CN109781395A - A kind of cylinder creep detection and life-span prediction method based on DENSENET - Google Patents
A kind of cylinder creep detection and life-span prediction method based on DENSENET Download PDFInfo
- Publication number
- CN109781395A CN109781395A CN201811615955.2A CN201811615955A CN109781395A CN 109781395 A CN109781395 A CN 109781395A CN 201811615955 A CN201811615955 A CN 201811615955A CN 109781395 A CN109781395 A CN 109781395A
- Authority
- CN
- China
- Prior art keywords
- cylinder
- creep
- life
- data
- network
- 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.)
- Granted
Links
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
- Testing Resistance To Weather, Investigating Materials By Mechanical Methods (AREA)
Abstract
The invention discloses a kind of cylinder creep detection and appraisal procedure based on DENSENET, this method is for detecting whether cylinder occurs creep, if creep failure has occurred, then shutdown inspection, and the follow-up operation state of cylinder is paid close attention to, prevent safety accident；Secondly, predicting remaining creep life and repairing inspection plan according to remaining life reasonable arrangement.Temperature, pressure and the strain information that the present invention passes through acquisition cylinder, it is translated into picture, establish creep detection and life prediction task of the fine tuning DENSENET network implementations to input cylinder measurement data, allow common operating personnel while guaranteeing not destroy cylinder body, immediately it effectively detects the creep behavior of cylinder and predicts creep life, reasonable arrangement unit operational plan, extends the service life of cylinder, ensure that the safe and effective operation of unit.
Description
Technical field
The invention belongs to steam turbine technology fields, and in particular to a kind of cylinder creep detection and service life based on DENSENET
Prediction technique.
Background technique
With the fast development of power industry, to the single-machine capacity of power plant, steam inlet condition and safety etc. are proposed more
High requirement, steam-turbine unit gradually develop towards the large-scale direction of high parameter.Steam turbine long service is inevitable
Meeting generate creep impairment, the high-temperature component mainly influenced by creep impairment includes super-pressure/high pressure/middle pressure rotor, superelevation
Room and reheating channel etc. occur for pressure/high pressure/middle pressure outer shell, super-pressure/high pressure/middle pressure steam.High-temperature component uses high temperature resistant
Material, cost is high and is not easily repaired, once creep impairment occurs, exists to the safe and reliable operation of entire unit equipment serious
Threat.
There is not direct effective method yet for the detection of steam-turbine unit creep impairment and life prediction at present, it is main
Detection means is to damage test and two kinds of non-destructive testing.Damage test and need to acquire the sample of material of cylinder, improve temperature and
In the case where pressure under evaluation work temperature and pressure cylinder creep life, this method is also referred to as outside the creep rupture test service life
Pushing manipulation mainly has thermoisopleth extrapolation, Larson-Miller parametric method and theta function method.But the detection method is needed in cylinder
Multiple positions are sampled test, in practical projects extremely difficult realization, still in experimental stage.Non-destructive testing is mainly ultrasonic wave
Detection carries out creep impairment monitoring and assessment by detection material internal micro-fractures.Existing ultrasonic detection technology is main
Applied to simple materials such as copper, if can be applied to steam turbine, there is still a need for further researchs, and ultrasonic detection technology misses
Difference is larger.
With the promotion of computer computation ability and the development of artificial intelligence, more and more traditional industry problems use people
Work intelligent algorithm has obtained new solution.The cylinder creep based on DENSENET that the invention proposes a kind of detects and the service life
It is compacted to obtain cylinder by DENSENET Web Mining data information in high pressure cylinder arrangement measuring point acquisition data for prediction technique
Change state and remaining life.The advantage of this method is that, on the one hand, operator is not required to it is to be understood that the correlation in terms of creep is known
Know, the operation of unit can be adjusted according to state according to the measurement data real-time monitoring of cylinder, be overhauled immediately to guarantee
The stable operation of unit safety, on the other hand, relative to traditional scheme, the program does not need to destroy cylinder collecting sample, not yet
The experiment instrument of complicated profession is needed, while accuracy rate is higher, it is more immediately reliable.
Summary of the invention
The cylinder creep detection and life-span prediction method that the purpose of the present invention is to provide a kind of based on DENSENET, the party
Temperature, pressure and strain information of the method based on cylinder carry out creep detection and life prediction, Ke Yiyou using DENSENET network
Effect improves detection efficiency, and calculating time cost is greatly reduced, and detects creep state immediately and predicts remaining creep life.
The present invention adopts the following technical scheme that realize:
A kind of cylinder creep detection and life-span prediction method based on DENSENET, comprising the following steps:
1) in cylinder disposed inboard measured zone, temperature sensor, pressure sensor is arranged in measured zone respectively and is answered
Become sensor, obtains the data that the temperature, pressure and strain of single cylinder change over time；
The data for needing to acquire include the measurement result under normal operating condition and under creep extended mode, and to normal
Operating status is marked with creep extended mode measurement result, Nn,i,j,tFor the measurement number under certain cylinder normal operating condition
According to Cn,i,j,tFor the measurement data under the cylinder creep extended mode, Cn,i,j,tAnd Nn,i,j,tIt corresponds, the cylinder creep is total
Service life is TLifen, format is { Y-M-D-H-MM-S }n, i.e. year-month-day-when-point-second }n；Wherein n=1,2,3..., N, N
For measurement cylinder, total, i=1,2,3, have respectively represented temperature data, pressure data and strain data, j=1,2,3..., J, J
For single cylinder measured zone sum, t=1,2,3..., T, T is the sum of time of measuring point, and each t is one corresponding
Specific time point, with { Yt-Mt-Dt-Ht-MMt-St}nIt indicates, i.e. { yeartThe moontDaytWhentPointtSecondt}n, all temperature,
Pressure is phase measurement result in the same time with strain data；
2) cylinder measurement data pre-processes, and temperature, pressure and strain measurement data are normalized respectively；
By collected cylinder measurement data Cn,i,j,tAnd Nn,i,j,tThe cylinder that M length is L is cut at random respectively to measure
Short and small data set Cn′×m,i,j,tAnd Nn′×m,i,j,t, wherein m=1,2,3..., M, M are the short and small data count of single cylinder, t=
1,2,3..., L, L are the length of short and small data, i.e. the sum at time point, i and j is remained unchanged, and respectively indicate the property of measurement data
The measured zone sum of matter, temperature, pressure and strain and cylinder；To all cylinder measurement data { N 'i,j,t}n×mWith
{C′i,j,t}n×mIt is normalized, three kinds of data temperatures, pressure and strain correspond to picture rgb pixel value, and accordingly will
Data are converted into picture；
3) cylinder measurement data label is set；
Measurement data { the N ' of normal operating conditioni,j,t}n×mLabel there are two, be class label Class respectivelyn×m=
0, it indicates to operate normally, service life label Lifen×m=TLifen-OLifen×m, indicate the remaining life of cylinder this moment, wherein
TLifenIndicate the creep entire life of corresponding measured cylinder, OLifen×mIndicate corresponding cylinder runing time；Creep extended mode
Measurement data { C 'i,j,t}n×mClass label Classn×m=1, indicate that cylinder has been in the state of creep extension, service life
Label Lifen×m=TLifen-OLifen×m, indicate the remaining life of cylinder at this time；
4) training set is divided for cylinder measurement data and verifying collects；
Network is detected for creep, is inputted as X2n×m={ N 'i,j,t}n×m+{C′i,j,t}n×m, X will be inputted2n×mUpset at random,
According to the ratio of 4:1 by X2n×mIt is divided into training set (X2n×m×0.8)trainCollect (X with verifying2n×m×0.2)validate, corresponding label is
(Class2n×m×0.8)trainWith (Class2n×m×0.2)validate；For predicting network, the processing method of input and creep detect net
Network is identical, X '2n×m={ N 'i,j,t}n×m+{C′i,j,t}n×m, X ' will be inputted2n×mUpset at random, is classified as according to the ratio of 4:1
Creep life predicts network training collection (X '2n×m×0.8)trainCollect (X ' with verifying2n×m×0.2)validate, corresponding label is surplus for cylinder
Remaining service life (Life2n×m×0.8)trainWith (Life2n×m×0.2)validate；
5) cylinder creep detection network and creep life prediction network based on DENSENET are built respectively；
DENSENET network structure is adjusted according to the pond layer parameter of picture size classification layer first；Secondly, modification cylinder
The full articulamentum structure of creep detection and life prediction network detects network for creep, modifies the output of wherein full articulamentum
Categorical measure is 2, respectively normal operating condition and creep failure state；Network is predicted for creep life, is modified wherein complete
The output categorical measure of articulamentum is 1, as remaining life；
6) creep detection network and creep life based on DENSENET is trained to predict network；
7) creep detection and life prediction；
Cylinder measurement data is handled according to step 1), step 2) and step 3) respectively, the vapour after being normalized
Cylinder test data is translated into the form of picture；It is input with cylinder measurement picture, first using creep detection network judgement
Whether the steam turbine cylinder is in normal operating condition at this time, if sorter network output result is 0, cylinder is normally transported
Row, otherwise, cylinder has been in creep extended mode at this time, needs instant shutdown inspection, and carry out to the operating status of cylinder
Observation closely, to prevent serious accident occurs；Secondly, obtaining the remaining longevity of the cylinder at this time using creep life prediction network
Life arranges the inspection of repairing of unit, adjustment operation according to remaining life.
A further improvement of the present invention lies in that further comprising the steps of:
8) algorithm is safeguarded；
During practical application, if arrangement effective measuring area is less than estimated arrangement measured zone or according to need
Increase measured zone is sought, then is handled measurement data according to step 1), 2), 3), use is trained to be finished
DENSENET network restarts to train on this basis as pre-training model to whole neural network.
A further improvement of the present invention lies in that according to the shape and operating status of cylinder, existing respectively first in step 1)
Steam turbine disposes J according to circumferencial direction into vapour section middle section respectively1A measured zone, in the exit portion into vapour section
It is evenly arranged J2A measured zone；Secondly as upper and lower cylinder is bound up by traditional cylinder using bolt, it is coupled in cylinder
Split on be uniformly distributed J3A measured zone；Finally, being evenly arranged J respectively at the both ends of cylinder according to cylinder shape4It is a
Measured zone；According to analysis of finite element method as a result, find respectively the highest part of cylinder temperature, stress the best part and
The best part is strained as numerical analysis measured zone, i.e., temperature sensor, pressure sensor and the strain of each cylinder pass
Sensor is distributed total J=J1+J2+J3+J4× 2+3 measured zone；
Normal operating condition is that cylinder is on active service the stage for starting to cylinder first macroscopical engineering crackle occur；And creep is expanded
Exhibition state is from there is first macroscopical engineering crackle to the stage for first critical crack occur；Macroscopical engineering crackle is length
The crackle for being 0.1~0.15mm for 0.3~0.5mm, depth；Critical crack is the crackle that depth reaches 5mm, and cylinder has pole at this time
Big possibility is broken；
Temperature sensor, pressure sensor and strain transducer are sampled with identical sample frequency f, are guaranteed all
Data obtain under synchronization, record sampling instant t, and format is { Yt-Mt-Dt-Ht-MMt-St}n, obtain respectively original
Data Cn,i,j,tAnd Nn,i,j,t；It is since the temperature of cylinder under operation and pressure change are smaller, then primary every acquisition in 5 seconds
Measurement result, sample frequency 1/5Hz.
A further improvement of the present invention lies in that in step 2), the raw data set C of cylinder measurement datan,i,j,tAnd Nn,i,j,t
The length of individual data is T, the i.e. sum of time of measuring point, for the measurement data { C of each cylinderi,j,t}nWith
{Ni,j,t}n, take M start time point t at random between 0≤t of time point≤T-L respectively0, prolong L time point of Shen, t backward0≤t
≤t0Data in+L are a short and small data, obtain the cylinder that N × M length is L and measure short and small data set C 'n×m,i,j,tWith
N′n×m,i,j,t；
Wherein, L value range is 15~30min, i.e. 160~360 time points.
A further improvement of the present invention lies in that in step 2), with the data set { C ' under creep extended modei,j,t}n×m's
Data instance, the data { N ' under normal operating conditioni,j×k,t}n×mSimilarly, fixed n, m and i, normalized method are as follows:
{Maxi}n×m=Max [({ C 'i,j,t}n×m|0≤j≤J,0≤t≤L)}]
{Mini}n×m=Min [({ C 'i,j,t}n×m|0≤j≤J,0≤t≤L)]
Wherein, { C 'i,j,t}n×mIndicate the data before normalization,Indicate the data after normalization,
{ui}n×mIndicate the average value of all data of i-th kind of data character, { Maxi}n×mIndicate all data of i-th kind of data character
Maximum value, { Mini}n×mIndicate the minimum value of all data of i-th kind of data character, Max_value indicates model after normalization
The upper limit enclosed is herein the rgb pixel value upper limit 255；Min_value indicates the lower limit of range after normalization, is herein rgb pixel
It is worth lower limit 0；In order to facilitate describing and indicate, the data after normalization use { Ci,j,t}n×mIt indicates.
A further improvement of the present invention lies in that in step 3), service life be cylinder measurement data paragraph the last one when
Between at the time of put corresponding, i.e.,
OLifen={ YL-ML-DL-HL-MML-SL}n
Lifen×m=TLifen-OLifen×m
={ Y-M-D-H-MM-S }n-{YL-ML-DL-HL-MML-SL}n×m。
A further improvement of the present invention lies in that in step 5), creep detection and life prediction network based on DENSENET
A part of fine tuning, original DENSENET are carried out in traditional DENSENET network structure according to the feature of cylinder measurement data
Network has carried out 5 size reductions to picture altogether, using the step-length in convolution and pond step, each time by the width of picture and
Height reduces half, in order to enable cylinder measurement data is able to use in DENSENET network, it will be according to cylinder measurement data
The height J and width L for being converted into picture are adjusted network as adjustment standard:
A) according to the sum of the measured zone of cylinder measurement data, it is first determined whether needing to adjust pond layer, such as 1 institute of table
Show, the step-length selection of pond layer is divided into three kinds of situations:
As the total J >=32 of cylinder measured zone, the convolution 1 of network, Chi Hua 1, transition zone 2-4 uniform pond in walk
Length is disposed as [1,2,2,1]；
As sum 16≤J < 32 of cylinder measured zone, the convolution 1 of network, Chi Hua 1, transition zone 3-4 uniform pond
The step-length changed in step is set as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2 is set as [1,1,2,1]；
As sum 8≤J < 16 of cylinder measured zone, in the uniform pond step of convolution 1, Chi Hua 1 and transition zone 4
Step-length is disposed as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2-3 is set as [1,1,2,1]；
Wherein step-length [1,2,2,1] indicates that convolution or Chi Huahe carry out in the both direction of picture height and width
The operation that step-length is 2, and [1,1,2,1] indicates that convolution or Chi Huahe only carry out step-length on this direction of the width of picture
For 2 operation, and the operation that step-length is 1 is carried out on this direction of picture height；
B) according to dimension of picture and network structure, the Chi Huahe size of pondization 2 is determined:
C network) is detected for creep, in last full articulamentum, exports and is set as 2, and connect softmax classifier；
And creep life predicts network, in last full articulamentum, exports and is set as 1, the as remaining life of cylinder.
A further improvement of the present invention lies in that during training network, setting optimizer first is in step 6)
Adam, initial learning rate are set as 0.01,20 steps of training；Then SGD is set by optimizer, is training 100,80,50 later,
When 30 step, learning rate is reduced to original 1/10.
A further improvement of the present invention lies in that in step 8), using the network parameter that training is completed before as pre-training
Model, optimizer uses SGD gradient descent algorithm in training process, and initial learning rate is set as 0.001, later every 20 step study
Rate decays to the 1/10 of original learning rate.
The present invention has following beneficial technical effect:
A kind of cylinder creep detection and life-span prediction method based on DENSENET provided by the invention, for detecting cylinder
Whether creep occurs, is extended if creep has occurred, then shutdown inspection；If predicting residue still in normal operating condition
Creep life simultaneously repairs inspection plan according to remaining life reasonable arrangement.The present invention passes through the temperature of acquisition cylinder, pressure and strain letter
Breath, is translated into picture, classification and prediction task of the DENSENET network implementations to input data is established, so that normal operations
Personnel can effectively detect the creep behavior of cylinder immediately and predict the creep longevity while guaranteeing not destroy cylinder body
Life, reasonable arrangement unit operational plan extend the service life of cylinder, ensure that the safe and effective operation of unit.
Detailed description of the invention
Fig. 1 is the flow chart of the cylinder creep detection and life-span prediction method the present invention is based on DENSENET；
Fig. 2 is the flow chart that the present invention carries out creep detection and life prediction.
Specific embodiment
Below according to summary of the invention, combine that the present invention will be described in further detail with embodiment.It is as described below
To a kind of application of the invention, however, it is not limited to this, and implementation personnel can as the case may be modify to wherein parameter.
Assuming that needing to carry out certain power plant's cylinder creep detection and life prediction, N number of examined is shared in the power plant
The mesohigh cylinder of survey, specific implementation step are as follows:
The first step is passed in measured zone arrangement temperature respectively in 4 mesohigh cylinder disposed inboard measured zones of steam turbine
Sensor, pressure sensor and strain transducer obtain the data that the temperature, pressure and strain of single cylinder change over time.Root
According to the shape and operating status of cylinder, distinguishing according to circumferencial direction into vapour section middle section in steam turbine respectively first
3 measured zones are disposed, are evenly arranged 3 measured zones in the exit portion into vapour section；Secondly as traditional cylinder uses spiral shell
Upper and lower cylinder is bound up by bolt, is uniformly distributed 3 measured zones on the split that cylinder is coupled；Finally, according to cylinder
Shape is evenly arranged 3 measured zones at the both ends of cylinder respectively.According to analysis of finite element method as a result, finding cylinder respectively
The highest part of temperature, stress the best part and strain the best part are as numerical analysis measured zone.I.e. each vapour
Every kind of sensor (temperature, pressure and strain) of cylinder is distributed total J=3+3+3+3 × 2+3=18 measured zone.
The data for needing to acquire include the measurement result under normal operating condition and under creep extended mode.To normal fortune
Row state is marked with creep extended mode measurement result, Nn,i,j,tFor the measurement data under certain cylinder normal operating condition,
Cn,i,j,tFor the measurement data under the cylinder creep extended mode, Cn,i,j,tAnd Nn,i,j,tIt corresponds, the cylinder creep entire life
For TLifen, format is { Y-M-D-H-MM-S }n, i.e. year-month-day-when-point-second }n.Wherein n=1,2,3..., 4, i=1,
2,3, temperature data, pressure data and strain data are respectively represented, j=1,2,3..., 18, J be single cylinder measured zone
Sum, t=1,2,3..., T, T is the sum of time of measuring point, and each t corresponds to a specific time, with { Yt-Mt-
Dt-Ht-MMt-St}nIt indicates, i.e. { yeartThe moontDaytWhentPointtSecondt}n, all temperature, pressure and strain datas are identical
Moment measurement result.
Second step, the pretreatment of cylinder measurement data, and place is normalized to temperature, pressure and strain measurement data respectively
Reason.By collected cylinder measurement data Cn,i,j,tAnd Nn,i,j,tIt is short to be cut into the cylinder measurement that M length is 240 at random respectively
Small data set C 'n×m,i,j,tWith N 'n×m,i,j,t, to all cylinder measurement data { N 'i,j,t}n×m{ C 'i,j,t}n×mAccording to as follows
Formula is normalized, and three kinds of data temperatures, pressure and strain correspond to picture rgb pixel value, and accordingly turns data
Turn to picture.
{Maxi}n×m=Max [({ C 'i,j,t}n×m|0≤j≤18,0≤t≤240)}]
{Mini}n×m=Min [({ C 'i,j,t}n×m|0≤j≤18,0≤t≤240)]
In order to facilitate describing and indicate, the data after normalization still use { Ci,j,t}n×mIt indicates.
Cylinder measurement data label is arranged in third step.Measurement data { the N ' of normal operating conditioni,j,t}n×mLabel have
Two, be class label Class respectivelyn×m=0, it indicates to operate normally, service life label Lifen×m=TLifen-OLifen×m, table
Show the remaining life of cylinder this moment, wherein TLifenIndicate the creep entire life of corresponding measured cylinder, OLifen×mIt indicates to correspond to
Cylinder runing time.Measurement data { the C ' of creep extended modei,j,t}n×mClass label Classn×m=1, indicate cylinder
Through the state in creep extension, service life label Lifen×m=TLifen-OLifen×m, indicate the remaining life of cylinder at this time.Tool
The life formula of body is as follows:
OLifen={ YL-ML-DL-HL-MML-SL}n
Lifen×m=TLifen-OLifen×m
={ Y-M-D-H-MM-S }n-{YL-ML-DL-HL-MML-SL}n×m
4th step divides training set for cylinder measurement data and verifying collects.Network is detected for creep, is inputted as X2n×m
={ N 'i,j,t}n×m+{C′i,j,t}n×m, X will be inputted2n×mUpset at random, according to the ratio of 4:1 by X2n×mIt is divided into training set
(X2n×m×0.8)trainCollect (X with verifying2n×m×0.2)validate, corresponding label is (Class2n×m×0.8)trainWith
(Class2n×m×0.2)validate.For predicting network, the processing method of input is identical as creep detection network, X '2n×m=
{N′i,j,t}n×m+{C′i,j,t}n×m, X ' will be inputted2n×mUpset at random, is classified as creep life pre- survey grid according to the ratio of 4:1
Network training set (X '2n×m×0.8)trainCollect (X ' with verifying2n×m×0.2)validate.Corresponding label is cylinder remaining life
(Life2n×m×0.8)trainWith (Life2n×m×0.2)validate。
5th step builds cylinder creep detection network and creep life prediction network based on DENSENET respectively.First
Classified the pond layer parameter adjusting DENSENET network structure of layer according to picture size, the convolution 1 of network, Chi Hua 1, transition zone
Step-length in the uniform pond step of 3-4 is set as [1,2,2,1], and the step-length setting in the uniform pond step of transition zone 2
Chi Huahe for [1,1,2,1], Chi Hua 2 is sized toSecondly, modification cylinder creep detection and life prediction
The full articulamentum structure of network detects network for creep, and the output categorical measure for modifying wherein full articulamentum is 2, and is used
Softmax classifier, two classifications are respectively normal operating condition and creep failure state；Network is predicted for creep life,
The output categorical measure for modifying wherein full articulamentum is 1, as remaining life.
Wherein, the creep detection based on DENSENET and life prediction network are according to the feature of cylinder measurement data in tradition
DENSENET network structure on carry out a part of fine tuning, original DENSENET network has carried out 5 sizes contractings to picture altogether
It is small, using the step-length in convolution and pond step, the width of picture and height are reduced into half each time, in order to enable cylinder
Measurement data can be used in DENSENET network, will be converted into according to cylinder measurement data the height J and width L of picture as
Adjustment standard is adjusted network:
A) according to the sum of the measured zone of cylinder measurement data, it is first determined whether needing to adjust pond layer, such as 1 institute of table
Show, the step-length selection of pond layer is divided into three kinds of situations:
As the total J >=32 of cylinder measured zone, the convolution 1 of network, Chi Hua 1, transition zone 2-4 uniform pond in walk
Length is disposed as [1,2,2,1]；
As sum 16≤J < 32 of cylinder measured zone, the convolution 1 of network, Chi Hua 1, transition zone 3-4 uniform pond
The step-length changed in step is set as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2 is set as [1,1,2,1]；
As sum 8≤J < 16 of cylinder measured zone, in the uniform pond step of convolution 1, Chi Hua 1 and transition zone 4
Step-length is disposed as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2-3 is set as [1,1,2,1]；
Wherein step-length [1,2,2,1] indicates that convolution or Chi Huahe carry out in the both direction of picture height and width
The operation that step-length is 2, and [1,1,2,1] indicates that convolution or Chi Huahe only carry out step-length on this direction of the width of picture
For 2 operation, and the operation that step-length is 1 is carried out on this direction of picture height；
B) according to dimension of picture and network structure, the Chi Huahe size of pondization 2 is determined.
C network) is detected for creep, in last full articulamentum, exports and is set as 2, and connect softmax classifier；
And creep life predicts network, in last full articulamentum, exports and is set as 1, the as remaining life of cylinder.
6th step, training creep detection and creep life prediction model.It is pre- in training creep detection and creep life
During survey grid network, setting optimizer is Adam first, and initial learning rate is set as 0.01,20 steps of training or so；Then will
Optimizer is set as SGD, and later in training 100,80,50,30 step, learning rate is reduced to original 1/10.
7th step, creep detection and life prediction.By cylinder measurement data respectively according to step 1), step 2) and step 3)
It is handled, the cylinder test data after being normalized is translated into the form of picture.It is defeated with cylinder measurement picture
Enter, using the judgement of creep detection network, whether the steam turbine cylinder is in normal operating condition at this time first, if the classification net
It is 0 that network, which exports result, then cylinder operates normally, and otherwise, cylinder has been in creep extended mode at this time, needs to shut down inspection immediately
It looks into, and the operating status of cylinder is observed closely, to prevent serious accident occurs.Secondly, being predicted using creep life
Network obtains the remaining life of the cylinder at this time, and the inspection of repairing of unit is arranged according to remaining life, adjusts operational mode to extend the longevity
Life.
8th step, algorithm maintenance.During practical application, if arrangement effective measuring area is surveyed less than estimated arrangement
Amount region increases measured zone according to demand, then handles measurement data according to step 1), 2), 3), using before
The network parameter that training is completed restarts to train as pre-training model.Optimizer is declined using SGD gradient in training process
Algorithm, initial learning rate are set as 0.001, and every 20 step learning rate decays to the 1/10 of original learning rate later.
Table 1 is that the present invention is based on the creep detections of DENSENET and life prediction network fine tuning structure.
Claims (9)
1. a kind of cylinder creep detection and life-span prediction method based on DENSENET, which comprises the following steps:
1) in cylinder disposed inboard measured zone, arrange that temperature sensor, pressure sensor and strain pass in measured zone respectively
Sensor obtains the data that the temperature, pressure and strain of single cylinder change over time；
The data for needing to acquire include the measurement result under normal operating condition and under creep extended mode, and to normal operation
State is marked with creep extended mode measurement result, Nn,i,j,tFor the measurement data under certain cylinder normal operating condition,
Cn,i,j,tFor the measurement data under the cylinder creep extended mode, Cn,i,j,tAnd Nn,i,j,tIt corresponds, the cylinder creep entire life
For TLifen, format is { Y-M-D-H-MM-S }n, i.e. year-month-day-when-point-second }n；Wherein n=1,2,3..., N, N are to survey
It is total to measure cylinder, i=1,2,3, temperature data, pressure data and strain data, j=1,2,3..., J are respectively represented, J is single
A cylinder measured zone sum, t=1,2,3..., T, T is the sum of time of measuring point, and each t corresponding one specific
Time point, with { Yt-Mt-Dt-Ht-MMt-St}nIt indicates, i.e. { yeartThe moontDaytWhentPointtSecondt}n, all temperature, pressure
It is phase measurement result in the same time with strain data；
2) cylinder measurement data pre-processes, and temperature, pressure and strain measurement data are normalized respectively；
By collected cylinder measurement data Cn,i,j,tAnd Nn,i,j,tIt is short and small to be cut into the cylinder measurement that M length is L at random respectively
Data set C 'n×m,i,j,tWith N 'n×m,i,j,t, wherein m=1,2,3..., M, M are the short and small data count of single cylinder, t=1,2,
3..., L, L are the length of short and small data, i.e. the sum at time point, i and j is remained unchanged, and respectively indicate the property of measurement data,
The measured zone of temperature, pressure and strain and cylinder sum；To all cylinder measurement data { N 'i,j,t}n×mWith
{C′i,j,t}n×mIt is normalized, three kinds of data temperatures, pressure and strain correspond to picture rgb pixel value, and accordingly will
Data are converted into picture；
3) cylinder measurement data label is set；
Measurement data { the N ' of normal operating conditioni,j,t}n×mLabel there are two, be class label Class respectivelyn×m=0, table
Show normal operation, service life label Lifen×m=TLifen-OLifen×m, indicate the remaining life of cylinder this moment, wherein TLifenTable
Show the creep entire life of corresponding measured cylinder, OLifen×mIndicate corresponding cylinder runing time；The measurement number of creep extended mode
According to { C 'i,j,t}n×mClass label Classn×m=1, indicate that cylinder has been in the state of creep extension, service life label
Lifen×m=TLifen-OLifen×m, indicate the remaining life of cylinder at this time；
4) training set is divided for cylinder measurement data and verifying collects；
Network is detected for creep, is inputted as X2n×m={ N 'i,j,t}n×m+{C′i,j,t}n×m, X will be inputted2n×mUpset at random, according to
The ratio of 4:1 is by X2n×mIt is divided into training set (X2n×m×0.8)trainCollect (X with verifying2n×m×0.2)validate, corresponding label is
(Class2n×m×0.8)trainWith (Class2n×m×0.2)validate；For predicting network, the processing method of input and creep detect net
Network is identical, X '2n×m={ N 'i,j,t}n×m+{C′i,j,t}n×m, X ' will be inputted2n×mUpset at random, is classified as according to the ratio of 4:1
Creep life predicts network training collection (X '2n×m×0.8)trainCollect (X ' with verifying2n×m×0.2)validate, corresponding label is surplus for cylinder
Remaining service life (Life2n×m×0.8)trainWith (Life2n×m×0.2)validate；
5) cylinder creep detection network and creep life prediction network based on DENSENET are built respectively；
DENSENET network structure is adjusted according to the pond layer parameter of picture size classification layer first；Secondly, modification cylinder creep
The full articulamentum structure of detection and life prediction network detects network for creep, modifies the output classification of wherein full articulamentum
Quantity is 2, respectively normal operating condition and creep failure state；Network, modification wherein full connection are predicted for creep life
The output categorical measure of layer is 1, as remaining life；
6) creep detection network and creep life based on DENSENET is trained to predict network；
7) creep detection and life prediction；
Cylinder measurement data is handled according to step 1), step 2) and step 3) respectively, the cylinder after being normalized is surveyed
Data are tried, the form of picture is translated into；It is input with cylinder measurement picture, first at this time using the judgement of creep detection network
Whether the steam turbine cylinder is in normal operating condition, if sorter network output result is 0, cylinder is operated normally, no
Then, cylinder has been in creep extended mode at this time, needs instant shutdown inspection, and seen closely to the operating status of cylinder
It surveys, to prevent serious accident occurs；Secondly, obtaining the remaining life of the cylinder at this time, root using creep life prediction network
The inspection of repairing of unit, adjustment operation are arranged according to remaining life.
2. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 1, feature
It is, further comprising the steps of:
8) algorithm is safeguarded；
During practical application, if arrangement effective measuring area is less than estimated arrangement measured zone or increases according to demand
Add measured zone, then handles measurement data according to step 1), 2), 3), the trained DENSENET net finished of use
Network restarts to train on this basis as pre-training model to whole neural network.
3. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 1 or 2, special
Sign is, in step 1), according to the shape and operating status of cylinder, first respectively in steam turbine into vapour section middle part
Divide and disposes J respectively according to circumferencial direction1A measured zone is evenly arranged J in the exit portion into vapour section2A measured zone；Its
It is secondary, since upper and lower cylinder is bound up by traditional cylinder using bolt, J is uniformly distributed on the split that cylinder is coupled3It is a
Measured zone；Finally, being evenly arranged J respectively at the both ends of cylinder according to cylinder shape4A measured zone；According to finite element side
Method analysis as a result, find the highest part of cylinder temperature, stress the best part and strain the best part as number respectively
Value analysis measured zone, i.e., temperature sensor, pressure sensor and the strain transducer of each cylinder are distributed total J=J1+J2+
J3+J4× 2+3 measured zone；
Normal operating condition is that cylinder is on active service the stage for starting to cylinder first macroscopical engineering crackle occur；And creep extends shape
State is from there is first macroscopical engineering crackle to the stage for first critical crack occur；Macroscopical engineering crackle is that length is
0.3~0.5mm, the crackle that depth is 0.1~0.15mm；Critical crack is the crackle that depth reaches 5mm, and cylinder has greatly at this time
Possibility be broken；
Temperature sensor, pressure sensor and strain transducer are sampled with identical sample frequency f, guarantee all data
It is obtained under synchronization, records sampling instant t, format is { Yt-Mt-Dt-Ht-MMt-St}n, initial data is obtained respectively
Cn,i,j,tAnd Nn,i,j,t；Since the temperature of cylinder under operation and pressure change are smaller, then every 5 seconds acquisition one-shot measurements
As a result, sample frequency is 1/5Hz.
4. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 3, feature
It is, in step 2), the raw data set C of cylinder measurement datan,i,j,tAnd Nn,i,j,tThe length of individual data is T, that is, when measuring
Between the sum put, for the measurement data { C of each cylinderi,j,t}n{ Ni,j,t}n, respectively between 0≤t of time point≤T-L
M start time point t is taken at random0, prolong L time point of Shen, t backward0≤t≤t0Data in+L are a short and small data, are obtained
The cylinder that N × M length is L measures short and small data set C 'n×m,i,j,tWith N 'n×m,i,j,t；
Wherein, L value range is 15~30min, i.e. 160~360 time points.
5. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 4, feature
It is, in step 2), with the data set { C ' under creep extended modei,j,t}n×mData instance, the number under normal operating condition
According to { N "i,j×k,t}n×mSimilarly, fixed n, m and i, normalized method are as follows:
{Maxi}n×m=Max [({ C 'i,j,t}n×m|0≤j≤J,0≤t≤L)}]
{Mini}n×m=Min [({ C 'i,j,t}n×m|0≤j≤J,0≤t≤L)]
Wherein, { C 'i,j,t}n×mIndicate the data before normalization,Indicate the data after normalization, { ui}n×mTable
Show the average value of all data of i-th kind of data character, { Maxi}n×mIndicate the maximum of all data of i-th kind of data character
Value, { Mini}n×mIndicate the minimum value of all data of i-th kind of data character, range is upper after Max_value expression normalization
Limit, is herein the rgb pixel value upper limit 255；Min_value indicates the lower limit of range after normalization, is herein rgb pixel value lower limit
0；In order to facilitate describing and indicate, the data after normalization use { Ci,j,t}n×mIt indicates.
6. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 5, feature
It is, in step 3), at the time of correspondence at the last one time point of cylinder measurement data paragraph, i.e., service life is
OLifen={ YL-ML-DL-HL-MML-SL}n
Lifen×m=TLifen-OLifen×m
={ Y-M-D-H-MM-S }n-{YL-ML-DL-HL-MML-SL}n×m。
7. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 6, feature
It is, in step 5), the creep detection based on DENSENET is being passed with life prediction network according to the feature of cylinder measurement data
A part of fine tuning is carried out in the DENSENET network structure of system, original DENSENET network has carried out 5 sizes to picture altogether
It reduces, using the step-length in convolution and pond step, the width of picture and height is reduced into half each time, in order to enable vapour
Cylinder measurement data is able to use in DENSENET network, and height J and width L that picture is converted into according to cylinder measurement data are made
Network is adjusted for adjustment standard:
A) according to the sum of the measured zone of cylinder measurement data, it is first determined whether need to adjust pond layer, as shown in table 1,
The step-length selection of pond layer is divided into three kinds of situations:
As the total J >=32 of cylinder measured zone, convolution 1, Chi Hua 1, the uniform Chi Huazhong step-length of transition zone 2-4 of network are equal
It is set as [1,2,2,1]；
As sum 16≤J < 32 of cylinder measured zone, the convolution 1 of network, Chi Hua 1, transition zone 3-4 uniform pondization step
Step-length in rapid is set as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2 is set as [1,1,2,1]；
Step-length as sum 8≤J < 16 of cylinder measured zone, in the uniform pond step of convolution 1, Chi Hua 1 and transition zone 4
It is disposed as [1,2,2,1], and the step-length in the uniform pond step of transition zone 2-3 is set as [1,1,2,1]；
Wherein step-length [1,2,2,1] indicates that convolution or Chi Huahe carry out step-length in the both direction of picture height and width
For 2 operation, and it is 2 that [1,1,2,1], which indicates that convolution or Chi Huahe only carry out step-length on this direction of the width of picture,
Operation, and on this direction of picture height carry out step-length be 1 operation；
B) according to dimension of picture and network structure, the Chi Huahe size of pondization 2 is determined:
C network) is detected for creep, in last full articulamentum, exports and is set as 2, and connect softmax classifier；And it is compacted
Become life prediction network, in last full articulamentum, exports and be set as 1, the as remaining life of cylinder.
8. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 7, feature
It is, in step 6), during training network, setting optimizer is Adam first, and initial learning rate is set as 0.01, instruction
Practice 20 steps；Then SGD is set by optimizer, later in training 100,80,50,30 step, learning rate is reduced to original
1/10。
9. a kind of cylinder creep detection and life-span prediction method based on DENSENET according to claim 8, feature
It is, in step 8), pre-training model is used as using the network parameter that training is completed before, optimizer use in training process
SGD gradient descent algorithm, initial learning rate are set as 0.001, and every 20 step learning rate decays to the 1/10 of original learning rate later.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811615955.2A CN109781395B (en) | 2018-12-27 | 2018-12-27 | Cylinder creep detection and life prediction method based on DENSENET |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811615955.2A CN109781395B (en) | 2018-12-27 | 2018-12-27 | Cylinder creep detection and life prediction method based on DENSENET |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109781395A true CN109781395A (en) | 2019-05-21 |
| CN109781395B CN109781395B (en) | 2020-07-28 |
Family
ID=66498565
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811615955.2A Active CN109781395B (en) | 2018-12-27 | 2018-12-27 | Cylinder creep detection and life prediction method based on DENSENET |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109781395B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20150276548A1 (en) * | 2014-04-01 | 2015-10-01 | The Regents Of The University Of California | Condition monitoring and analytics for machines |
| JP6131539B2 (en) * | 2012-07-13 | 2017-05-24 | 富士電機株式会社 | Degradation evaluation method for machine parts |
| CN108388841A (en) * | 2018-01-30 | 2018-08-10 | 浙江大学 | Cervical biopsy area recognizing method and device based on multiple features deep neural network |
| CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
-
2018
- 2018-12-27 CN CN201811615955.2A patent/CN109781395B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP6131539B2 (en) * | 2012-07-13 | 2017-05-24 | 富士電機株式会社 | Degradation evaluation method for machine parts |
| US20150276548A1 (en) * | 2014-04-01 | 2015-10-01 | The Regents Of The University Of California | Condition monitoring and analytics for machines |
| CN108388841A (en) * | 2018-01-30 | 2018-08-10 | 浙江大学 | Cervical biopsy area recognizing method and device based on multiple features deep neural network |
| CN108764065A (en) * | 2018-05-04 | 2018-11-06 | 华中科技大学 | A kind of method of pedestrian's weight identification feature fusion assisted learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109781395B (en) | 2020-07-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105241660B (en) | High-speed rail large bridge performance test methods based on health monitoring data | |
| CN109242049B (en) | Water supply pipe network multipoint leakage positioning method and device based on convolutional neural network | |
| CN106761668B (en) | Oil well failure intelligent analysis decision system and method | |
| CN102944416B (en) | Multi-sensor signal fusion technology-based fault diagnosis method for wind turbine blades | |
| CN101660969B (en) | Intelligent fault diagnosis method for gear box | |
| CN110319982B (en) | Buried gas pipeline leakage judgment method based on machine learning | |
| CN104331595B (en) | The mobile principal component correlation coefficient process of bridge damnification early warning | |
| CN104700321A (en) | Analytical method of state running tendency of transmission and distribution equipment | |
| CN107111311A (en) | Utilize the combustion gas turbine Transducer fault detection of sparse coding method | |
| CN109388774A (en) | A kind of thermal power plant NOx prediction model characteristics of variables extracting method based on method of comparison | |
| CN107657110A (en) | The fatigue damage evaluation method of large scale wind power machine blade | |
| CN102607641A (en) | Cluster anomaly detection method of combustion gas turbine | |
| CN113029616B (en) | Compressor surge early fault feature extraction method based on enhanced entropy weight | |
| US20230087465A1 (en) | Short-term AE Monitoring to Identifying ASR Progression in Concrete Structures | |
| CN101539137B (en) | Residual life gray prediction method of aerospace hydraulic pump based on delta filter | |
| CN117523177A (en) | Gas pipeline monitoring system and method based on artificial intelligent hybrid big model | |
| CN106845748A (en) | A kind of INTELLIGENT IDENTIFICATION method of low-voltage collecting meter reading system data deviation reason | |
| CN109781395A (en) | A kind of cylinder creep detection and life-span prediction method based on DENSENET | |
| CN105157986A (en) | Reliability monitoring method for hot-end component of gas turbine | |
| CN104483965B (en) | Liquid-propellant rocket engine ground run steady-state process malfunction determination parameter system of selection | |
| CN104462787A (en) | RST/IAHP-based aeronautical maintenance human factor reliability assessment method | |
| CN106706599A (en) | Ageing grade analysis method of heated surface of boiler | |
| CN207112772U (en) | Boiler hearth temperature on-line prediction system based on core offset minimum binary | |
| CN115310693A (en) | Comprehensive energy data cleaning method, system, equipment and medium | |
| CN108805192A (en) | Monitoring data analysis method based on hierarchical network structure |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |