METHOD FOR DERIVING FAULT DIAGNOSIS RULES OF BLAST FURNACE BASED ON DEEP NEURAL NETWORK

Abstract

The present disclosure discloses a method for deriving fault diagnosis rules of a blast furnace based on a deep neural network, which relates to the field of industrial process monitoring, modeling and simulation. Firstly, a deep neural network is used to model historical fault data of the blast furnace. Then, for each kind of fault, the process starts from the output layer of the network, wherein sub-models of nodes in the adjacent layers in the deep neural network are established by using the decision tree in sequence, and the if-then rule is derived. Finally, the if-then rules are merged layer by layer, so as to finally obtain fault diagnosis rules of the blast furnace with blast furnace process variables being the rule antecedents and with fault categories being the rule consequents.

Claims

1. A method for deriving fault diagnosis rules of a blast furnace based on a deep neural network, comprising: Step 1: a weight training is performed on a deep neural network by utilizing historical fault data of a blast furnace, and a deep neural network model is established for the blast furnace fault diagnosis, namely learning, and expressing, blast furnace fault diagnosis knowledge from the historical fault data of the blast furnace the knowledge as an abstract nonlinear mapping from a blast furnace process variable to a blast furnace fault category; Step 2: for each kind of fault, the process starts from a last layer of a hidden layer in the deep neural network, wherein a rule antecedent of the if-then rule in a rule subset formed by nodes in the current layer and the next layer is extracted, and duplicate items are removed to form a rule antecedent set; Step 3: each element in the rule antecedent set is a combination of node conditions in the current layer, the decision tree is used to establish an input and output sub-model of the element and nodes in a previous layer, and the decision tree formed in this way makes the nonlinear mapping, from blast furnace process variables to blast furnace fault categories, characterized by the deep neural network model into an intuitive fault diagnosis mode, thus enabling human-computer interaction; Step 4: the decision tree sub-model is used to derive the if-then rule with the node condition in the previous layer being a rule antecedent and with the nodes in the current layer being the rule consequent, and the if-then rule is added into the rule subset of nodes in the previous layer and the current layer; Step 5: Step 3 and Step 4 are repeated until all elements in the rule antecedent set are processed and a rule subset is formed for nodes in the previous layer and in the current layer; Step 6: by setting the previous layer in Step 5 to the current layer, and setting the current layer in Step 5 to the next layer, Step 2, Step 3, Step 4 and Step 5 are repeated again and so on until reaching an input layer of the deep neural network, namely the input layer of the blast furnace process variable parameters; Step 7: according to the sequence of formation of each rule subset, rules are searched in turn for matching the rule antecedents with the rule consequents in adjacent subsets, and are continuously linked to form new rules, so that a if-then rule with the blast furnace process variables as the rule antecedent and with the fault categories as the rule consequent can be obtained finally, and thus a fault diagnosis rule set of the blast furnace can be obtained finally.

2. The method according to claim 1, wherein the structure of the deep neural network in Step 1 is as follows: the depth neural network includes three parts: an input layer, a hidden layer and an output layer; the input layer is a layer for inputting blast furnace process variable parameters including air permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, hot air pressure, and other industrial process parameters characterizing the blast furnace production status; the output layer is a layer of the blast furnace fault categories, including tight furnace operation, hanging, pipeline, material slip, furnace heating, cooling, and other furnace faults related to the blast furnace production; the hidden layer functions to establish a nonlinear mapping from blast furnace process variables to blast furnace fault categories, and a blast furnace fault diagnosis model can be established by leaning the knowledge of blast furnace fault diagnosis from historical fault data of the blast furnace; neurons in the same layer are not connected, but the neurons between layers are fully connected, wherein each connection has a weight value that characterizes the strength of connection between neurons; a mathematical model of deep neural network is: $\begin{array}{cc}{h}_i^j=f\left(\underset{l=1}{\overset{z_{i-1}}{.Math.}}W\left(i,j\right)h_{i-1}^l+b_i^j\right)i=1,\mathrm{.Math.},k;j=1,\mathrm{.Math.},z_{i}y=g\left(\underset{l=1}{\overset{z_k}{.Math.}}W\left(i,j\right)h_k^l+b_{k+1}\right)& \left(1\right)\end{array}$ In the formula, h.sub.i.sup.j is the output of the jth hidden layer cell in the ith layer of the neural network; when h.sub.i is labeled as the ith layer of the neural network, h.sub.0 is the input layer of the neural network and h.sub.k+1 is the output layer of the neural network; the value of j depends on the number of neurons in the ith layer of the network; if the number of neurons in the ith layer is z.sub.i, the value of j in each layer is from 1 to z.sub.i; W(i, j) is a weight matrix corresponding to the jth neuron in the ith layer; b.sub.i.sup.j is an offset item corresponding to the jth neuron in the ith layer, and b.sub.k+1 is an offset item corresponding to the output layer cell; y represents the output of the neural network; f(●) and g(●) are activation functions of the hidden layer cell and the output cell, respectively.

3. The method according to claim 1, wherein steps of forming the decision tree in Step 3 are as follows: the rule subset formed by extracting nodes from the current layer and the next layer in Step 2 is labeled as R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v, wherein v represents the blast furnace fault category, h.sub.j is the current layer, h.sub.j+1 is the next layer and h.sub.j−1 is the previous layer; the historical data samples of the blast furnace participating in the deep neural network training is labeled as x.sub.1, x.sub.2, . . . , x.sub.m, wherein each sample includes air permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, hot air pressure, and other industrial parameters characterizing the blast furnace production status, as well as corresponding categories of blast furnace faults, including tight furnace operation, hanging, pipeline, material slip, furnace heating and cooling; the extracted rule antecedent set is T, and for t∈T, the steps of forming the decision tree are as follows: a. For the blast furnace fault training sample x.sub.1, x.sub.2, . . . , x.sub.m, calculating an activation function value of each neuron of each sample in the h.sub.j−1 layer of the network layer, namely the output of neurons, which is labeled as x.sub.1.sup.j−1, x.sub.2.sup.j−1, . . . , x.sub.m.sup.j−1; b. For the blast furnace fault training sample x.sub.1, x.sub.2, . . . x.sub.m, calculating an activation function value of each neuron of each sample in the h.sub.j layer of the network layer, namely the output of neurons, which is labeled as x.sub.1.sup.j, x.sub.2.sup.j, . . . , x.sub.m.sup.j; c. For x.sub.1.sup.j, x.sub.2.sup.j, . . . , x.sub.m.sup.j, determining that whether it meets the condition of the rule antecedent t, and, if so, it is labeled as 1, otherwise 0, so as to obtain a binary output labeled as y.sub.1.sup.j, y.sub.2.sup.j, . . . , y.sub.m.sup.j; d. Combining x.sub.1.sup.j−1, x.sub.2.sup.j−1, . . . , x.sub.m.sup.j−1 with y.sub.1.sup.j, y.sub.2.sup.j, . . . , y.sub.m.sup.j correspondingly into a data sample (x.sub.1.sup.j−1, y.sub.1.sup.j), (x.sub.2.sup.j−1, y.sub.2.sup.j), . . . , (x.sub.m.sup.j−1, y.sub.m.sup.j) for training the decision tree, wherein the decision tree is trained by a C4.5 algorithm, and the decision tree is recursively constructed by selecting features based on the criterion of maximum information gain rate at each node of the decision tree.

4. The method according to claim 1, wherein steps of deriving the if-then rule by using the decision tree mentioned in Step 4 are as follows: the rule subset of nodes in the previous layer and in the current layer is labeled as R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v, wherein after the decision tree is generated in Step 3, each leaf node is taken as a rule consequent in order to fine all paths from the root node to respective leaf node; nodes and their respective conditions that each path passes through are recorded, wherein an “and” combination is used to form a rule antecedent so that each path reaching the leaf node may form a rule antecedent; an if-then rule are formed by combining the rule antecedent and the rule consequent correspondingly, and added into the rule subset, with each subset characterizing a part of the nonlinear mapping from blast furnace process variables to blast furnace fault categories contained in the deep neural network model of fault diagnosis, which is an expression understandable to blast furnace operators.

5. The method according to claim 1, wherein steps of finding the rule for matching the rule antecedent with the rule consequent in adjacent subsets and forming new rules by continuous linking are as follows: the current layer is labeled as h.sub.j, the later layer as h.sub.j+1, and the previous layer as h.sub.j−1; the rule subset formed by nodes in the current layer and in the next layer is R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v, and the rule subset formed by nodes in the previous layer and in the current layer node is R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v; the initially formed rule subset R.sub.h.sub.k.sub..fwdarw.h.sub.k+1.sup.v takes the blast furnace fault categories as the rule consequent, and the last formed rule subset R.sub.h.sub.0.sub..fwdarw.h.sub.1.sup.v takes the blast furnace process variables as the rule antecedent; the rule antecedent of each rule is extracted from the rule subset R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v for comparison with the rule consequent of the rule in R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v if the rule antecedent is consistent with the rule consequent, the rules corresponding to R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v and R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v are changed into new rules taking the rule antecedent of the rules in R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v as the rule antecedent and taking the rule consequent of the rules in R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v as the rule consequent, which are then added into the new rule subset R.sub.h.sub.j−1.sub..fwdarw.h.sub.j+1.sup.v; it proceeds in this way until all rules in the rule subset R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v are processed and a new rule subset R.sub.h.sub.j−1.sub..fwdarw.h.sub.j+1.sup.v is formed, so that after continuous linking performed in adjacent rule subsets, the if-then rule can be finally obtained which takes blast furnace process variables as rule antecedents and takes fault categories as rule consequents, and finally an expert fault diagnosis rule set R.sub.h.sub.0.sub..fwdarw.h.sub.k+1.sup.v of the blast furnace is obtained for the fault category v.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0027] FIG. 1 is an flow block diagram of the method according to the present disclosure;

[0028] FIG. 2 is a structural schematic view of the deep neural network;

[0029] FIG. 3 is an AUC curve of prediction effects of blast furnace faults by the method according to the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] The present disclosure aims at providing a method for deriving fault diagnosis rules of a blast furnace based on a deep neural network, with the flow block diagram as shown in FIG. 1. Considering the vulnerability and incomplete information of the blast furnace system, in addition to using the advantage of high diagnosis precision of the deep neural network, the abstract knowledge represented by the deep neural network model is also converted into rules readily understandable to blast furnace operators, thereby greatly facilitating the blast furnace operators to understand, modify and reference the fault diagnosis rules of the blast furnace, and being very practical. The method can acquire, and convert, the knowledge from historical fault data of the blast furnace into a form understandable to operators, and realized knowledge and decision-making enhancement in blast furnace fault diagnosis through man-machine cooperation, thus ensuring the confidence levels and accuracy for blast furnace fault diagnosis. In the description below, the effectiveness of the method of the present disclosure is verified by the blast furnace fault data collected from a No. 2 blast furnace in a steel plant.

[0031] The blast furnace is divided into five parts from top to bottom: throat, shaft, bosh, belly and hearth. Cokes, ores and fluxes will go through different changes in different parts within the furnace during the settlement, until they reach the hearth bottom and are completely converted into molten iron and slag. Because of the huge volume of a blast furnace and complex chemical reactions within the furnace, it is extremely important to ensure its safe and stable operation. The blast furnace faults mainly consist of 4 categories: tight operation, hanging, pipeline, and material slip. Data collected during the production includes 29 parameters such as air permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure. During the actual production, workers are organized by a three-shift duty system to monitor and operate the blast furnace iron-making process, which is costly in human resources, and the control mode is relatively extensive which mainly relies on several parameters to determine the furnace condition, so it is difficult to timely diagnose the problems existing in the operation of the blast furnace and carry out accurate control. The method of the present disclosure can solve this problem to a certain extent, and has a usage value in practical application.

[0032] The implementation steps of the present disclosure are described in detail in combination with the specific process as below:

[0033] I Establishing a Deep Neural Network Model

[0034] 80% of samples are taken as training data of the neural network, and the remaining 20% are taken as data for verifying the validity of the finally derived rules. The training data is used to perform the weight training of the deep neural network and establish the deep neural network model of the fault diagnosis. The structure of the deep neural network is as shown in FIG. 2;

[0035] II Utilizing a decision tree to export fault diagnosis rules of blast furnace

[0036] When h.sub.i is labeled as the ith layer of the neural network, h.sub.0 is the input layer of the neural network and h.sub.k+1 is the output layer of the neural network; the input layer is a layer for inputting blast furnace process variable parameters including 29 industrial process parameters characterizing the blast furnace production status, such as air permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure; the output layer is a layer of the blast furnace fault categories, including 4 kinds of furnace faults related to the blast furnace production, such as tight furnace operation, hanging, and pipeline; when h.sub.i.sup.j is the output of the jth hidden layer cell in the ith layer of the neural network, if the number of neurons in the ith layer is z.sub.i, the value of j in each layer is set to 1 to z.sub.i; the rule subset formed by nodes in the current layer and in the next layer is R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v, with v representing the fault category, h.sub.j is the current layer, h.sub.j+1 is the next layer, and h.sub.j−1 is the previous layer; the historical data samples of the blast furnace participating in the deep neural network training is labeled as x.sub.1, x.sub.2, . . . , x.sub.m, wherein each sample includes blast furnace process variables and corresponding blast furnace fault categories; the extracted rule antecedent set is T, and for t∈T, the steps of forming the decision tree are as follows:

[0037] (1) for each kind of fault v, the process starts from a final hidden layer in the deep neural network, wherein a rule antecedent of the if-then rule in a rule subset formed by nodes in the current layer and the next layer is extracted, and duplicate items are removed to form a rule antecedent set T;

[0038] (2) for t∈T, the steps of forming the decision tree are as follows:

[0039] a. For the blast furnace fault training sample x.sub.1, x.sub.2, . . . , x.sub.m, calculating an activation function value of each neuron of each sample in the h.sub.j−1 layer of the network layer, namely the output of neurons, which is labeled as x.sub.1.sup.j−1, x.sub.2.sup.j−1, . . . , x.sub.m.sup.j−1;

[0040] b. For the blast furnace fault training sample x.sub.1, x.sub.2, . . . , x.sub.m, calculating an activation function value of each neuron of each sample in the h.sub.j layer of the network layer, namely the output of neurons, which is labeled as

[0041] c. For x.sub.1.sup.j, x.sub.2.sup.j, . . . , x.sub.m.sup.j, determining that whether it meets the condition of the rule antecedent t, and, if so, it is labeled as 1, otherwise 0, so as to obtain a binary output labeled as y.sub.1.sup.j, y.sub.2.sup.j, . . . , y.sub.m.sup.j;

[0042] d. Combining with x.sub.1.sup.j−1, x.sub.2.sup.j−1, . . . , x.sub.m.sup.j−1 with y.sub.1.sup.j, y.sub.2.sup.j, . . . , y.sub.m.sup.j correspondingly into a data sample (x.sub.1.sup.j−1, y.sub.1.sup.j), (x.sub.2.sup.j−1, y.sub.2.sup.j), . . . , (x.sub.m.sup.j−1, y.sub.m.sup.j) for training the decision tree, wherein the decision tree is trained by a C4.5 algorithm, and the decision tree is recursively constructed by selecting features based on the criterion of maximum information gain rate at each node of the decision tree.

[0043] (3) The rule subset of nodes in the previous layer and in the current layer is labeled as wherein after the decision tree is generated in Step (2), each leaf node is taken as a rule consequent in order to fine all paths from the root node to respective leaf node; nodes and their respective conditions that each path passes through are recorded, wherein an “and” combination is used to form a rule antecedent so that each path reaching the leaf node may form a rule antecedent; an if-then rule are formed by combining the rule antecedent and the rule consequent correspondingly, and added into the rule subset.

[0044] (4) Step (2) and Step (3) are repeated until all elements in the rule antecedent set are processed and a rule subset R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v is formed by nodes in the previous layer and in the current layer;

[0045] (5) By setting the previous layer in Step (4) to the current layer, and setting the current layer in Step (4) to the next layer, Step (1), Step (2), Step (3) and Step (4) are repeated again and so on until reaching an input layer of the deep neural network, namely the input layer of the blast furnace process variable parameters;

[0046] (6) According to the sequence of formation of each rule subset, rules are searched in turn for matching the rule antecedents with the rule consequents in adjacent subsets, wherein the current layer is labeled as h.sub.j, the later layer as h.sub.j+1, and the previous layer as h.sub.j−1; the rule subset formed by nodes in the current layer and in the next layer is R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v, and the rule subset formed by nodes in the previous layer and in the current layer node is R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v; the initially formed rule subset R.sub.h.sub.k.sub..fwdarw.h.sub.k+1.sup.v takes the blast furnace fault categories as the rule consequent, and the last formed rule subset R.sub.h.sub.0.sub..fwdarw.h.sub.1.sup.v takes the blast furnace process variables as the rule antecedent; the rule antecedent of each rule is extracted from the rule subset R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v for comparison with the rule consequent of the rule in R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v; if the rule antecedent is consistent with the rule consequent, the rules corresponding to R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v and R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v are changed into new rules taking the rule antecedent of the rules in R.sub.h.sub.j−1.sub..fwdarw.h.sub.j.sup.v as the rule antecedent and taking the rule consequent of the rules in R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v as the rule consequent, which are then added into the new rule subset R.sub.h.sub.j−1.sub..fwdarw.h.sub.j+1.sup.v; it proceeds in this way until all rules in the rule subset R.sub.h.sub.j.sub..fwdarw.h.sub.j+1.sup.v are processed and a new rule subset R.sub.h.sub.j−1.sub..fwdarw.h.sub.j+1.sup.v is formed, so that after continuous linking performed in adjacent rule subsets, the if-then rule can be finally obtained which takes blast furnace process variables as rule antecedents and takes fault categories as rule consequents, and finally a fault diagnosis rule set R.sub.h.sub.0.sub..fwdarw.h.sub.k+1.sup.v of the blast furnace is obtained for the fault category v.

[0047] III Substituting Industrial Actual Data for Verification

[0048] We take the fault data from January 2018 to December 2019 of the No. 2 blast furnace with a capacity of 2650 m.sup.3 in an iron-making plant as samples, which contain 29 parameters and are sampled at the same sampling rate. The deep neural network training is perform on 80% of data for extracting rules, and the remaining 20% are used to verify the validity of expert rules.

[0049] As shown in FIG. 3, it is an AUC curve of prediction effects of blast furnace faults by the method according to the present disclosure. It can be seen from the fault diagnosis result that the model shows a good effect. The true positive rate of classification accuracy reaches 93%, and the false positive rate keeps low, which indicates that the blast furnace fault samples can be accurately classified, so this method can be applied to actual industrial production.