CN103413028A - Intelligent fault diagnosis method for polymerizing pot - Google Patents

Abstract

The invention relates to an intelligent fault diagnosis method for a polymerizing pot on the basis of the combination of a rough set with improved discernibility matrix attribute reduction and an improved LMBP neural network, and aims at meeting requirements of real-time fault diagnosis and optimization monitoring of the polymerizing pot in a polyvinyl chloride production process. An improved discernibility matrix algorithm is applied to the attribute reduction of the rough set, and dimensionality reduction is effectively performed on a large number of data; and fault diagnosis is performed with an improved Levenberg-Marquardt BP neural network according to a reduced decision table, and nonlinear mapping between a symptom set and a fault set is realized. The result of a fault diagnosis simulation experiment performed by combining historical data of an industrial site of the polymerizing pot shows the effectiveness of a fault diagnosis strategy of the neural network of the rough set.

Description

A kind of intelligent failure diagnosis method of polymeric kettle

Technical field

The present invention relates to a kind of polymeric kettle method for diagnosing faults, particularly relate to a kind of intelligent failure diagnosis method of polymeric kettle.

Background technology

Corvic (PVC) is important organic synthesis material, and its product has good physical property and chemical property, is widely used in the fields such as industry, building, agricultural, electric power, public utilities.Polymeric kettle is the key equipment of Production of PVC device, and can polymeric kettle stable operation be directly connected to the operation conditions of whole Production of PVC device.And motor, reductor and machine envelope are to guarantee the key equipment of the normal operation of polymeric kettle device, in a single day they break down and will bring serious loss to production.

Summary of the invention

The object of the present invention is to provide a kind of intelligent failure diagnosis method of polymeric kettle, the method that the method combines with improved LMBP neural network by the improvement differential matrix realizes the fault diagnosis of polymeric kettle, diagnose early polymeric kettle fault type and position, the tremendous economic loss of avoiding the polymeric kettle parking to cause, improve simultaneously the polyvinyl chloride product quality, reduce production costs.

The objective of the invention is to be achieved through the following technical solutions:

A kind of intelligent failure diagnosis method of polymeric kettle, be a kind of rough set based on improving the differential matrix attribute reduction and the method for diagnosing faults that improvement LMBP neural network combines, and said method comprising the steps of:

Step 1: image data, the data of correlated variables in gatherer process, for each fault, produce two groups of data, i.e. training data and real-time working condition data; Training data is for setting up model, and the real-time working condition data are for on-line monitoring; And the data that gather by the standardized method standard;

Step 2: adopt the rough set method of improving differential matrix to carry out similar yojan to data, remove the data that similarity is stronger;

Step 3: utilize improved LMBP to train one group of fault sample, to determine the structure and parameter of network;

Step 4: the pattern to fault is classified, and according to one group of given sign, realizes that the sign collection is to the Nonlinear Mapping between the fault collection;

Step 5: according to the LMBP network trained, test data is carried out to emulation, the observation Output rusults is fault diagnosis result.

Advantage of the present invention and effect are:

1. the front-end system using rough set method as neural network, simplify the complicacy of nerve network system, improves efficiency and the precision of fault diagnosis.

2. use improved LMBP neural network as rearmounted information identification system, be widely used in the industrial circles such as electric power, metallurgy and chemical industry.

3. the combination of two kinds of methods has more improved the accuracy and efficiency of polymeric kettle fault diagnosis.

Embodiment

The present invention is described in detail below in conjunction with embodiment.

One, at first adopt the rough set attribute reduction method that improves differential matrix to carry out yojan to a large amount of training datas, wherein, be described below based on the old attribute reduction algorithms of improving differential matrix

Step 1: the differential matrix that calculates decision table C _D

Step 2: find out the single element in differential matrix, retain this element element, and this element element is the core of attribute reduction, and other is contained to all elements c of this single element _IjChange 0 into;

Step 3: for all values in differential matrix, be non-zero, 1 element c _Ij(c _Ij

0, c _Ij

1), set up the logical expression of extracting accordingly:

（1）

Step 4: by all logical expression L that extracts _IjCarry out the conjunction computing, obtain conjunctive normal form:

（2）

Step 5: the form that conjunctive normal form L is transferred to disjunctive normal form:

（3）

Step 6: output attribute yojan result.The conditional attribute set after into about letter of the result of the corresponding attribute reduction of each conjunct in disjunctive normal form, the set of properties comprised in each conjunct.

The improvement algorithm of this paper is to have increased step 2 on the basis of canonical algorithm, namely in differential matrix, find out single element, retain this element element, and this element element is the core of attribute reduction, the value that comprises the element of core attribute in differential matrix all is modified to 0, obtains a new differential matrix.

Two. then adopt improvement algorithm LMBP in neural network to train these data, build neural network model, can utilize the LU direct breakdown method to carry out the symmetric triangular decomposition to A.Reduced greatly the calculated amount of LMBP.It is below the development of LMBP algorithm

Tradition LMBP algorithmic procedure is as follows: establishing the error objective function is

（4）

Wherein:

(5)

For the network error vector,

For error vector.

By Newton method

(6)

(7)

Although having, Newton method restrains advantage rapidly, owing in each iterative computation, can not guaranteeing the Hessian matrix

All reversible, available

The approximate replacement

, in formula

For

Jacobi (Jacobian) matrix.

For

Hession matrix (error matrix).

（8）

Can prove (9)

When separating near extreme point

(10)

(11)

(8) formula is improved to the mixed form that makes it not only comprise Gauss-Newton method but also have gradient descent method.Formula is

(12)

In formula,

For unit matrix,

For scale-up factor, if

Close to 0 o'clock, be Gauss-Newton method, if Be worth when larger, be similar to gradient descent method

, common adjustment strategy is that algorithm is while starting Get one little on the occasion of, if a certain step can not reduce error order function

Value,

Be multiplied by one and be greater than 1 the stepping factor

, namely

If a certain step has produced less

,

Next step divided by

, namely

.

The improvement of LMBP algorithm is as follows:

The LMBP algorithm is furtherd investigate, found the matrix wherein related to

Be the principal element that affects its convergence, remove matrix inversion operation consuming time by using the LU direct breakdown method, reduced greatly the calculated amount of LMBP.

Can make

(13)

(14)

(15)

(9)

(16)

Can utilize LU direct breakdown method pair

Carry out the symmetric triangular decomposition.Ask

Problem just be equivalent to and obtain

(17)

Basis again

(18)

Have again

(19)

Can obtain

.

The use LU factorization solves

Do not need finding the inverse matrix, now only need

Inferior multiplication and division computing, arithmetic speed can improve more than three times.Due to the minimizing of operation times, not only

To save operation time, can also reduce round-off error, therefore, this improvement makes the value of calculating more accurate.

In actual computation, along with

Increase, the problem of little step-length can occur, cause an iteration in the circulation of small step, to need circulation repeatedly, the flower long time could finish.In order to address this problem, by original fixing

Value is designed to variable step mode, i.e. step factor

It is a variable amount.The variable step formula is defined as

（20）

In formula

For only entering the number of times of this step partial circulating,

For adjusting variable.

If a certain step can not reduce error order function

Value, Be multiplied by a new stepping factor

, If a certain step has produced less

,

At next step divided by the new stepping factor

, namely

.

Three. after model buildings success, the test data below the yojan that uses the same method, with the model training of putting up, watch Output rusults.

Claims (1)

1. the intelligent failure diagnosis method of a polymeric kettle, for a kind of rough set based on improving the differential matrix attribute reduction and the method for diagnosing faults that improvement LMBP neural network combines, is characterized in that, said method comprising the steps of:

Step 1: image data, the data of correlated variables in gatherer process, for each fault, produce two groups of data, i.e. training data and real-time working condition data; Training data is for setting up model, and the real-time working condition data are for on-line monitoring; And the data that gather by the standardized method standard;

Step 2: adopt the rough set method of improving differential matrix to carry out similar yojan to data, remove the data that similarity is stronger;

Step 3: utilize improved LMBP to train one group of fault sample, to determine the structure and parameter of network;

Step 4: the pattern to fault is classified, and according to one group of given sign, realizes that the sign collection is to the Nonlinear Mapping between the fault collection;

Step 5: according to the LMBP network trained, test data is carried out to emulation, the observation Output rusults is fault diagnosis result.