighborhood relation, which dictates the topology of the map. During the iterative training of SOM, the SOM forms an elastic net that folds onto the "cloud* formed by the data. The net tends to approximate the probability density of the data: the codebook vectors tend to drift where the data is dense, while there are only a few codebook vectors where the data are sparse. The training of SOM can be accomplished generally with a competitive learning rule as

Where A k, i is a spatial neighborhood function and a is the learning rate, and the (t) index in bracket denotes the iteration step. Usually, the neighborhood function is

-26-

which represents the Euclidean distance in the low dimensional output space between the /-th vector and the iok winner neuron (BMU). o(t)2 accounts for the degree of participation of the neighboring neurons in the learning process. There are two phases during learning. First, in the ordering phase, the algorithm should cover the full input data space and establish neighborhood relations that preserve the input data structure. This requires competition among the majority of the weights and a large learning rate such that the weights can orient themselves to preserve beat relationships. Hence, in the first phase relatively large initial o(t)2 is used. The second phase of learning is the convergence phase where the local detail of the input space is preserved. Hence the neighborhood function should cover just one unit and the learning rate should be also small. In order to achieve these properties, both the neighborhood function and the learning rate are scheduled during learning.

-27-
K-MEANS:
K-means clustering method partitions the observations in data into K mutually exclusive clusters, and returns a vector of indices indicating to which of the k clusters it has assigned each observation.
k-means treats each observation in the data as an object having a location in space. It finds a partition in which objects within each cluster are as close to each other as possible, and as far from objects in other clusters as possible. Each cluster in the partition is defined by its member objects and by its centroid, or centre. The centroid for each cluster is the point to which the sum of distances from all objects in that cluster is minimised. K-means computes cluster centroids differently for each distance measure, to minimise the sum with respect to the measure that you specify.

-28-
K-means uses an iterative algorithm that minimises the sum of distances (sumdist) from each object to its cluster centroid, over ail clusters. This algorithm moves objects between clusters until the sum cannot be decreased further. The result is a set of clusters that are as compact and well-separated as possible. The details of the minimisation using several optional input parameters to k-means, including ones for the initial values of the cluster centroids, and for the maximum number of iterations are controlled:
To observe how well-separated the resulting clusters are, a silhouette plot is prepared using the cluster indices output from k-means. The silhouette plot displays a measure of how close each point in one cluster is to points in the neighboring clusters. This measure i.e. silhouette value ranges from +1, indicating points that are very distant from neighboring clusters, through 0, indicating points that are not distinctly in one cluster or another, to -1, indicating points that are probably assigned to the wrong cluster. The silhouette value for each point is a measure of how similar that point is to points in its own cluster vs. points in other clusters. It is defined as

-29-

where AVGD_WITHIN(i) is the average distance from the i-th point to the other points in its own cluster, and AVGD_BETWEEN(i,k) is the average distance from the i-th point to points in another cluster k. A more quantitative way to compare the solutions of clustering is to look at the average sifcouette values for each of the cases.
In Figure 4 set of significant variables have been shown in the outline of the model.
The generated model will have resultant graphical user interface (GUI) for the operators. This GUI will depict the online ABP and the suggested target ABP. The model will have an interface for the existing level-2 database for the storage of the results as well as retrieval of the online data.
Figure 5 is an example of the resulted profiles from the generated model showing the target ABP corresponding to 10 clusters.

-30-
Ftofiles depicted in Figure 5 represent classes for: W-shape with some peripheral gas flow; : W-shape with considerable peripheral gas flow; sharp central gas flow with some peripheral flow; L-shape with little peripheral flow; e.g. cluster 7: W-shape with some peripheral gas flow; this pattern represents a common profile, with a quite balanced distribution. The central gas flow makes it possible to maintain permeability and shaft efficiency. While slight flow along the wall prevents the formation of accretions. Cluster 10: W-shape with considerable peripheral gas flow; this class corresponds to a furnace state with excessive peripheral gas flow, which may be used for cleaning accretions from the wall. Cluster 3: sharp central gas flow with some peripheral flow; this pattern has a more pronounced central gas flow, is the target profile for the furnace in question, and corresponds to low-fuel-rate operation. Cluster 2: L-shape with little peripheral flow; this pattern is a good alternative if the furnace is hot apt to form accretions on the wall e.g. if the alkali load is small.
For a certain group of raw materials properties, each ABP class is correlated to Burden Distribution (B.D.) class and performance indices.
At a certain instance, if ABP attained belongs to cluster 7, then the system will suggest the required change in B.O. so as to achieve target ABP belonging to cluster 3 (if according to the calculation of the hybrid model, ABP corresponding to the cluster 3 should be the target ABP) such that the corresponding set of process parameters and conditions are favourable and achievable as compared to the current ones, with an aim to achieve better Blast Furnace performance. So, the system would suggest the amount of coke, sinter and ore to be charged and at what chute angle can they be charged so as to achieve target ABP corresponding to cluster 3.

-31-
Steps involved in the development of hybrid burden distribution model are narrated as below:
A. Algorithm for classification/ grouping of similar data:
1. Initialisation of weights and other model parameters and network
topology: Generally, the weights (Wo) associated to neurons, arranged in
a grid of SOM, are initialised randomly and at very small values. Initial
network topology is selected e.g. hexagonal, random or rectangular.
Other model parameters like 'learning rate for ordering phase', Yiumber of
steps for ordering phase', 'learning rate for tuning phase' and
Vie ighbor hood distance for tuning phase'are initialised at optimum values.
2. As the range of data in the input vector, X, used for training and
development of SOM, is very large, the inputs are normalised by Yninmax'
method to bring it down to compressed scale of [-1, +1]
i

-32-
3. Training algorithm for development of SOM:
SOM is developed and weights are updated according to equation (2). Network topology and network parameters are initialised, as mentioned in step 1.
For the above mentioned parameters, SOM is trained for different combinations of parameters like number of neurons that make up the grid of SOM, distance function, initialised weights associated to SOM and number of epochs upto which SOM is trained.
4. Final weights (Wf) of developed SOM based classifier:
When SOM shows some consistency tn results during training i.e. even on increasing the number of epochs, the outputs and weights of SOM remains almost same, weights of SOM are trapped. Abo, it is studied as to how many outputs of SOM did not match with that of the previous number of epochs. Lower is the number of mismatches of the outputs of the SOM trained for successive number of epochs, better is SOM developed. The weights captured at this point is denoted by Wf.

-33-
5. Clustering of SOM:
If the number of neurons of the finally developed SOM is large and similar neurons/weight vectors is to be further grouped into smaller number of clusters, it is done by k-means. So, the final vector, Wfl, is described by the vector containing centroid of each cluster.
B. Shift average of inputs is used for training and development of SOM for classification/grouping of A8P profile, Stack temperature pattern and Heat Flux. Let, ith data vector of the input vector is PI,
(where PI = [p_abp(l), p_abp(2), , p_abp(nl)]) used for the
development of ABP classifier (let it be denoted by SOM.abp) be given by,
p_abp(i) where p_abp(i) = [Tl_abp(i), T2_abp(i) Tll_abp(i)],
where the inputs are the temperature measurements taken from the 11 points at
the probe fixed above the burden level in the Blast Furnace, as seen in figure 2,
and i=l, 2, n1

-34-
nl = total number of data vectors in the input vector, PI.
SI = number of neurons arranged tn the grid of SOM_abp,
Wa = Final weight vector (of the order of Slxll) associated to SOM_abp,
(Wa is obtained by using algorithm given in step A)

2. Layer wise sectorial average (of 6 sectors, 3 layers i.e. layer 9,10 and 11) is used for the development of classifier in this case. It may be noted that for the case study taken up at T Blast Furnace, layer number 9,10,11, as seen in figure 3 (a), are considered to be most effective for changing Burden Distribution but the layers may be different for other, Blast furnaces.

-35-
Let,
jth data vector (consist of 3 elements) of the input vector, P2,
(where P2 = [p_stck(l), p_stck(2), , p_stck(n2)]) used for
development of Stack temperature classifier (let it be denoted by SOM_stck), is described by
p_stck(j) = [Tl_stck(j), T2_stck(j)/' T3_stck(j)]; where the elements are the average stack temperature measurements of the 6 sectors (AS SHOWN IN
FIGURE 3(b)) at 3 layers respectively in the Blast Furnace and 1=1,2, p2
n2 = total number of data vectors in P2
S2 = number of neurons arranged in the grid of SOM_stck,
Ws = Final weight matrix for SOM_stck, where the order of Ws is S2x3;
(Ws is obtained by using algorithm given in step A)


-3*-
3. Sector wise area average (of 4 areas i.e. areas 1,2,3,4) is used for the development of SOM in this step; it may be noted that heat flux at areas 1,2,3,4 are considered to be most effective for changing Burden Distribution in the case study that has been taken up in the current investigation but the case may be different for other furnaces.
Let, mth data vector (consists of 4 elements) of the input vector, P3,
(where P3 = [p_hf(l), p_hf(2), , p_hf (n3)]) used for the development
of SOM (let it be denoted by SOM_hf) for grouping of heat flux, is described by
p_hf(m) = [Hl_hf(m), H2_hf(m), H3_hf(m), H4_hf(m)], where the inputs are
the average heat flux measurements of the 4 areas, for the 4 sectors
respectively in the Blast Furnace and m»l,2 /i3;
n3 = total number of data vectors in P3; S3 = no. of neurons of the SOM_hf,

-37-
Wh = Final weight matrix associated to finally developed SOM_hf, where the order of the Wh is S3x4; (Wh is obtained by using algorithm given in step A)


-38-
1. In this step, SOM (let it be denoted by SOM_1) is developed for grouping of all the process parameters that call for a change in Burden Distribution. (Let,the input vector used for developing SOM_1 be described by, P4,
(where P4 = [p4(l), p4(2), , p4(n4)]) where n4=total no. of
data vectors in the input vector, P4.
Where, p4(k) = kth data vector in the input vector, P4 k=l,2 ,n4;
= [r(k), P_k(k), u_k(k), l_k(k), m_k(k), hf(k), stck(k), pci(k), tg_T(k), abp(k), cwi(k), pwi(k), slc(k), tg_C(k)],
where p4(k) is made up of following elements and vectors, r(k) = vector containing elements describing raw material properties (composition and granubmetry of coke, sinter and coal respectively; also, Coke Strength Ratio (C.S.R.) and Coke Reduction Index (C.R.I.) are considered)
P_k(k) = perm K, U_k(k) = upper K,

-39-
l_k(k) = lower K,
M_k(k) = mid K,
hf(k) = vector describing Heat flux required for changing burden distribution
= [Hl_hf(k), H2_hf(k), H3Jif(k), H4_hf(k)],
stck(k) = vector describing stack temperature required for changing burden
distribution
= [ Tl_stck(k)f T2_stck(k), T3_stck(k)]....,
pci(k) = Pulverized Coal Injection (PCI),
t0_T(k) - top gas temperature
abp(k) = fector describing 11 temperature measurements from ABP
= [ Tl_abp(k), T2_abp(k), ,TH_abp(k) ]
cwi(k) = Central Working Index (CWI) pwi(k) = Peripheral Working Index (PWI), sic = solution Loss Carbon (SIX), *9_C(k) = top gas composition

-40-
In this case, SOM_1 is developed using input vector P4, so that it can learn to group similar data. SOMs developed in stage B i.e. SOM.abp, SOM_stck and SOMJif, are used to identify the class number or group number (given by activated node number of the three SOMs respectively) that best describes the ABP profile, Stack Temperature pattern and heat flux group respectively (as described by abp(k), stck(k) and hf(k) in the input vector, p4(k).
Let,
11 = node number activated for input abp(k) presented to SOM.abp;
12 = node number activated for input stck(k) presented to
SOM_stck;
13 = node number activated for tnputhf(k) presented to SOM_hf;
Now, in the input vector, p4(k); abp, stck and hf are replaced by il, »2 and i3 respectively and thus, p4(k) is transformed to p4(k)new. Now, input matrix made up of p4(k)ne* is used for training and development of SOM_1.

-41-
2. Now, SOM (denoted by SOM_2) is developed for identifying similar group of data for the following input vector, P5,
(where P5 = [p5(l), p5(2), p5(n5)j), where n5 = total
number of data vectors in the input vector, P5, where, p5(i) is the ith input vector and P5(i) = [abpl(i),bd(i),pi(i)],
where abpl(i) = vector describing 11 temperature measurements from ABP
= [Tl_abpl(i), T2_abpl(i), ,Tll_abpl(i)]
abpl(i) is replaced by the activated node number of SOM_abp and the transformed input vector, pS(i), (let (he transformed input vector, p5(i) be denoted by p5(i)new) is used for the development of S0M_2. bd(i) = vector describing burden distribution = [ca, amt, type], where ca = chute angle,

amt = amount of the burden to be charged,
type = type of the burden to be charged (i.e. coke, sinter or coal, to be
denoted by 0, 1 and 2 respectively)
pi(i) = vector describing performance indices = [ pr, hm_si, cr], where pr = production rate, hm_si = hot metal Si, cr = coke rate
Let, Wk = weights associated to SOM_2. The nodes of SOM_2 are labeled with performance indices. Burden Distribution classes are correlated with ABP classes. As SOM produces component planes, it helps to understand corelation between all process parameters, ABP, B.D. and performance indices of Blast Furnace.
0. When an input vector po at a certain time instance (where po == [r, p_k, u_k, Lk, m_k, hf, stck, pci, tg_T, abp, cwi,

-43-
pwi, sic, tg_C ]) (I.e. corresponding to a input vector p4(k) of
step 1 of stage C) is fed to the hybrid model, the outline of which is given in
figure 4.
a) First, inputs corresponding to ABP, Stack temperature and heat flux are
presented to SOM_abp, SOM.stck and SOMJtf respectively. Then, the node
number of the activated nodes are substituted for the data vectors corresponding
to ABP, Stack temperature and Heat flux respectively. Let, I_abp = node number
activated for input abp presented to SOM.abp ; i.stck = node number activated
for input stck presented to SOM_stck ; i_hf = node number activated for input hf
presented to SOM.hf; Now, substituting i_abp, i.stck and i_hf for abp, stck and
hf in input vector, po, generates po(new) «= [r, p_k, u_k, l_k, m_k, i_hf,
i_stck,pci, tg_T, i_abp, cwi, pwi, sic, tg_C].
b) po(new) vector is fed to SOM_1 and activated neuron number is identified.

-44-
c) then, in SOM_1 optional target neurons (based on the distance from the
current one (in SOM_1)) and associated performance indices are identified.
Optional target neurons are suggested (in the descending order of performance
indices). The weight vector and class of ABP corresponding to target neurons are
also captured.
d) Based on the results obtained in the previous step, and using SOM.2, Burden
Distribution and performane indices corresponding to target ABP is obtained.
Finally, B.D. and corresponding BF performance is suggested to the operator in a
table in a descending order of performance of Blast furnace. As the suggested
B.D. should not change the conditions in the Blast Furance in an abrupt way,
suggestion is such that the Euclidean distance between the node (of SOM_1)
corresponding to current conditions and the node (of SOM_1) corresponding to
suggested conditions is minimum and resulting tn better ABP and performance.
The nodes of SOM_1 are already labeled with performance indices and correlated
to B.D. classes, with the help of S0M_2. Transition from current to suggested
state should be smooth and feasible.

-45-
The model would suggest the best possible B.D. that will help to achieve target ABP. The model would also suggest other optional B.D.s which will result in better ABP than the current one and also better performance of Blast performance. The final solution would be presented in a table containing the suggested Burden Distribution (in the descending order of preference) with corresponding ABP and performance indices of suggested state.
The invention as herein described should not be read in a restrictive manner as various adaptations, modifications and changes are possible within the scope and limit of the invention as encampused in the appended claims.

-46-
WE CLAIM
1. A process of controlling blast furnace operations to optimize its productivity by generating a hybrid distribution model through two stage investigating procedure i.e. Self Organizing Map (SOM) and followed by K-means to develop a classifier for classification of above burden probe (ABP), stack temperature pattern and heat flux, the said classifier being formed by preparing a matrix and cluster of matrix in which for a certain set of raw materials properties like composition and granulometry of coke, sinter and coal, coal strength ratio (CSR) and coke reduction index (C.R.I.), certain process parameters like perm k, upper k, lower k, mid k, heat flux (HF), stack temperature (ST), pulverized coal injection (PCI) and top gas temperature (TGT) being correlated with ABP profile and hence with burden distribution by algorithm comprising the steps of initialization of weights associated to neurons and other model parameters in different combinations and net work topology arranged in a grid of SOM; normalizing the inputs tn input vector (X) by Vninrmax' method; developing SOM and updating weights; capturing weights (Wf) of the said SOM based classifier,

-47-
developing final vector Wf! into further group containing smaller number of clusters through K-means; using shift average inputs formed by using layerwise sectorial average and sectorial area averages for training and development of SOM for classification/grouping of ABP profile, stack temperature pattern and heat flux; grouping all the process parameters to develop SOM_1 for a input vector P4, the deviation in which calls for a change in Burden Distribution; developing S0M_2 for identifying similar group of data for an input vector P5; wherein SOM_1 is being formed by communicating input vectors of a set of significant variable parameters such as r (raw material properties), p_k (perm k), u_k (upper k), \_k (lower k), m_k (mid k), pci (pulverized coal injection), tg_T (top gas temperature), cwi (central working index), pwl (peripheral working index), sic (solution loss carbon), tg_c (top gas composition), using som.abp (above burden probe classifier) SOM.stck (stack temperature classifier) and SOM_hf (heat flux classifier) generated from the input vector to identify and describe the class of abp profile [abp (k)], stack temperature pattern [stck (k)] and heat flux group [hf (k)] through transmitting them to SOM_1 by activating

node number of the said three SOMs respectively; wherein SOM_1 helps to Identify the target ABP, with the corresponding group of process parameters, which is achieved so as to move the current group of significant process parameters in a favourable direction and S0M_2 helping to identify the Burden Distribution that needs to be carried put to achieve target ABP and performance indices that are to be achieved; the hybrid burden distribution model thus formed controls the steps of blast furnace operation to optimize productivity by indicating the amount of coke, sinter and ore to be charged at chute angle to achieve target ABP corresponding to a cluster generated through the model.
2. A process of controlling blast furnace operation as claimed in claim 1, wherein the ABP classifier is developed by projection of input space of process parameters and ABP on proto types (Vectors representing each cluster of cluster map produced by SOM) through SOM algorithms wherein each observation being represented as matrix and effectively utilized to visualize and explore properties of the data via SOM followed by k-means.

-49-
3. A process as claimed in claims 1 and 2, wherein weights and other model
parameters such as learning rate for tuning phase, number of steps for ordering
phase, learning rate for ordering phase, neighborhood distance for tuning phase,
network topology and initial weight associated to neurons and arranged in a grid
of SOM are initialized at optimum values by input vector X.
4. A process as claimed in the preceeding claims wherein SOM is developed and
weight are updated according to the following equation mi (t+1) = mi (t) + a
A**, i (Xk - mi ((t)) in which Atok, i is a spatial neighborhood function and a is the
learning rate and the (t) index denotes the iteration step, i represents a neuron,
Xk represents a matrix for k th observation, mi represents an l-dimenstonal
weight or model vector of the SOM, in which SOM is trained for different
combination of parameters like number of neurons that make up the grid of
SOM, distance function, initialized weights associated to SOM and number of
epochs upto which SOM is trained.

-50-
5. A process as claimed in the preceeding claims wherein, from the trained SOM
outputs and weights of SOM is trapped, record the number of mismatch of
outputs of SOM with that of the previous number of epochs and further grouping
into smaller number of clusters by k means to result final weight vector
containing centroid of each cluster.
6. A process as claimed in the preceeding claims wherein S0M_2 is developed for
identifying similar group of data for the following input vector, P5, (where P5 =
[p5(l)/P5(2), ,p5(n5)J),
Where n5 = total
number of data vectors in the input vector, P5, where, p5(i) is the ith input vector and P5(i)-[abpl(0,bd(i),pi(i)],
where abpl(i) = vector describing 11 temperature measurements from ABP =
[Tl_abpl(i), T2_abpl(i), , Tll_abpl(i)] abpl(i) being

-51-
replaced by the activated node number of SOM.abp and the transformed input
vector, p5(i), (let the transformed input vector, p5(i) be denoted by p5(i)mw) is
used for the development of S0M_2,
wherein bd(i) = vector describing burden distribution
= [ca, amt, type], where ca = chute angle,
amt = amount of the burden to be charged,
type = type of the burden to be charged (i.e. coke, sinter or coal, to be denoted
by 0,1 and 2 respectively).
Pi(i) = vector describing performance indices
= [ pr, hm_si, cr], where
pr = production rate,
hm_si = hot metal Si,
cr = coke rate
let Wk = weights associated to SOM_2, the nodes of SOM_2 being labeled with performance indices, burden distribution classes being correlated with ABP classes and as SOM produces component planes, it helps to understand correlation between all process parameters, ABP, B.D. and performance indices of Blast Furnace.

-52-
7. A process as claimed in the preceedhg claims, wherein
a. first, inputs corresponding to. ABP, Stack temperature and heat flux are
presented to SOM_abp, SOM_stck and SOM_hf respectively, then, the node
number of the activated nodes are substituted for the data vectors corresponding
to ABP, Stack temperature and Heat Flux respectively, where
iabp = node number activated for input abp presented to
SOM_abp ;
i_stck = node number activated for input stck presented to
SOM_stck;
i_hf = node number activated for input hf presented to SOM_hf ; and,
substituting i_abp, i_stck and ijhf for abp, stck and hf in input vector, po,
generates po(new) = [r, p_k, u_k, l_k, m_k, i_hf, i_stck, pci, tg_T, i_abp, cwi,
pwi, sic, tg_C ].

-53-
b. po (new) vector being fed to SOM_1 and activated neuron number is
identified,
c. then, in SOM_1 optional target neurons (based on the distance from the
current one (in SOM_1)) and associated performance indices are identified and
optional target neurons are suggested (in the descending order of performance
indices), the weight vector and class of ABP corresponding to target neurons are
also being captured.
d. and finally burden distribution and corresponding blast furnace performance
is suggested by presenting a table containing suggested BO with the
corresponding ABP and performance indices formed from the results obtained in
the proceeding steps.

-54-
8. A process as claimed in the proceeding claims wherein ABP profile described
by the measurements of temperatures taken from eleven points from the
horizontal probe fixed above the burden level is an input to the model.
9. A process as claimed in the preceeding claims wherein from layers 9, 10 and
11 temperature measurements are used for the development of stack
temperature classifies which layers are considered to be the most effective for
changing burden distribution.
10. A process for controlling to Blast Furnace operations to optimise its
productivity by generating a hybrid distribution model as herein described and
illustrated.