CN111783354A - Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model - Google Patents

Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model Download PDF

Info

Publication number
CN111783354A
CN111783354A CN202010441538.1A CN202010441538A CN111783354A CN 111783354 A CN111783354 A CN 111783354A CN 202010441538 A CN202010441538 A CN 202010441538A CN 111783354 A CN111783354 A CN 111783354A
Authority
CN
China
Prior art keywords
flow
induced corrosion
neural network
tube bundle
network model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
CN202010441538.1A
Other languages
Chinese (zh)
Inventor
金浩哲
高帅棋
顾镛
刘骁飞
王超
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202010441538.1A priority Critical patent/CN111783354A/en
Publication of CN111783354A publication Critical patent/CN111783354A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Testing Resistance To Weather, Investigating Materials By Mechanical Methods (AREA)

Abstract

The invention discloses a flow-induced corrosion characteristic prediction and service life evaluation method of a RBF neural network model. The method comprises the steps of training sample data acquisition, testing sample data acquisition, building an RBF neural network model and utilizing the built RBF neural network model to carry out flow corrosion ammonium salt crystallization characteristic prediction analysis and life evaluation, so that the flow corrosion ammonium salt crystallization characteristic prediction and life evaluation of the high-risk tube bundle system are realized. The method can be used for rapidly and quantitatively predicting the crystallization temperature and the crystallization rate of the ammonium salt in the complex variable working condition environment, dividing a five-level flow induced corrosion risk matrix, calculating the residual service life of equipment and pipelines, providing scientific guidance for the safe closed-loop management of in-service inspection, risk evaluation, service life prediction, prevention and control optimization and the like of a high-risk tube bundle system, and promoting the safe, stable and long-period operation of the flow corrosion high-risk equipment system.

Description

Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model
Technical Field
The invention relates to flow-induced corrosion characteristic prediction and service life evaluation of a flow-type industrial cold-exchange equipment tube bundle, in particular to a flow-induced corrosion characteristic prediction and service life evaluation method of an RBF neural network model.
Background
The medium stored or conveyed by the petrochemical equipment system is flammable, explosive, toxic and harmful, has potential leakage and explosion risks, and directly harms public safety once a safety accident occurs. China is the first major country for importing and refining high-sulfur, high-acid, chlorine and heavy crude oil, along with the diversification of crude oil properties, the frequent change of working conditions and even overload operation process, petrochemical equipment generally faces the major potential safety hazards of strong corrosion and splicing equipment, and accidents such as unplanned shutdown, fire, explosion and the like caused by corrosion failure due to flow seriously threaten the daily safety production of refining enterprises, for example: in the 5 th month in 2013, 10 coastal petrochemical enterprises are subjected to unplanned shutdown for more than 40 times within one month of 30 sets of hydrogenation devices due to refining of high-chlorine crude oil, and the loss is heavy. A large number of failure case research results show that: the flow-induced corrosion failure of phase change condensation, ammonium salt crystallization, wall deposition, under-scale corrosion and the like of a cold exchange tube bundle in the transport process of the multi-component fluid containing easily crystallized components is a main cause of corrosion thinning, leakage perforation or tube explosion of the tube wall under the combined action of flowing, heat transfer, phase change and corrosion, the flow-induced corrosion characteristic mechanism is complex, the critical conversion condition is unclear, and accurate quantitative prediction is difficult to realize.
It can be known from the review of relevant documents at home and abroad that along with the increase of crude oil import quantity in China year by year, the deterioration degree of crude oil is deepened year by year, the content of corrosive media such as sulfur, nitrogen, chlorine and the like in raw materials is higher and higher, the high-risk cold tube bundle system is aggravated, particularly the corrosion risk caused by the flow of a hydrogenation reaction effluent heat exchanger and an air cooler is aggravated, and particularly, the ammonium salt crystallization corrosion is the most serious. However, at present, a better method and technology for the online monitoring and quantitative prediction of the flow-induced corrosion characteristics directly related to the ammonium salt crystallization corrosion at home and abroad is lacked. Although the flow induced corrosion temperature can be simulated through the process industrial software, the time consumption is long, the difference with the actual working condition is large, the timeliness is not realized, the flow induced corrosion characteristic prevention and control method and technology which are sequentially used as the basis are lack of instantaneity, and the protection effect is poor. Considering that safety accidents such as fire, explosion and the like of high-risk cold-exchange tube bundle systems such as hydrogenation reaction effluent heat exchangers, air coolers and the like of current petrochemical enterprises are related to flow-induced corrosion characteristics, the academic and engineering circles need an accurate prediction and scientific evaluation method for flow-induced corrosion temperature and flow-induced corrosion rate with short calculation time and high prediction precision, so that the flow-induced corrosion condition of equipment can be estimated more accurately, risk evaluation is carried out on the service life of high-risk cold-exchange tube bundle equipment, and the operation reliability of the equipment in the service process is improved.
Disclosure of Invention
In order to accurately predict the flow-induced corrosion temperature and the flow-induced corrosion rate of a high-risk cold exchange tube bundle system, the invention provides a flow-induced corrosion characteristic prediction and service life evaluation method of a RBF neural network model.
In order to achieve the purpose of the invention, the invention adopts the technical scheme that:
1) establishing an RBF neural network model;
2) training sample data collection
The acquisition of training sample data is to acquire real-time operation condition data of a Distributed Control System (DCS) of the high-risk cold-exchange tube bundle system and Laboratory Information Management System (LIMS) test analysis data, namely to acquire tube bundle inlet average flow velocity V1, tube bundle outlet average flow velocity V2, water injection quantity m1, raw material chlorine content m2, raw material nitrogen content m3, raw material sulfur content m4 and tube bundle inlet temperature T1 corresponding to different time sequences in a certain high-risk cold-exchange tube bundle system in real time, to use the seven variables as input variables of seven RBF neural network models, and to use the flow induced corrosion temperature TJ and the flow induced corrosion rate G as two output variables of the RBF neural network models;
the whole RBF neural network model is a multi-input multi-output model, input variables and output variables of training sample data are input into the RBF neural network model for optimization training, and the trained RBF neural network model is obtained.
In specific implementation, the test data set is also acquired, test verification is performed by using the test data set, and the acquisition of the test data set is also obtained by processing in the same way.
3) And processing real-time operation condition data and assay analysis data of the high-risk cold-exchange tube bundle system to be detected by using the trained RBF neural network model to obtain the results of flow-induced corrosion characteristic prediction and service life evaluation of the high-risk cold-exchange tube bundle system to be detected.
The RBF neural network model in 1) is a single hidden layer multilayer neural network structure, and comprises an input layer, a hidden layer and an output layer, wherein an I-P-O structure is adopted, namely the number of the input layer neural nodes is I, the number of the hidden layer neural nodes is P, the number of the output layer neural nodes is O, I is set to be 7, P is 5, O is 2, the hidden layer activation function is a Gaussian radial basis function, and the output layer activation function is a linear function.
In the step 2), the training sample data acquisition step specifically includes:
2.1) first, the average flow velocity V of the tube bundle inlet acquired in real time is measured1Tube bundle outlet average flow velocity V2Water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Performing discrete point outlier statistical analysis, removing any discrete point data with deviation exceeding 40% between the geometric mean value of the discrete points of the same data type adjacent to the discrete point outlier, performing data cleaning pretreatment, constructing training sample data of sequences at different moments, and dividing the training sample data into two groups;
2.2) then, the average flow velocity V at the inlet of the tube bundle1Tube bundle outlet average flow velocity V2Water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Inputting the temperature into the RBF neural network model to obtain the flow-induced corrosion temperature TJPrediction of sum flow induced corrosion rate GA value;
2.2.1) represent the sample data set of the RBF neural network model as a vector form of (X, y), where X ═ Xi1，Xi2，Xi3，Xi4，Xi5，Xi6，Xi7]T，y＝[yi1,yi2]T；
In the formula XinFor the nth feature of the ith sample, in the RBF neural network model, one input neural node represents one feature of the sample, and seven input variables are set as the feature of Xi1、Xi2Respectively represents the average flow velocity V of the inlet and outlet of the tube bundle1、V2The unit: m/s; xi3For the amount of water injected m1The unit: t/h; xi4As the raw material, the chlorine content m2In ppm; xi5As the raw material nitrogen content m3The unit: g/kg; xi6As raw material sulfur content m4The unit: percent; xi7For the inlet temperature T of the tube bundle1The unit: DEG C; y isiFor the output value corresponding to the ith sample, YimFor the m-th feature of the i-th sample, the feature values of two output variables, y, are seti1Temperature T for corrosion by flowJThe unit: DEG C; y isi2Flow induced corrosion G, unit: g/t;
2.2.2) setting the hidden layer to be a j-dimensional space vector consisting of radial basis functions
Figure BDA0002504172590000036
Selecting a Gaussian function as a kernel function of the RBF neural network model, and expressing as follows:
Figure BDA0002504172590000031
wherein u isjThe central vector of the kernel function of the neural node j is taken as the width function of the kernel function, x represents the characteristics of 7 input samples of the RBF neural network model, and P is the total number of the neural nodes of the hidden layer;
ujand the calculation is expressed as:
Figure BDA0002504172590000032
Figure BDA0002504172590000033
the min i and the max i respectively represent the minimum value and the maximum value of all input information of the ith characteristic in the training sample data, and j represents the jth hidden layer node; p is the total number of the hidden layer neural nodes; dmaxRepresenting a selected kernel function center vector ujMaximum euclidean distance of (a), arbitrary two vectors (x)1,x2,…,xn) And (y)1,y2,…,yn) The calculation method of the Euclidean distance d is as follows:
Figure BDA0002504172590000034
the hidden layer output matrix is:
Figure BDA0002504172590000035
weight vector omega from hidden layer neural node j to output layer neural node kjkAnd a bias vector cjkRespectively as follows:
ωjk＝[ωj1,ωj2]T，j＝1,2,…,P
cjk＝[cj1,cj2]T，j＝1,2,…,P
wherein, ω isj1,ωj2Respectively representing the component weight vectors from the jth hidden layer neural node to the 1 st and 2 nd output layer neural nodes, cj1,cj2Respectively representing bias vectors from a jth hidden layer neural node to a 1 st output layer neural node and a 2 nd output layer neural node, wherein T is a transposed symbol of the matrix;
2.2.3) prediction of the flow induced corrosion behaviour in a high risk cold exchanger tube bundle system is performed using the following formula, expressed as:
Figure BDA0002504172590000041
Figure BDA0002504172590000042
wherein y represents the flow induced corrosion temperature T obtained by predicting in step 2.2.3) for all groups of training sample data groupsJOr the prediction of the flow induced erosion rate G, k represents the group number of the training sample data packet.
For the flow induced corrosion temperature TJThe calculation flow is as follows:
a) testing of component NH in high-risk cold-exchange tube bundle system3Partial pressure of HCl and interpolation calculation of flow-induced corrosion temperature T under different working conditionsJAnd determining the convection induced corrosion temperature T by controlling the variable methodJHas a remarkable influence of that the water injection quantity m is about to be injected1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4And system pressure Psys；
b) Flow induced corrosion temperature TJThe average flow velocity of the inlet and the outlet of the tube bundle is not involved in the calculation process, and only the water injection quantity m is selected1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4And system pressure PsysCharacteristic parameter of (2) to the flow-induced corrosion temperature TJThe following relationship is established:
TJ1＝f1(m1,α)＝α1×m1+α2
Figure BDA0002504172590000043
Figure BDA0002504172590000044
Figure BDA0002504172590000045
Figure BDA0002504172590000046
wherein α, β, θ, η and τ are the first, second, third, fourth and fifth coefficient matrixes, respectively, α ═ α1,α2]，β＝[β1,β2]，θ＝[θ1,θ2]，η＝[η1,η2]，τ＝[τ1,τ2](ii) a f () represents the stream induced corrosion temperature TJα function of relation between characteristic parameters1,a2Representing a slope matrix and an intercept matrix, respectively, β1,β2Respectively representing a first amplitude matrix and a first exponential matrix, theta1,θ2Respectively representing a second amplitude matrix and a second index matrix, h1,h2Respectively representing a third amplitude matrix and a third exponential matrix, τ1,τ2Respectively representing a fourth amplitude matrix and a fourth exponential matrix;
fitting the above five formulas as the flow induced corrosion temperature TJThe calculation formula of (2):
Figure BDA0002504172590000047
A＝[m1,m2,m3,m4,Psys]
λ＝[λ1,λ2,…λ10]
wherein A is a characteristic parameter matrix, λ is a sixth coefficient matrix, λ1,λ2,…λ10Respectively represent a first coefficient to a tenth coefficient;
c) characteristic parameter matrix A and flow induced corrosion temperature T of each training sample data groupJSubstituting into step b) to train the flow-induced corrosion temperature T of the f (A, lambda) function valueJAnd predicting the obtained flow-induced corrosion temperature TJThe sum of the squares of the errors between them is minimal, and the optimal value to obtain the coefficient matrix λ satisfying the condition is determined, expressed as:
Figure BDA0002504172590000051
wherein, YkRepresenting the flow induced corrosion temperature T obtained by corresponding calculation of the kth training sample data groupJ，ykFlow induced corrosion temperature T obtained by prediction in step 2.2.3) for the kth set of training sample data packetsJ。
The flow-induced corrosion rate G and the flow-induced corrosion temperature TJThe calculation flow is the same.
In the step 3), after the flow-induced corrosion characteristic prediction of the high-risk cold exchange tube bundle system to be tested is obtained, the following processes are carried out to obtain a service life result:
3.1) predicting the flow-induced corrosion characteristic of the high-risk cold exchange tube bundle system to be tested to obtain the influence factor m which can obviously influence the flow-induced corrosion characteristic1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4System pressure PsysThen dividing a five-stage flow induced corrosion risk matrix as five characteristic parameters
The specific implementation obtains the water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4、PsysAnd then dividing the five-level flow induced corrosion risk matrix.
And 3.2) comprehensively evaluating the risk level by combining five characteristic parameters, and setting the comprehensive risk level as R:
R＝r(m4)+r(m3)+r(m2)+r(m1)+r(Psys)
wherein m is1、m2、m3、m4、PsysRespectively representing a first influencing factor, a second influencing factor, a third influencing factor, a fourth influencing factor and a fifth influencing factor which can obviously influence the flow-induced corrosion characteristics; r () represents the contribution function of each influencing factor to the risk level R;
3.3) the remaining life L of the device is calculated as:
Figure BDA0002504172590000052
wherein L ismaxDesign life for the apparatus, LsAnd (5) enabling the equipment to be in service for time.
In the specific implementation, a five-level flow induced corrosion risk matrix R belongs to [1, 5], and then R belongs to [5, 25 ].
The high-risk cold exchange tube bundle system particularly relates to a system containing NH3And multi-component fluid transportation of easily crystallized components such as HCl and the like.
The method comprises the steps of training sample data acquisition, testing sample data acquisition, building an RBF neural network model and utilizing the built RBF neural network model to carry out flow corrosion ammonium salt crystallization characteristic prediction analysis and life evaluation, so that the flow corrosion ammonium salt crystallization characteristic prediction and life evaluation of the high-risk tube bundle system are realized. Aiming at the problem of ammonium salt crystallization corrosion of a petrochemical process type industrial high-risk tube bundle system, the invention establishes a rapid quantitative prediction and residual life evaluation method of ammonium salt crystallization characteristics based on a RBF neural network model, can rapidly and quantitatively predict the crystallization temperature and crystallization rate of ammonium salt in a complex variable working condition environment, divides a five-level flow induced corrosion risk matrix, calculates the residual life of equipment and pipelines, can provide scientific guidance for the safety closed-loop management of in-service inspection, risk evaluation, life prediction, prevention and control optimization and the like of the high-risk tube bundle system, and promotes the safe, stable and long-period operation of the flow corrosion high-risk equipment system.
The invention has the beneficial effects that:
aiming at the problem of flow-induced corrosion failure of a flow-type industrial high-risk cold exchange tube bundle system, the RBF neural network model in the artificial intelligence technology is adopted to predict the distribution characteristics of flow-induced corrosion characteristic parameters, so that the flow-induced corrosion characteristic of a complex variable working condition environment can be rapidly and quantitatively predicted, and the problems that the traditional simulation calculation speed is low, the precision is low, and the prediction and early warning cannot be carried out on the flow-induced corrosion of the high-risk cold exchange tube bundle through flow process software are solved.
By adopting the life evaluation method based on the method, the flow-induced corrosion characteristic parameters can be accurately and quickly predicted only by inputting the relevant parameters on the established RBF neural network model, scientific guidance can be provided for safe closed-loop management such as in-service inspection, risk evaluation, life prediction, prevention and control optimization and the like of the high-risk cold exchange tube bundle system flow-induced corrosion in the petrochemical industry, and the safe, stable and long-period operation of an equipment system is promoted.
Drawings
FIG. 1 is a diagram of a model of a RBF-based neural network architecture constructed in accordance with the present invention;
FIG. 2 is a fitting scatter plot of predicted values of flow induced corrosion temperature in an embodiment of the present invention;
FIG. 3 is a comparison of predicted samples and actual samples in an embodiment of the present invention.
In the figure, 1, average flow velocity V of the tube bundle inlet1(ii) a 2. Average flow velocity V at outlet of tube bundle2(ii) a 3. Water injection quantity m1(ii) a 4. Chlorine content m of the raw material2(ii) a 5. Nitrogen content m of raw material3(ii) a 6. Sulfur content m of raw material4(ii) a 7. Tube bundle inlet temperature T1(ii) a 8. Temperature T of flow induced corrosionJ(ii) a 9. Flow induced corrosion rate G.
Detailed Description
The invention is further explained by the figures and the examples.
The examples of the invention are as follows:
fig. 1 is a block diagram of a RBF-based neural network model according to the present invention. Aiming at a specific high-risk cold tube bundle exchanging system, such as a hydrogenation reaction effluent heat exchanger or a hydrogenation reaction effluent air cooler, a Distributed Control System (DCS) is utilized to collect real-time operation condition data and Laboratory Information Management System (LIMS) test analysis data, and tube bundle inlet average flow velocity V corresponding to different time sequences is obtained1Tube bundle outlet average flow velocity V2Water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Respectively corresponding the 7 variables to 1, 2, 3, 4, 5, 6 and 7 in the figure 1 to be used as an input layer of the RBF neural network model; while maintaining the flow induced corrosion temperature TJThe flow induced corrosion rate G corresponds to 8 and 9 respectively and is used as an output layer of the RBF neural network model; the RBF neural network model is of a single hidden layer multilayer neural network structure and comprises 7 input layers in total, 5 hidden layer neurons in total, 2 output layer neurons in total, wherein the hidden layer activation function is a Gaussian radial basis function, and the output layer activation function is a linear function.
The method comprises two parts of training sample data acquisition and RBF neural network model establishment. The above section is detailed for the construction of the RBF neural network model, and the training sample data acquisition includes training data and test data acquisition. Firstly, collecting a flow-induced corrosion characteristic database related to flow-induced corrosion temperature, flow-induced corrosion rate and the like to construct a sample database, so as to realize the collection and pretreatment of sample data; then, constructing a neural network model based on RBF for flow corrosion characteristic prediction analysis; and finally, performing comparative analysis on a prediction sample obtained by prediction analysis and an actual sample, verifying the prediction precision and reliability of the constructed RBF neural network model, and providing a basis for service life evaluation of the high-risk cold-exchange tube bundle system.
Based on the high-risk cold-exchange tube bundle system in the field of petrochemical industry, such as a heat exchanger, an air cooler and an associated pipeline system, relevant data of a Distributed Control System (DCS) and a Laboratory Information Management System (LIMS) are collected in real time, the relevant data mainly comprise test analysis data of temperature, pressure, flow and raw materials, and NH in a unit volume is tested and calibrated3Partial pressure PNH3And partial pressure P of HClHClSo as to establish a flow induced corrosion characteristic database; real-time acquisition of water injection quantity m corresponding to time sequence from DCS1Tube bundle inlet temperature T1Total flow Q of tube bundle inlet and outlet1/Q2The chlorine content m of the raw material is acquired from LIMS2Nitrogen content m of the raw material3Sulfur content m of source stream4Calculating to obtain the average flow velocity V of the tube bundle inlet by combining the tube bundle cross section area A of the cold exchange tube bundle system1Tube bundle outlet average flow velocity V2And monitoring and recording in real time to form a state database.
Corresponding different moments acquired in real timeWater injection amount m1Average flow velocity V at tube bundle inlet1Tube bundle outlet average flow velocity V2Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1And performing discrete point outlier statistical analysis and rejecting abnormal outlier points. The abnormal outlier point data judgment method is as follows, and the current discrete point is assumed to be x, and data meeting the following conditions is considered to be abnormal data:
Figure BDA0002504172590000071
in the formula (I), the compound is shown in the specification,
Figure BDA0002504172590000072
represents the geometric mean of the five discrete points adjacent to the selected current discrete point x, namely:
Figure BDA0002504172590000073
m represents the position number where the discrete point is located. The specific physical value of x is the water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Average flow velocity V at tube bundle inlet1Or average flow velocity V at the outlet of the tube bundle2。
Calibration of NH in unit volume based on testing3Partial pressure PNH3And partial pressure P of HClHClSpecific value of (1), interpolation calculation of NH3Product W of partial pressure and partial pressure of HCl and flow induced corrosion temperature TJAnd performing multivariate function nonlinear fitting on the corresponding relation of the water injection quantity m under different conditions1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4System pressure PsysIsovariant and flow induced corrosion temperature TJThe associated expression of (1):
TJ＝-135.4364-0.2027m1+2.2947m2 1.2842+35.6813m3 -0.0215-52.4008m4 1.4459+426.6395·Psys 0.0348
based on a state parameter library formed by real-time monitoring records, 50 groups of data are randomly selected from the state parameter library to serve as a training sample data set, and the system randomly divides the training sample data set into 40 groups of training data and 10 groups of test data. Defining a sample format of a RBF neural network model front-end input as a form of (X, y), wherein X ═ Xi1，Xi2，Xi3，Xi4，Xi5，Xi6，Xi7]TWherein X isinFor the nth feature of the ith sample, the feature values of seven input variables, X, are seti1Is the average flow velocity V of the tube bundle inlet1The unit: m/s; xi2Is the average flow velocity V of the outlet of the tube bundle2The unit: m/s; xi3For the amount of water injected m1The unit: t/h; xi4As the raw material, the chlorine content m2The unit: ppm; xi5As the raw material nitrogen content m3The unit: g/kg; xi6As raw material sulfur content m4The unit: percent; xi7For the inlet temperature T of the tube bundle1The unit: DEG C. For simplicity of illustration, three sets of samples in the total training sample are taken for the following example:
Figure BDA0002504172590000081
in the formula, each column represents a set of training samples.
Definition y ═ yi1,yi2]TSetting the characteristic values of two output variables, yi1Indicating the temperature T of flow induced corrosionJThe unit: DEG C, yi2Represents the flow induced corrosion rate G, in units: g/t; at flow induced corrosion temperature TJFor example, the training sample output data is y ═ 195,188,179]。
Based on RBF neural network model, convection induced corrosion temperature TJAnd (6) performing prediction. Wherein the activation function of the hidden layer neural node is:
Figure BDA0002504172590000082
in the formula, a kernel function center vector u of a neural node jjThe width function of the kernel function is calculated as follows:
Figure BDA0002504172590000083
Figure BDA0002504172590000084
min i and max i respectively represent the minimum value and the maximum value of all input information of the ith characteristic in the training sample data; dmaxRepresenting the distance from the selected center point ujP is the total number of neurons in the hidden layer, and P is 5;
Figure BDA0002504172590000091
the weight vector and the bias vector for initializing the hidden layer to the output layer are respectively:
ωj1＝[187 195 203 211 219]T
cj1＝[-0.6952 0.6516 0.0767 0.9923 -0.8436]T
the prediction model of the induced corrosion temperature in the high-risk cold exchange tube bundle system is expressed as follows:
Figure BDA0002504172590000092
based on the steps, test data are input into the RBF neural network model for training to obtain the flow-induced corrosion temperature T of the high-risk cold exchange tube bundle systemJThe scatter fit of the prediction of (2) is shown in fig. 2. Fitting is carried out according to the prediction result to obtain a prediction curve of the flow induced corrosion temperature, wherein y is 0.9901x + 20207.
Similarly, the flow induced corrosion rate G can also be predicted in the same way.
For flow induced corrosion temperature TJThe following actual sample data may be adoptedThe method comprises the following steps:
a) testing and calibrating component NH in unit volume of high-risk cold exchange tube bundle system3Partial pressure of HCl PNH3、PHClDetermining NH3Product W of partial pressure and partial pressure of HCl and flow induced corrosion temperature TJThe corresponding relationship of (a) is as follows:
Figure BDA0002504172590000093
determining the flow induced corrosion temperature T under different working conditions by interpolationJ. Determination of a characteristic parameter for the temperature T of corrosion by convection by means of controlled variablesJThe tendency of influence of (2) is precisely on NH3HCl partial pressure PNH3、PHClSelecting characteristic parameters which have obvious influence on the induced corrosion temperature, namely: water injection quantity m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4System pressure Psys。
b) Convection induced corrosion temperature T according to the characteristic parameters of step a)JDetermining a functional relation:
TJ1＝f1(m1,α)＝α1×m1+α2
Figure BDA0002504172590000094
Figure BDA0002504172590000095
Figure BDA0002504172590000096
Figure BDA0002504172590000097
wherein α, β, θ, h, τ are coefficient matrices, α ═ α1,α2]，β＝[β1,β2]，θ＝[θ1,θ2]，h＝[h1,h2]，τ＝[τ1,τ2]. Fitting the above five formulas as the flow induced corrosion temperature TJThe calculation formula of (2):
Figure BDA0002504172590000101
a is a characteristic parameter matrix, and A ═ m1,m2,m3,m4,Psys](ii) a λ is coefficient matrix, λ ═ λ1,λ2,…λ10]。
c) A plurality of groups A and TJSubstituting into the above equation, the sum of squares of the error of the function value of f (a, λ) and the prediction data y is minimized, that is:
min∑(f(A,λ)-y)2
the lambda satisfying the above condition is the optimum parameter obtained by fitting the multivariate nonlinear function.
Similarly, the flow-induced corrosion rate G and the flow-induced corrosion temperature TJThe calculation flow is the same.
And (3) carrying out error analysis on the prediction result based on the flow induced corrosion characteristics and the actual sample data, and expressing as follows:
Figure BDA0002504172590000102
Ykrepresents the flow induced corrosion temperature T obtained by calculationJOr flow induced corrosion G, ykAnd predicting the obtained flow-induced corrosion characteristic data based on the RBF neural network model.
FIG. 3 shows the corrosion-causing temperature T under the same operating conditionsJThe result of the comparative analysis graph of the predicted data and the actual sample data shows that the predicted value of the flow-induced corrosion characteristic is well matched with the actual sample data value, and the established RBF neural network prediction model has high prediction precision.
The method for predicting the flow-induced corrosion characteristic and evaluating the service life of the RBF neural network model comprises the following steps of:
a) based on influencing factors of flow-induced corrosion behavior, i.e. on the amount of water injected m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4、PsysDividing a five-level flow induced corrosion risk matrix; according to the risk influence factors, the risk grades corresponding to the influence factors are divided, and the following table shows that:
Figure BDA0002504172590000103
the risk grades in the table are divided into 5 grades of 1, 2, 3, 4 and 5, and the larger the number is, the higher the risk grade is; and 5 risk influence factors are used, and the risk grade corresponding to each influence factor is subjected to interval division.
b) And comprehensively evaluating the risk level by combining 5 characteristic influence factors, and setting the comprehensive risk level as R:
R＝r(m4)+r(m3)+r(m2)+r(m1)+r(Psys)
c) according to five-level flow induced corrosion risk matrix r ∈ [1, 5]]It is known that R ∈ [5, 25]](ii) a Defining the residual life of the equipment as L, LmaxDesign life for the apparatus, LsFor a device active time, then L may be expressed as:
Figure BDA0002504172590000111
if the design life of a certain high-risk cold-exchange tube bundle system is known to be 5 years and the system is used for 2 years, the current equipment runs with the risk of corrosion caused by flow, and the water injection quantity m is assumed112t/h, chlorine content m of raw material22.5ppm, feed nitrogen content m35g/kg, raw material sulfur content m42.3%, system pressure PsysWhen the pressure is 10.8 MPa:
R＝r(m4)+r(m3)+r(m2)+r(m1)+r(Psys)＝2+3+2+3+3＝13
Figure BDA0002504172590000112
i.e. the equipment can run for 2.25 years under the current service condition. The method can help enterprises to realize the preliminary estimation of the residual life of the high-risk cold exchange tube bundle system, and help the enterprises to take relevant life-prolonging measures in time.

Claims (6)

1. A flow-induced corrosion characteristic prediction and service life assessment method of an RBF neural network model is characterized by comprising the following steps:
1) establishing an RBF neural network model;
2) training sample data collection
The acquisition of training sample data is to acquire real-time operation condition data of a Distributed Control System (DCS) of the high-risk cold-exchange tube bundle system and Laboratory Information Management System (LIMS) test analysis data, namely to acquire tube bundle inlet average flow velocity V1, tube bundle outlet average flow velocity V2, water injection quantity m1, raw material chlorine content m2, raw material nitrogen content m3, raw material sulfur content m4 and tube bundle inlet temperature T1 corresponding to different time sequences in a certain high-risk cold-exchange tube bundle system in real time, to use the seven variables as input variables of seven RBF neural network models, and to use the flow induced corrosion temperature TJ and the flow induced corrosion rate G as two output variables of the RBF neural network models; the whole RBF neural network model is a multi-input multi-output model, input variables and output variables of training sample data are input into the RBF neural network model for optimization training, and the trained RBF neural network model is obtained;
3) and processing real-time operation condition data and assay analysis data of the high-risk cold-exchange tube bundle system to be detected by using the trained RBF neural network model to obtain the results of flow-induced corrosion characteristic prediction and service life evaluation of the high-risk cold-exchange tube bundle system to be detected.
2. The method for predicting flow-induced corrosion characteristics and evaluating lifetime of an RBF neural network model as claimed in claim 1, wherein: the RBF neural network model in the step 1) is of a single hidden layer multilayer neural network structure and comprises an input layer, a hidden layer and an output layer, wherein an I-P-O structure is adopted, namely the number of input layer neural nodes is I, the number of hidden layer neural nodes is P, the number of output layer neural nodes is O, the hidden layer activation function is a Gaussian radial basis function, and the output layer activation function is a linear function.
3. The method for predicting flow-induced corrosion characteristics and evaluating lifetime of an RBF neural network model according to claim 1 or 2, wherein: in the step 2), the training sample data acquisition step specifically includes:
2.1) first, the average flow velocity V of the tube bundle inlet acquired in real time is measured1Tube bundle outlet average flow velocity V2Water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Performing discrete point outlier statistical analysis, removing any discrete point data with deviation exceeding 40% between the geometric mean value of the discrete points of the same data type adjacent to the discrete point outlier, performing data cleaning pretreatment, constructing training sample data of sequences at different moments, and dividing the training sample data into two groups;
2.2) then, the average flow velocity V at the inlet of the tube bundle1Tube bundle outlet average flow velocity V2Water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4Tube bundle inlet temperature T1Inputting the temperature into the RBF neural network model to obtain the flow-induced corrosion temperature TJAnd a predicted value of the flow induced corrosion rate G;
2.2.1) represent the sample data set of the RBF neural network model as a vector form of (X, y), where X ═ Xi1，Xi2，Xi3，Xi4，Xi5，Xi6，Xi7]T，y＝[yi1,yi2]T；
In the formula XinFor the nth feature of the ith sample, in the RBF neural network model, one input neural node represents one feature of the sample, and seven input variables are set as the feature of Xi1、Xi2Respectively represents the average flow velocity V of the inlet and outlet of the tube bundle1、V2The unit: m/s；Xi3For the amount of water injected m1The unit: t/h; xi4As the raw material, the chlorine content m2In ppm; xi5As the raw material nitrogen content m3The unit: g/kg; xi6As raw material sulfur content m4The unit: percent; xi7For the inlet temperature T of the tube bundle1The unit: DEG C; y isiFor the output value corresponding to the ith sample, YimFor the m-th feature of the i-th sample, the feature values of two output variables, y, are seti1Temperature T for corrosion by flowJThe unit: DEG C; y isi2Flow induced corrosion G, unit: g/t;
2.2.2) setting the hidden layer to be a j-dimensional space vector consisting of radial basis functions
Figure FDA0002504172580000021
Selecting a Gaussian function as a kernel function of the RBF neural network model, and expressing as follows:
Figure FDA0002504172580000022
wherein u isjThe central vector of the kernel function of the neural node j is taken as the width function of the kernel function, x represents the characteristics of 7 input samples of the RBF neural network model, and P is the total number of the neural nodes of the hidden layer;
ujand the calculation is expressed as:
Figure FDA0002504172580000023
Figure FDA0002504172580000024
the min i and the max i respectively represent the minimum value and the maximum value of all input information of the ith characteristic in the training sample data, and j represents the jth hidden layer node; p is the total number of the hidden layer neural nodes; dmaxRepresenting a selected kernel function center vector ujMaximum euclidean distance of (a), arbitrary two vectors (x)1,x2,…,xn) And (y)1,y2,…,yn) The calculation method of the Euclidean distance d is as follows:
Figure FDA0002504172580000025
the hidden layer output matrix is:
Figure FDA0002504172580000026
weight vector omega from hidden layer neural node j to output layer neural node kjkAnd a bias vector cjkRespectively as follows:
ωjk＝[ωj1,ωj2]T，j＝1,2,…,P
cjk＝[cj1,cj2]T，j＝1,2,…,P
wherein, ω isj1,ωj2Respectively representing the component weight vectors from the jth hidden layer neural node to the 1 st and 2 nd output layer neural nodes, cj1,cj2Respectively representing bias vectors from a jth hidden layer neural node to a 1 st output layer neural node and a 2 nd output layer neural node, wherein T is a transposed symbol of the matrix;
2.2.3) prediction of the flow induced corrosion behaviour in a high risk cold exchanger tube bundle system is performed using the following formula, expressed as:
Figure FDA0002504172580000031
Figure FDA0002504172580000032
wherein y represents the flow induced corrosion temperature T obtained by predicting in step 2.2.3) for all groups of training sample data groupsJOr the prediction of the flow induced erosion rate G, k represents the group number of the training sample data packet.
4. The method for predicting flow-induced corrosion characteristics and evaluating lifetime of an RBF neural network model as claimed in claim 3, wherein: for the flow induced corrosion temperature TJThe calculation flow is as follows:
a) testing of component NH in high-risk cold-exchange tube bundle system3Partial pressure of HCl and interpolation calculation of flow-induced corrosion temperature T under different working conditionsJ；
b) Only selecting the water injection amount m1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4And system pressure PsysCharacteristic parameter of (2) to the flow-induced corrosion temperature TJThe following relationship is established:
TJ1＝f1(m1,α)＝α1×m1+α2
Figure FDA0002504172580000033
Figure FDA0002504172580000034
Figure FDA0002504172580000035
Figure FDA0002504172580000036
wherein α, β, θ, η and τ are the first, second, third, fourth and fifth coefficient matrixes, respectively, α ═ α1,α2]，β＝[β1,β2]，θ＝[θ1,θ2]，η＝[η1,η2]，τ＝[τ1,τ2](ii) a f () represents the stream induced corrosion temperature TJα function of relation between characteristic parameters1,α2Representing a slope matrix and an intercept matrix, respectively, β1,β2Respectively representing a first amplitude matrix and a first exponential momentArray, theta1,θ2Representing a second amplitude matrix and a second index matrix, respectively, η1,η2Respectively representing a third amplitude matrix and a third exponential matrix, τ1,τ2Respectively representing a fourth amplitude matrix and a fourth exponential matrix;
fitting the above five formulas as the flow induced corrosion temperature TJThe calculation formula of (2):
Figure FDA0002504172580000037
A＝[m1,m2,m3,m4,Psys]
λ＝[λ1,λ2,…λ10]
wherein A is a characteristic parameter matrix, λ is a sixth coefficient matrix, λ1,λ2,…λ10Respectively represent a first coefficient to a tenth coefficient;
c) characteristic parameter matrix A and flow induced corrosion temperature T of each training sample data groupJSubstituting into step b) to train the flow-induced corrosion temperature T of the f (A, lambda) function valueJAnd predicting the obtained flow-induced corrosion temperature TJThe sum of the squares of the errors between them is minimal, and the optimal value to obtain the coefficient matrix λ satisfying the condition is determined, expressed as:
Figure FDA0002504172580000041
wherein, YkRepresenting the flow induced corrosion temperature T obtained by corresponding calculation of the kth training sample data groupJ，ykFlow induced corrosion temperature T obtained by prediction in step 2.2.3) for the kth set of training sample data packetsJ。
5. The method for predicting flow-induced corrosion characteristics and evaluating lifetime of an RBF neural network model as claimed in claim 3, wherein: in the step 3), after the flow-induced corrosion characteristic prediction of the high-risk cold exchange tube bundle system to be tested is obtained, the following processes are carried out to obtain a service life result:
3.1) predicting the flow-induced corrosion characteristic of the high-risk cold exchange tube bundle system to be tested to obtain the influence factor m which can obviously influence the flow-induced corrosion characteristic1Raw material chlorine content m2Nitrogen content m of the raw material3Sulfur content m of raw material4System pressure PsysThen dividing a five-stage flow induced corrosion risk matrix as five characteristic parameters
And 3.2) comprehensively evaluating the risk level by combining five characteristic parameters, and setting the comprehensive risk level as R:
R＝r(m4)+r(m3)+r(m2)+r(m1)+r(Psys)
wherein m is1、m2、m3、m4、PsysRespectively representing a first influencing factor, a second influencing factor, a third influencing factor, a fourth influencing factor and a fifth influencing factor which can obviously influence the flow-induced corrosion characteristics; r () represents the contribution function of each influencing factor to the risk level R;
3.3) the remaining life L of the device is calculated as:
Figure FDA0002504172580000042
wherein L ismaxDesign life for the apparatus, LsAnd (5) enabling the equipment to be in service for time.
6. The method for predicting flow-induced corrosion characteristics and evaluating lifetime of an RBF neural network model as claimed in claim 1, wherein: the high-risk cold exchange tube bundle system particularly relates to a system containing NH3And multi-component fluid transportation of easily crystallized components such as HCl and the like.
CN202010441538.1A 2020-05-22 2020-05-22 Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model Withdrawn CN111783354A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010441538.1A CN111783354A (en) 2020-05-22 2020-05-22 Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010441538.1A CN111783354A (en) 2020-05-22 2020-05-22 Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model

Publications (1)

Publication Number Publication Date
CN111783354A true CN111783354A (en) 2020-10-16

Family

ID=72753799

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010441538.1A Withdrawn CN111783354A (en) 2020-05-22 2020-05-22 Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model

Country Status (1)

Country Link
CN (1) CN111783354A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113239504A (en) * 2021-06-30 2021-08-10 西南石油大学 Pipeline corrosion defect prediction method based on optimized neural network
CN115330094A (en) * 2022-10-14 2022-11-11 成都秦川物联网科技股份有限公司 Intelligent gas pipeline service life prediction method, internet of things system, device and medium
CN115983116A (en) * 2022-12-22 2023-04-18 新疆敦华绿碳技术股份有限公司 Carbon dioxide miscible flooding corrosion detection method
US11879599B2 (en) 2022-12-16 2024-01-23 Chengdu Qinchuan Iot Technology Co., Ltd. Methods, Internet of Things systems, and mediums for assessing electrochemical corrosion of smart gas pipeline

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113239504A (en) * 2021-06-30 2021-08-10 西南石油大学 Pipeline corrosion defect prediction method based on optimized neural network
CN113239504B (en) * 2021-06-30 2022-01-28 西南石油大学 Pipeline corrosion defect prediction method based on optimized neural network
CN115330094A (en) * 2022-10-14 2022-11-11 成都秦川物联网科技股份有限公司 Intelligent gas pipeline service life prediction method, internet of things system, device and medium
US11898704B2 (en) 2022-10-14 2024-02-13 Chengdu Qinchuan Iot Technology Co., Ltd. Methods and Internet of Things systems for smart gas pipeline life prediction based on safety
US11879599B2 (en) 2022-12-16 2024-01-23 Chengdu Qinchuan Iot Technology Co., Ltd. Methods, Internet of Things systems, and mediums for assessing electrochemical corrosion of smart gas pipeline
CN115983116A (en) * 2022-12-22 2023-04-18 新疆敦华绿碳技术股份有限公司 Carbon dioxide miscible flooding corrosion detection method
CN115983116B (en) * 2022-12-22 2024-05-24 新疆敦华绿碳技术股份有限公司 Carbon dioxide miscible flooding corrosion detection method

Similar Documents

Publication Publication Date Title
CN111783354A (en) Flow-induced corrosion characteristic prediction and service life evaluation method of RBF neural network model
Wang et al. A two‐layer ensemble learning framework for data‐driven soft sensor of the diesel attributes in an industrial hydrocracking process
CN105069486B (en) Petrochemical equipment corrosion prediction method based on extreme learning machine
Shaik et al. Remaining useful life prediction of a piping system using artificial neural networks: A case study
CN107491840B (en) Flow wear characteristic prediction and service life evaluation method based on ELM neural network model
Shaik et al. Recurrent neural network-based model for estimating the life condition of a dry gas pipeline
CN113077039A (en) Task-driven RBF neural network-based water outlet total nitrogen TN soft measurement method
CN114841080A (en) Gas pipeline internal corrosion prediction method based on BP neural network and SD coupling
Amiri-Ramsheh et al. Modeling wax deposition of crude oils using cascade forward and generalized regression neural networks: Application to crude oil production
CN114692322A (en) Method, device and equipment for monitoring scaling of storage and heavy oil heat exchanger
Gu et al. Predictive study of flow-accelerated corrosion characteristic parameters based on the neural network
Loyola-Fuentes et al. Classification and estimation of unmeasured process variables in crude oil pre-heat trains subject to fouling deposition
Ye et al. Data‐driven soft‐sensor modelling for air cooler system pH values based on a fast search pruned‐extreme learning machine
Andrijić et al. Fouling detection in industrial heat exchanger using number of transfer units method, neural network, and nonlinear finite impulse response models
Dias et al. Forecasting the research octane number in a Continuous Catalyst Regeneration (CCR) reformer
Li et al. Petrochemical equipment corrosion prediction based on BP artificial neural network
Loyola-Fuentes et al. Fouling modelling in crude oil heat exchanger networks using data reconciliation and estimation of unmeasured process variables
Zhao et al. A machine learning methodology for reliability evaluation of complex chemical production systems
Ishiyama et al. Importance of data reconciliation on improving performances of crude refinery preheat trains
Nelsia Priya Dharsini et al. Probabilistic model development for estimating construction labor productivity optimization integrating with fuzzy logic approach systems
Loyola-Fuentes et al. Estimation of fouling model parameters for shell side and tube side of crude oil heat exchangers using data reconciliation and parameter estimation
Gudmundsson et al. Fouling detection in a cross-flow heat exchanger based on physical modeling
Mohammed et al. Predicting Oil and Gas Pipe Wall Thickness Using Machine Learning
CN111650894A (en) Bayesian network complex industrial process soft measurement method based on hidden variables
Zhang et al. Health Condition Assessment of Marine Systems Based on an Improved Radar Chart

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
WW01 Invention patent application withdrawn after publication

Application publication date: 20201016

WW01 Invention patent application withdrawn after publication