CN105929812A - Strip steel hot continuous rolling quality fault diagnosis method and device - Google Patents

Abstract

The invention proposes a strip steel hot continuous rolling quality fault diagnosis method and device, and the method comprises the steps: collecting historical sample data of a strip steel in a hot continuous rolling production process, wherein the historical sample data comprises a data set generated in a conventional generation production process of the hot continuous rolling of the strip steel; carrying out the clustering division based on the historical sample data, obtaining a plurality of modes, and building models corresponding to all modes; obtaining the data set in a current production process of the hot continuous rolling of the strip steel in real time, and dividing the data in the data set obtained in real time to be divided into the corresponding mode; and carrying out fault diagnosis corresponding to the data in each mode based on a detection index and the models corresponding to all modes. The method solves problems in the prior art that the stability and safety of the production process cannot be guaranteed because the quality of a product is usually controlled by a more skillful worker according to his experience, and the quality of the product is difficult to guarantee through a delay lagged feedback control strategy once a fault happens.

Description

The method for diagnosing faults of hot strip rolling quality and device

Technical field

The present invention relates to steel manufacture process Detection & Controling field, particularly relate to the fault of a kind of hot strip rolling quality Detection method and device.

Background technology

In the industrial production, prediction and monitoring to product quality are the most necessary.Hot strip rolling production process is produced Quality is controlled by the experience of oneself by the most skilled operator often so that the stability of production process and safety Property can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that product quality.

Summary of the invention

The invention provides method for diagnosing faults and the device of a kind of hot strip rolling quality, solve in prior art and produce Quality is controlled by the experience of oneself by the most skilled operator often so that the stability of production process and safety Property can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that the problem of product quality.

First aspect present invention provides the method for diagnosing faults of a kind of hot strip rolling quality, including: gather strip steel heat even Rolling the historical sample data in production process, described historical sample data includes in the generation production process that hot strip rolling is conventional Data set, described data set includes roll gap, roll-force, bending roller force and rack outlet thickness；Based on described historical sample number Obtain multiple mode according to carrying out clustering, and set up model corresponding to each mode；Obtain hot strip rolling in real time and work as previous existence Data in data set during product, and the data set that will obtain in real time are divided in the mode of correspondence；Corresponding to each mould Data in state, the model corresponding based on Testing index and each mode carries out fault diagnosis.

Carry out clustering based on described historical sample data obtain multiple mode as it has been described above, described, including: according to person in servitude Euclidean distance between genus degree matrix and cluster centre and data point, builds core Fuzzy C-cluster (Kernel Fuzzy C- Cluster, KFCC) the first object function of algorithm, in subordinated-degree matrix, each element value is between [0,1]；According to the first mesh The Lagrange multiplier of scalar functions and constraint formula builds the second object function and minimizes value to obtain the second object function Essential condition；Value is minimized according to described first object function, described second object function and described second object function Essential condition, use to introduce and divide (Between-Within Proportion, BWP) index in the m-class of class and determine optimal mould State number, exports cluster result, and described cluster result includes effectiveness desired value and preferable clustering number.

As it has been described above, described first object function:

$J(U,c_1,\mathrm{...},c_c)=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{J}_i=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is to add Power index.

As it has been described above, described second object function:

$\begin{array}{c}\stackrel{\‾}{J}(U,c_1,\mathrm{...},c_c,{\mathrm{\λ}}_1,\mathrm{...},{\mathrm{\λ}}_n)=J(U,c_1,\mathrm{...},c_c)+{\mathrm{\Σ}}_{j=1}^n{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\\ =\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2+\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\end{array}$

Wherein, λ_j(j=1 ..., n) it is formulaThe Lagrange multiplier of n constraint formula, c_iFor mould The cluster centre of paste group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], d_ij=| | c_i-x_j| | it is Euclidean distance between i cluster centre and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

As it has been described above, the essential condition that described second object function minimizes value is:

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{x}_j}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m}$

With

$u_{ij}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\Σ}}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(m-1)}}$

Wherein, c_iFor the cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be [0,1] it Between, x_jFor jth data point, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, d_kj= ||c_k-x_j| | for the Euclidean distance between kth cluster centre and jth data point, and m ∈ [1, ∞) it is Weighted Index.

Model is set up corresponding to each mode as it has been described above, described, including: set up hot strip rolling process variable and outlet The parsing relation of thickness.

As it has been described above, the described parsing relation setting up hot strip rolling process variable and exit thickness, including: for each Class data, utilize core latent structure projection KPLS (Kernel Partial Least Squares, core latent structure projects) method pair Process variable X and quality variable Y is modeled respectively, and described process variable includes described data set.

As it has been described above, in the described mode that data in described real time data group are divided into correspondence, including: utilize KPLS Regression algorithm, and generation k (k=1 ..., 4) individual mode, determine each mode corresponding regression coefficient matrixWherein, i table Show the historical sample data classification that kth mode is corresponding, i=1 ..., k) for the data set x obtained in real time_new, utilize corresponding Average and covariance information data are normalized, obtain forecast quality variate-value:Wherein, x_newRepresenting the data matrix that new real-time online gathers, the transposition of T representing matrix, i represents the history sample that kth model is corresponding Notebook data classification；Utilize prior probability formula:Determine x_newBelong to The prior probability of the i-th class, wherein, i=1 ..., k；Represent the jth sample of the i-th class historical sample data,Represent The predictive value of the i-th class historical sample data jth sample；According to Bayesian formula, determine posterior probability:Wherein, P (i) is the number of samples ratio with total number of samples of the i-th class；By x_newDivide To that class that posterior probability is maximum.

As it has been described above, it is described corresponding to the data in each mode, based on the mould that Testing index and each mode are corresponding Type carries out fault diagnosis, including: Testing index T2 and Q that two kinds of quality based on KPLS method are relevant is synthesized an index φ, for fault detect；Wherein, T²Represent HotellingT²Statistics, detection is the fault directly related with quality variable, Q Representing the Q statistics in statistics, detection is process noise；Contribution rate method is utilized to realize the troubleshooting issue that quality is relevant.

Second aspect present invention provides the trouble-shooter of a kind of hot strip rolling quality, including:

Collecting unit, for gathering the historical sample data in hot strip rolling production process, described historical sample data Including the data set in the generation production process that hot strip rolling is conventional, described data set include roll gap, roll-force, bending roller force with And rack outlet thickness；

Clustering unit, obtains multiple mode for carrying out clustering based on described historical sample data, and corresponding Model is set up in each mode；

Division unit in real time, obtains the data set in the current production process of hot strip rolling, and will obtain in real time in real time Data in the data set taken are divided in the mode of correspondence；

Failure diagnosis unit, for corresponding to the data in each mode, corresponding based on Testing index and each mode Model carry out fault diagnosis.

As it has been described above, described clustering unit, including:

First builds module, for according to the Euclidean distance between subordinated-degree matrix and cluster centre and data point, structure Building the first object function of core Fuzzy C-cluster KFCC algorithm, in subordinated-degree matrix, each element value is between [0,1]；

Second builds module, for building the second target according to the Lagrange multiplier of first object function and constraint formula Function minimizes the essential condition of value to obtain the second object function；

Clustering module, for according to described first object function, described second object function and described second mesh Scalar functions minimizes the essential condition of value, uses division BWP index in introducing the m-class of class to determine optimal mode number, and output is poly- Class result, described cluster result includes effectiveness desired value and preferable clustering number.

As it has been described above, described first object function:

$J(U,c_1,\mathrm{...},c_c)=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{J}_i=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is to add Power index.

As it has been described above, described second object function:

$\begin{array}{c}\stackrel{\‾}{J}(U,c_1,\mathrm{...},c_c,{\mathrm{\λ}}_1,\mathrm{...},{\mathrm{\λ}}_n)=J(U,c_1,\mathrm{...},c_c)+{\mathrm{\Σ}}_{j=1}^n{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\\ =\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2+\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\end{array}$

Wherein, λ_j(j=1 ..., n) it is formulaThe Lagrange multiplier of n constraint formula, c_iFor The cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], d_ij=| | c_i-x_j| | for Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

As it has been described above, the essential condition that described second object function minimizes value is:

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{x}_j}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m}$

With

$u_{ij}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\Σ}}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(m-1)}}$

Wherein, c_iFor the cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be [0,1] it Between, x_jFor jth data point, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, d_kj= ||c_k-x_j| | for the Euclidean distance between kth cluster centre and jth data point, and m ∈ [1, ∞) it is Weighted Index.

As it has been described above, described clustering unit, it is additionally operable to the solution setting up hot strip rolling process variable with exit thickness Analysis relation.

As it has been described above, described clustering unit, specifically it is additionally operable to, for each class data, utilize core latent structure to project Process variable X and quality variable Y is modeled by KPLS method respectively, and described process variable includes described data set.

As it has been described above, described real-time division unit, including:

First determines module, is used for utilizing KPLS regression algorithm, produce k (k=1 ..., 4) individual mode, determine each mode Corresponding regression coefficient matrixWherein, i represents the historical sample data classification that kth mode is corresponding, i=1 ..., k)

Prediction module, for the data set x for obtaining in real time_new, utilize corresponding average and covariance information to data It is normalized, obtains forecast quality variate-value:Wherein, x_newThe real-time online representing new gathers Data matrix, the transposition of T representing matrix, i represents the historical sample data classification that kth model is corresponding；

Priori module, is used for utilizing prior probability formula:Really Determine x_newBelong to the prior probability of the i-th class, wherein, i=1 ..., k；Represent the jth sample of the i-th class historical sample data,Represent the predictive value of the i-th class historical sample data jth sample；

Posteriority module, for according to Bayesian formula, determines posterior probability:Wherein, P (i) is the number of samples ratio with total number of samples of the i-th class；

Divide module in real time, for by x_newIt is divided into that class that posterior probability is maximum.

As it has been described above, described failure diagnosis unit, including:

Synthesis module, for synthesizing a finger by Testing index T2 and Q that two kinds of quality based on KPLS method are relevant Mark φ, for fault detect；Wherein, T²Represent HotellingT²Statistics, detection is the fault directly related with quality variable, Q represents the Q statistics in statistics, and detection is process noise；

Diagnostic module, for utilizing contribution rate method to realize the troubleshooting issue that quality is relevant.

The fault detection method of the hot strip rolling quality that the embodiment of the present invention provides and device, utilize hot strip rolling existing The substantial amounts of data that can reflect production process that collect of field establish production process variable and end product quality thickness it Between relation, provide important theory support and technical support for the quality producing process monitoring, diagnosing and monitor product.Solve In prior art, product quality is controlled by the experience of oneself by the most skilled operator often so that production process Stability and safety can not be guaranteed, once break down, only by postpone delayed feedback control strategy it is difficult to ensure that The problem of product quality.

Accompanying drawing explanation

The flow chart of the method for diagnosing faults of a kind of hot strip rolling quality that Fig. 1 provides for the embodiment of the present invention；

In the fault detection method of the hot strip rolling quality that Fig. 2 provides for the embodiment of the present invention, the flow process of step 102 is shown It is intended to；

The technology arrangement figure of the hot strip rolling that Fig. 3 provides for the embodiment of the present invention；

The fault detection method process flow diagram of the hot strip rolling quality that Fig. 4 provides for the embodiment of the present invention；

Fig. 5 is the failure detection result that thickness based on the KPLS method improved is relevant；

Fig. 6 is the fault diagnosis result that thickness based on contribution rate method is relevant；

The structural representation of the trouble-shooter of a kind of hot strip rolling quality that Fig. 7 provides for the embodiment of the present invention.

Detailed description of the invention

The invention belongs to steel manufacture process Detection & Controling field, particularly propose a kind of hot strip rolling thickness matter Amount monitoring and fault diagnosis method, it is to ensure that final product thickness satisfactory quality, as target, establishes strip steel heat even Roll the relation of process variable and exit thickness, it is provided that a kind of this complex nonlinear of hot strip rolling, multi-modal batch process The new method that mode divides, provides theory support and technical support for hot strip rolling thickness qualities monitoring and fault diagnosis.

The flow chart of the method for diagnosing faults of a kind of hot strip rolling quality that Fig. 1 provides for the embodiment of the present invention.

As it is shown in figure 1, embodiments provide the method for diagnosing faults of a kind of hot strip rolling quality, including:

Step 101, the historical sample data gathered in hot strip rolling production process.

Wherein, described historical sample data includes the data set in the generation production process that hot strip rolling is conventional, described Data set includes roll gap, roll-force, bending roller force and rack outlet thickness.

Step 102, carry out clustering based on described historical sample data and obtain multiple mode, and corresponding to each mode Set up model.

Data set in step 103, in real time the acquisition current production process of hot strip rolling, and the data set that will obtain in real time In data be divided into correspondence mode in.

Step 104, corresponding to the data in each mode, the model corresponding based on Testing index and each mode is carried out Fault diagnosis.

The fault detection method of the hot strip rolling quality that the embodiment of the present invention provides, utilizes hot strip rolling collection in worksite To the substantial amounts of data that can reflect production process establish the pass between production process variable and end product quality thickness System, for producing process monitoring, the quality important theory support of offer diagnosing and monitoring product and technical support.Solve existing In technology, product quality is controlled by the experience of oneself by the most skilled operator often so that stablizing of production process Property and safety can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that product matter The problem of amount.

In the fault detection method of the hot strip rolling quality that Fig. 2 provides for the embodiment of the present invention, the flow process of step 102 is shown It is intended to.

As in figure 2 it is shown, in one embodiment, described carry out clustering based on described historical sample data and obtain multiple mould State, including:

Step 1021, according to the Euclidean distance between subordinated-degree matrix and cluster centre and data point, build KFCC and calculate The first object function of method, in subordinated-degree matrix, each element value is between [0,1].

In the present embodiment, described first object function can be expressed by below equation:

$J(U,c_1,\mathrm{...},c_c)=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{J}_i=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is to add Power index.

Step 1022, according to first object function and constraint formula Lagrange multiplier build the second object function to obtain Take the second object function and minimize the essential condition of value；

In the present embodiment, described second object function can be expressed by below equation:

$\begin{array}{c}\stackrel{\‾}{J}(U,c_1,\mathrm{...},c_c,{\mathrm{\λ}}_1,\mathrm{...},{\mathrm{\λ}}_n)=J(U,c_1,\mathrm{...},c_c)+{\mathrm{\Σ}}_{j=1}^n{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\\ =\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{d}_{ij}^2+\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_j(\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{u}_{ij}-1)\end{array}$

Wherein, λ_j(j=1 ..., n) it is formulaThe Lagrange multiplier of n constraint formula, c_iFor The cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], d_ij=| | c_i-x_j| | for Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

In the present embodiment, described second object function minimizes the essential condition of value and can pass through below equation table Reach:

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m{x}_j}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^m}$

With

$u_{ij}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\Σ}}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(m-1)}}$

Wherein, c_iFor the cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be [0,1] it Between, x_jFor jth data point, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, d_kj= ||c_k-x_j| | for the Euclidean distance between kth cluster centre and jth data point, and m ∈ [1, ∞) it is Weighted Index.

Step 1023, reach according to described first object function, described second object function and described second object function To the essential condition of minima, use division (Between-Within Proportion, BWP) index in introducing the m-class of class true Fixed optimal mode number, exports cluster result, and described cluster result includes effectiveness desired value and preferable clustering number.

In another embodiment, described set up model corresponding to each mode, including: set up hot strip rolling process variable with The parsing relation of exit thickness.

Concrete, in one embodiment, the described parsing relation setting up hot strip rolling process variable and exit thickness, bag Include: for each class data, utilize KPLS method that process variable X and quality variable Y is modeled respectively, described process variable Including described data set.

Wherein, when clustering, in class in terms of sample, should have more identical point, and between class in terms of sample, should There is more dissimilarity.The present invention is when clustering, and with distance measure as starting point, divides in introducing the m-class of class (Between-Within Proportion, BWP) index, each class data should mainly include roll gap, roll-force and bending roller force and Corresponding exit thickness.

Concrete, in the described mode that data in described real time data group are divided into correspondence, including: utilize KPLS to return Reduction method, and generation k (k=1 ..., 4) individual mode, determine each mode corresponding regression coefficient matrixWherein, i represents The historical sample data classification that kth mode is corresponding, i=1 ..., k) for the data set x obtained in real time_new, utilize corresponding Data are normalized by average and covariance information, obtain forecast quality variate-value:Wherein, x_new Representing the data matrix that new real-time online gathers, the transposition of T representing matrix, i represents the historical sample number that kth model is corresponding According to classification；Utilize prior probability formula:Determine x_newBelong to the i-th class Prior probability, wherein, i=1 ..., k；Represent the jth sample of the i-th class historical sample data,Represent the i-th class history The predictive value of sample data jth sample；According to Bayesian formula, determine posterior probability: Wherein, P (i) is the number of samples ratio with total number of samples of the i-th class；By x_newIt is divided into that class that posterior probability is maximum.

In one embodiment, described corresponding to the data in each mode, corresponding based on Testing index and each mode Model carry out fault diagnosis, including: Testing index T2 and Q relevant for two kinds of quality based on KPLS method is synthesized one Index φ, for fault detect；Wherein, T²Represent HotellingT²Statistics, detection is event directly related with quality variable Barrier, Q represents the Q statistics in statistics, and detection is process noise；The fault diagnosis utilizing contribution rate method to realize quality relevant is asked Topic.

The fault detection method of the hot strip rolling quality that the embodiment of the present invention provides, utilizes hot strip rolling collection in worksite To the substantial amounts of data that can reflect production process establish the pass between production process variable and end product quality thickness System, for producing process monitoring, the quality important theory support of offer diagnosing and monitoring product and technical support.Solve existing In technology, product quality is controlled by the experience of oneself by the most skilled operator often so that stablizing of production process Property and safety can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that product matter The problem of amount.

The technology arrangement figure of the hot strip rolling that Fig. 3 provides for the embodiment of the present invention；

The fault detection method process flow diagram of the hot strip rolling quality that Fig. 4 provides for the embodiment of the present invention；

In order to make those skilled in the art be better understood from the fault of the hot strip rolling quality that the embodiment of the present invention provides Detection method, is now described in detail hot strip rolling production process.Hot strip rolling process be one extremely complex Industrial processes.The technology arrangement figure of whole process is as it is shown on figure 3, comprise on the whole with lower part: heating furnace, roughing, heat Runout table, coiling machine, flying shear, finish rolling and cooling facility for laminar flow and coiling machine etc..Heating furnace ensures that strip steel enter roughing Temperature can reach 1200 DEG C, board rolling typically can be become the intermediate blank of 28～45mm, the thus length of slab by rough rolling step Also can proportional increase, heat output roller way by strip steel from roughing district send into finish rolling district, carry out the most accurate rolling.In Fig. 3, Mm finishing mill unit is made up of 7 frames, and each frame comprises two working rolls and two support rollers, and each frame has driving of oneself Moving cell to provide power for working roll, and the distance between two working rolls is referred to as roll gap, and it is by adjustable supporting roll Hydraulic pressing is adjusted.The speed of roll is set as meeting the rolling requirements of last frame.Relative to roughing Process only one of which frame, 7 frames in finish rolling district mean whole strip institute to be passed through organic frame in finishing stands.Owing to adding The plastic coefficient of the strip steel after heat is substantially reduced, and the tension force of finish rolling zone steel uses mild tension condition, this avoid strip steel relatively Big extension and deformation.To this end, be fitted with tension force between each frame in finish rolling district to control ring, so control to make monoblock steel The transmission of plate safety and steady.

Whole course of hot rolling, the quality variable of care is thickness, width, template and outlet temperature etc., especially exports thickness Degree, it is the most key factor directly affecting product quality.Exit thickness is at last machine by X-ray calibrator The exit of frame is measured, and the whole operation of rolling is controlled by AGC system.But, the input of monitoring AGC system is outlet thickness Degree, this exists for an obvious Delay control, occurs fault X-ray the to be waited until calibrator in above frame to measure exception One-tenth-value thickness 1/10 could be controlled for thickness fault.So, setting up can process variable and the exit thickness of Real-time Collection Relation, and thickness is carried out quality-monitoring just become the most meaningful with fault diagnosis by measuring in real time process variable.

As shown in Figure 4, particular content is as follows:

1. mode division methods based on KFCC and BWP index

According to degree of membership size, when utilizing KFCC algorithm to cluster, mainly judged by the size of degree of membership, certain The value of the degree of membership of individual data point certain cluster corresponding is the biggest, then belong to a certain class.KFCC algorithm is by n sample X={x₁, x₂,…,x_nIt is divided into c ambiguity group, and then solving the cluster centre often organized, the object function of non-similarity index to be made reaches To minimum.The value of subordinated-degree matrix U is the element between [0,1].Under the constraint of normalizing condition, being subordinate to of a data set The summation of degree is equal to 1, i.e.Object function is:

$J_{u_{i,j}}=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{i,j}^{\mathrm{\δ}}{d}_{i,j}^2---\left(1\right)$

In formula, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_i,j=| | φ (x_j)-c_i||_EIt is Euclidean distance between i cluster centre and jth data point, and δ ∈ [1, ∞) it is Weighted Index.In order to obtain data acquisition system Good fuzzy division, utilizes lagrange's method of multipliers, tries to achieve the essential condition making formula (1) minimize value:

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^{\mathrm{\δ}}\mathrm{\φ}\left(x_j\right)}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{u}_{ij}^{\mathrm{\δ}}}---\left(2\right)$

$u_{i,j}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\Σ}}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(\mathrm{\δ}-1)}}---\left(3\right)$

When the present invention carries out mode division, taking KFCC algorithm to be because K-means algorithm is a kind of hard clustering algorithm, certain What individual sample was definite belongs to a certain class, either-or.And KFCC algorithm is the fuzzy division of a kind of flexibility, judging certain sample When originally belonging to the degree of certain cluster, degree of membership is mainly utilized to weigh, finally according to subordinated-degree matrix, according in fuzzy set Maximum subjection principle determines the classification of each sample, and cluster result so can be made more accurate.Especially for hot strip rolling The batch process of this complexity, the transient process between different modalities can make process data present more uncertainty, logical Cross this fuzzy method, the impact on monitoring accuracy can be reduced as far as possible.

When determining optimal mode number, the present invention considers from distance measure, introduces a kind of based on sample geometry Division BWP index in the m-class of Cluster Validity Index class:

(4)

In formula,Represent pth the sample of m class,Represent the q-th sample of jth class,Represent the of jth class I sample；(j, i) is defined as the infima species spacing of the i-th sample of jth class to b, and (j i) defines the i-th sample of jth class to w Inter-object distance.

What deserves to be explained is, described distance refers to the Euclidean distance in n dimension theorem in Euclid space, refers to sample the most in the present invention Average distance to other each apoplexy due to endogenous wind sample.

By this index, determine multi-modal optimal mode number in conjunction with clustering algorithm.Then, KFCC algorithm is referred to BWP Mark combines, it is proposed that a kind of evaluate Clustering Effect, the method determining preferable clustering number.Implement step as follows:

1) hunting zone [k of cluster numbers is selected_min,k_max]；

2) from k_minStart, to k_maxTerminate, run KFCC algorithm, calculated by the BWP desired value calculating single sample The BWP desired value of all samples；

3) determine preferable clustering number by the BWP desired value of relatively all samples, i.e. its maximum is preferable clustering number；

4) output cluster result, comprises effectiveness desired value and preferable clustering number.

Such as, this historical data has 4 mode, and each mode has 8 batches, and each batch contains 800 samplings.Can To find out that the method is not on the premise of having priori, it is possible to realize the consistent poly-of hot strip rolling production process historical data Class.

2. the fault detection and diagnosis method that quality based on the KPLS improved is relevant

After determining classification number, reasonably sorting data into, for each class data, utilize KPLS method to history sample The process variable X and quality variable Y of notebook data are modeled respectively, and by detection relevant for two kinds of quality based on KPLS method Index T²Synthesizing index φ with Q is:

Can simply be expressed as:

$\mathrm{\φ}=\frac{K(x_{new},x_{new})}{{\mathrm{\δ}}_r}+K_{new}^T{\mathrm{\ΩK}}_{new}---\left(6\right)$

In formula, Ω ∈ R^n×n.From first order Taylor and the kernel function gradient of φ function, define a kind of new Diagnostic method, i.e. contribution rate method.The method has the clearest and the most definite physical significance, can represent that each variable is to Testing index φ's Influence degree, each variable described, including roll gap, roll-force and bending roller force and the rack outlet thickness of frame.

Above method achieves the fault detect that quality is relevant, and next step is accomplished by diagnosing the variable causing this fault.This Invention have employed contribution rate method and achieves the troubleshooting issue that quality is relevant.

3. the online classification method mixed with Bayes's classification is returned based on KPLS

After the above data to different modalities establish model, when data x having new real-time online to gather_newTime, first First need to be classified as it corresponding apoplexy due to endogenous wind, corresponding model so could be utilized to carry out fault detect.The present invention mainly utilizes Bayes method realizes classification.Assume that historical sample data is divided into k class, utilize KPLS regression algorithm, k mould can be produced Type, each model has corresponding regression coefficient matrixFor new data x_new, first with corresponding average and association Data are normalized by covariance information, by following formula, can obtain corresponding forecast quality variate-value:

${\hat{y}}_{new}^{\left(i\right)}={\left(C_{KPLSR}^{\left(i\right)}\right)}^T{x}_{new}^{\left(i\right)}---\left(7\right)$

It is defined as follows prior probability formula:

$P\left(x_{new}\right|i)=prob\left(\right||{\hat{y}}_{new}^{\left(i\right)}-y_j^{\left(i\right)}|{|}_E\≤\left|\right|{\hat{y}}_j^{\left(i\right)}-y_j^{\left(i\right)}|{|}_E---\left(8\right)$

This equation gives x_newBelong to the prior probability of the i-th class, i=1 ..., k.WhereinRepresent the i-th class historical sample number According to jth sample,Represent the predictive value of the i-th class historical sample data jth sample.

According to Bayesian formula, posterior probability is represented by:

$P\left(i\right|x_{new})=\frac{P\left(x_{new}\right|i)P\left(i\right)}{\underset{i=1}{\overset{k}{\mathrm{\Σ}}}P\left(x_{new}\right|i)P\left(i\right)}---\left(9\right)$

In formula, P (i) is the number of samples ratio with total number of samples of the i-th class.When posterior probability corresponding to a certain class Time big, x_newJust that class is belonged to, it may be assumed that

$i^{opt}=\underset{i}{max}P\left(i\right|x_{new})---\left(10\right)$

4. hot strip rolling thickness qualities monitoring and fault diagnosis method

In the operation of rolling, the specification of the steel of different periods is different.Therefore, the present invention is by hot strip rolling production line Regard a multi-modal batch production process as, method set out above is applied to this production process thickness qualities monitoring with In fault diagnosis, achieving good monitoring effect, fault diagnosis accuracy is higher.

Implement step as follows:

1) hot strip rolling data are prepared

In the present invention, quality variable is finishing mill outlet thickness, and process variable mainly considers to affect the factor of exit thickness, Including average roll gap, roll-force and the bending roller force of seven frames, totally 20 process variables, wherein, the first frame is without roller, between sampling It is divided into 0.01s, as shown in table 1.

Table 1 finishing mill process variable and quality variable

2) hot strip rolling optimal mode number is determined

Utilize KFCC and BWP index, it is achieved that the division of hot strip rolling creation data mode.

3) online classification based on KPLS recurrence with the hot strip rolling production process data of Bayes's classification mixed method

After the above data to different modalities establish model, when there being new data x_newTime, need to utilize KPLS to return With the method for Bayes's classification mixing, it is classified as corresponding apoplexy due to endogenous wind, corresponding model so could be utilized to carry out fault detect.

4) the hot strip rolling production process fault detection and diagnosis that quality based on the KPLS improved is relevant

After historical sample data is divided into different modalities, utilize the modeling method of KPLS that the data of different modalities are divided It is not modeled.

Illustrate: consider two different failure conditions:

Situation 1: between sampling 1≤i≤1400, mode 2 does not has fault to occur；<between i≤2100, mode 1 exists in sampling 1401 Break down between 1601 to 1800 samplings；<between i≤3500, mode 3 does not has fault to occur in sampling 2101.

Situation 2: between sampling 1≤i≤3000, the faulty generation of mode 1.

1) for both the above situation, KPLS algorithm and PLS, the KPLS of the improvement that the present invention is proposed in fault detect Rate (FDR) aspect does a relative analysis, it can be seen that the FDR of the KPLS algorithm of improvement will be apparently higher than other two kinds of methods.Right More as shown in table 2 than analysis result.

Table 2 FDR relative analysis result

2) the KPLS algorithm by improving realizes the fault of the relevant online hot strip rolling production process thickness of quality Detection and diagnosis.

What deserves to be explained is, traditional PLS method is suitable for processing linear data, and due to along with computer, sensing technology Development, substantial amounts of data are collected and store, and there is serious non-linear relation between data, KPLS method be pass Kernel function is introduced, it is possible to process the non-linear relation of High dimensional space data well, and improve on the basis of the PLS method of system KPLS method be to continue to improve on the basis of traditional KPLS, the most traditional method T2 and SPE detect fault, and improve KPLS method by T2 and SPE synthesize a Testing index φ, it is possible to preferably detect dependent failure.

Testing index φ is applied in the fault detect that hot strip rolling thickness qualities is relevant, is imitated by the emulation of Fig. 5 Fruit is it can be seen that can detect fault at the 2350th sampling, and remains higher fault detect rate, for the event of latency Barrier diagnosis is taken a firm foundation.

After detecting that fault occurs, contribution rate method is applied to the fault diagnosis that hot strip rolling thickness qualities is relevant In.By the simulated effect of Fig. 6 it can be seen that variable 5 (roll-force of the 5th frame) and variable 17 (the 4th frame average Roll gap) there is maximum contribution rate, just it is consistent with the fault defined before.

What deserves to be explained is, in the present embodiment

As seen through the above analysis, the hot strip rolling thickness qualities monitoring and fault diagnosis method that the present invention proposes Establish the parsing relation between production process variable and end product quality thickness, it is possible to the most accurately realize hot strip rolling The online mode of creation data divides, and fault detect rate is higher, it is ensured that the quality of product, improves the economic benefit of enterprise.

The fault detection method of the hot strip rolling quality that the embodiment of the present invention provides, utilizes hot strip rolling collection in worksite To the substantial amounts of data that can reflect production process establish the pass between production process variable and end product quality thickness System, for producing process monitoring, the quality important theory support of offer diagnosing and monitoring product and technical support.Solve existing In technology, product quality is controlled by the experience of oneself by the most skilled operator often so that stablizing of production process Property and safety can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that product matter The problem of amount.

The structural representation of the trouble-shooter of a kind of hot strip rolling quality that Fig. 7 provides for the embodiment of the present invention.

As it is shown in fig. 7, another embodiment of the present invention provides the trouble-shooter of a kind of hot strip rolling quality, bag Include:

Collecting unit 71, for gathering the historical sample data in hot strip rolling production process, described historical sample number The data set in generation production process conventional according to including hot strip rolling, described data set includes roll gap, roll-force, bending roller force And rack outlet thickness.

Clustering unit 72, obtains multiple mode for carrying out clustering based on described historical sample data, and right Model should be set up in each mode.

Division unit 73 in real time, obtains the data set in the current production process of hot strip rolling in real time, and will in real time Data in the data set obtained are divided in the mode of correspondence.

Failure diagnosis unit 74, for corresponding to the data in each mode, based on Testing index and each mode pair The model answered carries out fault diagnosis.

In one embodiment, described clustering unit, including:

First builds module, for according to the Euclidean distance between subordinated-degree matrix and cluster centre and data point, structure Building the first object function of core Fuzzy C-cluster KFCC algorithm, in subordinated-degree matrix, each element value is between [0,1].

Second builds module, for building the second target according to the Lagrange multiplier of first object function and constraint formula Function minimizes the essential condition of value to obtain the second object function.

Clustering module, for according to described first object function, described second object function and described second mesh Scalar functions minimizes the essential condition of value, uses division BWP index in introducing the m-class of class to determine optimal mode number, and output is poly- Class result, described cluster result includes effectiveness desired value and preferable clustering number.

In one embodiment, described first object function:

$J(U,c_1,\mathrm{...},c_c)=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{J}_i=\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is to add Power index.

In one embodiment, described second object function:

$\begin{array}{c}\stackrel{\&OverBar;}{J}(U,c_1,\mathrm{...},c_c,{\mathrm{\&lambda;}}_1,\mathrm{...},{\mathrm{\&lambda;}}_n)=J(U,c_1,\mathrm{...},c_c)+{\mathrm{\&Sigma;}}_{j=1}^n{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{ij}-1)\\ =\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}\underset{j}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{ij}^m{d}_{ij}^2+\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\mathrm{\&Sigma;}}}{u}_{ij}-1)\end{array}$

Wherein, λ_j(j=1 ..., n) it is formulaThe Lagrange multiplier of n constraint formula, c_iFor The cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], d_ij=| | c_i-x_j| | for Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

In one embodiment, described second object function minimizes the essential condition of value and is:

$c_i=\frac{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{ij}^m{x}_j}{\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{u}_{ij}^m}$

With

$u_{ij}=\frac{1}{\underset{k=1}{\overset{c}{\mathrm{\&Sigma;}}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(m-1)}}$

Wherein, c_iFor the cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be [0,1] it Between, x_jFor jth data point, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, d_kj= ||c_k-x_j| | for the Euclidean distance between kth cluster centre and jth data point, and m ∈ [1, ∞) it is Weighted Index.

In one embodiment, described clustering unit, it is additionally operable to set up hot strip rolling process variable and exit thickness Parsing relation.

In one embodiment, described clustering unit, specifically it is additionally operable to, for each class data, utilize core latent structure to throw Process variable X and quality variable Y is modeled by shadow KPLS method respectively, and described process variable includes described data set.

In one embodiment, described real-time division unit, including:

First determines module, is used for utilizing KPLS regression algorithm, produce k (k=1 ..., 4) individual mode, determine each mode Corresponding regression coefficient matrixWherein, i represents the historical sample data classification that kth mode is corresponding, i=1 ..., k)

Prediction module, for the data set x for obtaining in real time_new, utilize corresponding average and covariance information to data It is normalized, obtains forecast quality variate-value:Wherein, x_newThe real-time online representing new gathers Data matrix, the transposition of T representing matrix, i represents the historical sample data classification that kth model is corresponding；

Priori module, is used for utilizing prior probability formula:Really Determine x_newBelong to the prior probability of the i-th class, wherein, i=1 ..., k；Represent the jth sample of the i-th class historical sample data,Represent the predictive value of the i-th class historical sample data jth sample；

Posteriority module, for according to Bayesian formula, determines posterior probability:Wherein, P (i) is the number of samples ratio with total number of samples of the i-th class；

Divide module in real time, for by x_newIt is divided into that class that posterior probability is maximum.

In one embodiment, described failure diagnosis unit, including:

Synthesis module, for synthesizing a finger by Testing index T2 and Q that two kinds of quality based on KPLS method are relevant Mark φ, for fault detect；Wherein, T²Represent HotellingT²Statistics, detection is the fault directly related with quality variable, Q represents the Q statistics in statistics, and detection is process noise；

Diagnostic module, for utilizing contribution rate method to realize the troubleshooting issue that quality is relevant.

The failure detector of the hot strip rolling quality that the embodiment of the present invention provides, utilizes hot strip rolling collection in worksite To the substantial amounts of data that can reflect production process establish the pass between production process variable and end product quality thickness System, for producing process monitoring, the quality important theory support of offer diagnosing and monitoring product and technical support.Solve existing In technology, product quality is controlled by the experience of oneself by the most skilled operator often so that stablizing of production process Property and safety can not be guaranteed, once break down, only by postponing delayed feedback control strategy it is difficult to ensure that product matter The problem of amount.

Claims (10)

1. the method for diagnosing faults of a hot strip rolling quality, it is characterised in that including:

Gathering the historical sample data in hot strip rolling production process, described historical sample data includes that hot strip rolling is conventional Generation production process in data set, described data set includes roll gap, roll-force, bending roller force and rack outlet thickness；

Carry out clustering based on described historical sample data and obtain multiple mode, and set up model corresponding to each mode；

Obtain the data set in the current production process of hot strip rolling in real time, and the data in the data set that will obtain in real time divide In corresponding mode；

Corresponding to the data in each mode, the model corresponding based on Testing index and each mode carries out fault diagnosis.

Method the most according to claim 1, it is characterised in that described carry out clustering based on described historical sample data Obtain multiple mode, including:

According to the Euclidean distance between subordinated-degree matrix and cluster centre and data point, build core Fuzzy C-cluster KFCC algorithm First object function, in subordinated-degree matrix, each element value is between [0,1]；

Lagrange multiplier according to first object function and constraint formula builds the second object function to obtain the second target letter Number minimizes the essential condition of value；

Necessity of value is minimized according to described first object function, described second object function and described second object function Condition, uses division BWP index in introducing the m-class of class to determine optimal mode number, exports cluster result, and described cluster result includes Validity Index value and preferable clustering number.

Method the most according to claim 2, it is characterised in that described first object function:

$J(U,c_1,...,c_c)=\underset{i=1}{\overset{c}{\&Sigma;}}{J}_i=\underset{i=1}{\overset{c}{\&Sigma;}}\underset{j}{\overset{n}{\&Sigma;}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijValue be between [0,1], c_iFor the cluster centre of ambiguity group i, d_ij=| | c_i-x_j|| For the Euclidean distance between ith cluster center and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

Method the most according to claim 2, it is characterised in that described second object function:

$\begin{array}{c}\stackrel{\&OverBar;}{J}(U,c_1,\mathrm{...},c_c,{\mathrm{\&lambda;}}_1,\mathrm{...},{\mathrm{\&lambda;}}_n)=J(U,c_1,\mathrm{...},c_c)+{\&Sigma;}_{j=1}^n{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\&Sigma;}}{u}_{ij}-1)\\ =\underset{i=1}{\overset{c}{\&Sigma;}}\underset{j}{\overset{n}{\&Sigma;}}{u}_{ij}^m{d}_{ij}^2+\underset{j=1}{\overset{n}{\&Sigma;}}{\mathrm{\&lambda;}}_j(\underset{i=1}{\overset{c}{\&Sigma;}}{u}_{ij}-1)\end{array}$

Wherein, λ_j(j=1 ..., n) it is formulaThe Lagrange multiplier of n constraint formula, c_iIt is fuzzy The cluster centre of group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], d_ij=| | c_i-x_j| | it is i-th Euclidean distance between individual cluster centre and jth data point, x_jFor jth data point, and m ∈ [1, ∞) it is Weighted Index.

Method the most according to claim 2, it is characterised in that described second object function minimizes the essential condition of value For:

$c_i=\frac{\underset{j=1}{\overset{n}{\&Sigma;}}{u}_{ij}^m{x}_j}{\underset{j=1}{\overset{n}{\&Sigma;}}{u}_{ij}^m}$

With

$u_{ij}=\frac{1}{\underset{k=1}{\overset{c}{\&Sigma;}}{\left(\frac{d_{ij}}{d_{kj}}\right)}^{2/(m-1)}}$

Wherein, c_iFor the cluster centre of ambiguity group i, u_ijIt is the element in subordinated-degree matrix U, u_ijValue be between [0,1], x_j For jth data point, d_ij=| | c_i-x_j| | for the Euclidean distance between ith cluster center and jth data point, d_kj=| | c_k- x_j| | for the Euclidean distance between kth cluster centre and jth data point, and m ∈ [1, ∞) it is Weighted Index.

Method the most according to claim 1, it is characterised in that described set up model corresponding to each mode, including:

Set up the parsing relation of hot strip rolling process variable and exit thickness.

Method the most according to claim 6, it is characterised in that described set up hot strip rolling process variable and exit thickness Parsing relation, including:

For each class data, core latent structure is utilized to project KPLS (Kernel Partial Least Squares, core latent structure Projection) process variable X and quality variable Y is modeled by method respectively, and described process variable includes described data set.

Method the most according to claim 1, it is characterised in that described data in described real time data group are divided into right In the mode answered, including:

Utilize KPLS regression algorithm, produce k (k=1 ..., 4) individual mode, determine each mode corresponding regression coefficient matrixWherein, i represents the historical sample data classification that kth mode is corresponding, i=1 ..., k)

For the data set x obtained in real time_new, utilize corresponding average and covariance information that data are normalized, obtain pre- Mass metering variate-value:Wherein, x_newRepresenting the data matrix that new real-time online gathers, T represents square The transposition of battle array, i represents the historical sample data classification that kth model is corresponding；

Utilize prior probability formula:Determine x_newBelong to the elder generation of the i-th class Test probability, wherein, i=1 ..., k；Represent the jth sample of the i-th class historical sample data,Represent the i-th class history sample The predictive value of notebook data jth sample；

According to Bayesian formula, determine posterior probability:Wherein, P (i) is the sample of the i-th class The ratio of number and total number of samples；

By x_newIt is divided into that class that posterior probability is maximum.

Method the most according to claim 1, it is characterised in that described corresponding to the data in each mode, based on detection Index and model corresponding to each mode carry out fault diagnosis, including:

Testing index T2 and Q that two kinds of quality based on KPLS method are relevant is synthesized index φ, for fault detect； Wherein, T²Represent HotellingT²Statistics, detection is the fault directly related with quality variable, and Q represents the Q system in statistics Meter, detection is process noise；

Contribution rate method is utilized to realize the troubleshooting issue that quality is relevant.

10. the trouble-shooter of a hot strip rolling quality, it is characterised in that including:

Collecting unit, for gathering the historical sample data in hot strip rolling production process, described historical sample data includes Data set in the generation production process that hot strip rolling is conventional, described data set includes roll gap, roll-force, bending roller force and machine Frame exit thickness；

Clustering unit, obtains multiple mode for carrying out clustering based on described historical sample data, and corresponding to every Individual mode sets up model；

In real time division unit, obtain the data set in the current production process of hot strip rolling in real time, and will obtain in real time Data in data set are divided in the mode of correspondence；

Failure diagnosis unit, for corresponding to the data in each mode, based on the mould that Testing index and each mode are corresponding Type carries out fault diagnosis.