CN109781395B - Cylinder creep detection and life prediction method based on DENSENET - Google Patents
Cylinder creep detection and life prediction method based on DENSENET Download PDFInfo
- Publication number
- CN109781395B CN109781395B CN201811615955.2A CN201811615955A CN109781395B CN 109781395 B CN109781395 B CN 109781395B CN 201811615955 A CN201811615955 A CN 201811615955A CN 109781395 B CN109781395 B CN 109781395B
- Authority
- CN
- China
- Prior art keywords
- cylinder
- creep
- data
- network
- measurement
- 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.)
- Active
Links
Images
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 cylinder creep detection and evaluation method based on DENSENET, which is used for detecting whether a cylinder creeps or not, if the cylinder creeps and fails, stopping for inspection, closely paying attention to the subsequent operation state of the cylinder and preventing safety accidents; and secondly, predicting the residual creep life and reasonably arranging a repair plan according to the residual creep life. According to the invention, the temperature, pressure and strain information of the cylinder are collected and converted into pictures, and a fine-tuning DENSENET network is established to realize creep detection and service life prediction tasks on the input cylinder measurement data, so that common operators can timely and effectively detect the creep condition of the cylinder and predict the creep life while ensuring that the cylinder body is not damaged, a unit operation plan is reasonably arranged, the service life of the cylinder is prolonged, and the safe and effective operation of the unit is ensured.
Description
Technical Field
The invention belongs to the technical field of steam turbines, and particularly relates to a cylinder creep detection and service life prediction method based on DENSENET.
Background
With the rapid development of the power industry, higher requirements are provided for the unit capacity, the steam inlet parameters, the safety and the like of a power plant, and a steam turbine unit gradually develops towards a high-parameter large-scale direction. The steam turbine cylinder inevitably generates creep damage after long-term service, and high-temperature components mainly affected by the creep damage comprise an ultrahigh pressure/high pressure/medium pressure rotor, an ultrahigh pressure/high pressure/medium pressure outer cylinder, an ultrahigh pressure/high pressure/medium pressure steam generating chamber, a reheating channel and the like. The high-temperature components are made of high-temperature-resistant materials, are high in manufacturing cost and difficult to repair, and have serious threats to the safe and reliable operation of the whole unit equipment once creep damage occurs.
At present, a direct and effective method for detecting and predicting the creep damage of a steam turbine unit does not exist, and the main detection means comprises a lossy test and a nondestructive test, wherein the lossy test needs to collect a material sample of a cylinder, and the creep life of the cylinder under the working temperature and the working pressure is evaluated under the condition of improving the temperature and the working pressure, the method is also called as a durability strength test life extrapolation method, and mainly comprises an isothermal line extrapolation method, an L arson-Miller parameter method and a theta function method.
With the improvement of computer computing power and the development of artificial intelligence, more and more traditional industrial problems adopt artificial intelligence algorithms to obtain new solutions. The invention provides a cylinder creep detection and life prediction method based on DENSENET. The method has the advantages that on one hand, an operator can monitor the measured data of the cylinder in real time without knowing the relevant knowledge in creep deformation, adjust the operation of the unit according to the state and immediately overhaul to ensure the safe and stable operation of the unit, and on the other hand, compared with the traditional scheme, the scheme does not need to damage the cylinder to collect samples, does not need complicated professional experimental instruments, and is higher in accuracy and more immediate and reliable.
Disclosure of Invention
The invention aims to provide a DENSENET-based cylinder creep detection and life prediction method, which is based on the temperature, pressure and strain information of a cylinder and adopts a DENSENET network to perform creep detection and life prediction, so that the detection efficiency can be effectively improved, the calculation time cost can be greatly reduced, the creep state can be detected in real time, and the residual creep life can be predicted.
The invention is realized by adopting the following technical scheme:
a cylinder creep detection and life prediction method based on DENSENET comprises the following steps:
1) arranging a measuring area at the inner side of the cylinder, and arranging a temperature sensor, a pressure sensor and a strain sensor in the measuring area respectively to obtain the data of the temperature, the pressure and the strain of a single cylinder along with the time change;
the data to be collected includes measurements in normal operation and creep extensionMeasuring the results and marking the results of the normal operation state and creep extension state measurements, Nn,i,j,tMeasured data for normal operation of a certain cylinder, Cn,i,j,tMeasured data for creep expansion of the cylinder, Cn,i,j,tAnd Nn,i,j,tOne to one correspondence, the total creep life of the cylinder is T L ifenIn the format of { Y-M-D-H-MM-S }nI.e. { year-month-day-hour-minute-second }n(ii) a Where N is 1,2,3, N is the total number of cylinders to be measured, i is 1,2,3, which represents temperature data, pressure data, and strain data, respectively, J is 1,2,3, J is the total number of measurement regions of a single cylinder, T is 1,2,3, T is the total number of measurement time points, and each T corresponds to a specific time point, and { Y is used as the total number of measurement time pointst-Mt-Dt-Ht-MMt-St}nRepresentation, i.e. { yearst-monthtDay (c)tWhen istIs divided intotSecond(s) (-)t}nAll the temperature, pressure and strain data are measurement results at the same time;
2) preprocessing the measurement data of the cylinder, and respectively carrying out normalization processing on the measurement data of temperature, pressure and strain;
collecting cylinder measurement data Cn,i,j,tAnd Nn,i,j,tM cylinder measurement short data sets C with the length of L are randomly cut out respectivelyn′×m,i,j,tAnd Nn′×m,i,j,tWhere M1, 2,3, M is the total number of short data for a single cylinder, t 1,2,3, L is the length of the short data, i.e. the total number of time points, i and j remaining unchanged, representing the nature of the measured data, temperature, pressure and strain and the total number of measured areas of the cylinder, respectively, the measured data for all cylinders { N'i,j,t}n×mAnd { C'i,j,t}n×mCarrying out normalization processing, wherein the temperature, the pressure and the strain of the three data correspond to the RGB pixel values of the picture, and converting the data into the picture according to the RGB pixel values;
3) setting a cylinder measurement data label;
measurement data of Normal operating State { N'i,j,t}n×mThere are two of the tags of (a),respectively Class label Classn×m0, indicating normal operation, life label L ifen×m=TLifen-OLifen×mRepresenting the remaining life of the cylinder at that moment, where T L ifenIndicating the total creep life of the corresponding cylinder measured, O L ifen×mRepresenting the corresponding cylinder operating time; measurement data of creep extension State { C'i,j,t}n×m Class label Class n×m1, indicating that the cylinder has been in a state of creep expansion, life label L ifen×m=TLifen-OLifen×mIndicating the remaining life of the cylinder at that time;
4) dividing a training set and a verification set aiming at cylinder measurement data;
for a creep detection network, the input is X2n×m={N′i,j,t}n×m+{C′i,j,t}n×mWill input X2n×mRandom scrambling, according to 4: 1 ratio of X2n×mDivided into training sets (X)2n×m×0.8)trainAnd verification set (X)2n×m×0.2)validateThe corresponding label is (Class)2n×m×0.8)trainAnd (Class)2n×m×0.2)validate(ii) a For the prediction network, the input processing method is the same as that of the creep detection network, X'2n×m={N′i,j,t}n×m+{C′i,j,t}n×mX 'will be input'2n×mRandom scrambling, according to 4: 1 into a creep life prediction network training set (X'2n×m×0.8)trainAnd verification set (X'2n×m×0.2)validateThe correspondence label is the cylinder remaining life (L ife)2n×m×0.8)trainAnd (L ife)2n×m×0.2)validate;
5) Respectively building a cylinder creep detection network and a creep life prediction network based on DENSENET;
firstly, adjusting a DENSENET network structure according to parameters of a pooling layer of a picture size classification layer; secondly, modifying the structure of a full connection layer of the cylinder creep detection and life prediction network, wherein for the creep detection network, the number of output types of the full connection layer is modified to be 2, and the output types are respectively in a normal operation state and a creep failure state; for a creep life prediction network, modifying the number of output categories of a full connection layer to be 1, namely obtaining the residual life;
6) training a DENSENET-based creep detection network and a creep life prediction network;
7) creep detection and life prediction;
processing the cylinder measurement data according to the steps 1), 2) and 3) respectively to obtain normalized cylinder test data, and converting the normalized cylinder test data into a picture form; taking a cylinder measurement picture as input, firstly adopting a creep detection network to judge whether the cylinder of the steam turbine is in a normal operation state at the moment, if the output result of the classification network is 0, the cylinder normally operates, otherwise, the cylinder is in a creep expansion state at the moment, the cylinder needs to be stopped for inspection immediately, and the operation state of the cylinder is closely observed to prevent major safety accidents; and secondly, obtaining the residual life of the cylinder at the moment by adopting a creep life prediction network, arranging the repair and the maintenance of the unit according to the residual life, and adjusting the operation.
The invention is further improved in that the method also comprises the following steps:
8) maintaining an algorithm;
in the process of practical application, if the arrangement effective measurement area is less than the expected arrangement measurement area or the measurement area is increased according to the requirement, the measurement data is processed according to the steps 1), 2) and 3), and the trained DENSENET network is used as a pre-training model, so that the training of the whole neural network is restarted on the basis.
According to a further development of the invention, in step 1), depending on the shape and operating state of the cylinder, J is first respectively arranged in the circumferential direction in the middle of the inlet section of the turbine cylinder1A measuring region, wherein J is uniformly arranged at the outlet part of the steam inlet section2A measurement area; secondly, because the traditional cylinder adopts the bolt to connect the upper cylinder and the lower cylinder together, J is evenly distributed on the middle section of the cylinder connection3A measurement area; finally, J are uniformly arranged at two ends of the cylinder respectively according to the shape of the cylinder4Individual surveyA measurement region; according to the analysis result of the finite element method, a part with the highest cylinder temperature, a part with the highest stress and a part with the highest strain are respectively found as numerical analysis measurement areas, namely the temperature sensor, the pressure sensor and the strain sensor of each cylinder are distributed, and J is equal to J1+J2+J3+J4× 2+3 measurement regions;
the normal operation state is the stage from the beginning of the service of the cylinder to the occurrence of the first macro-engineering crack of the cylinder; the creep expansion state is a stage from the occurrence of a first macro-engineering crack to the occurrence of a first critical crack; the macro-engineering cracks are cracks with the length of 0.3-0.5 mm and the depth of 0.1-0.15 mm; the critical crack is a crack with the depth of 5mm, and the cylinder is extremely likely to break at the moment;
the temperature sensor, the pressure sensor and the strain sensor are sampled at the same sampling frequency f, all data are guaranteed to be obtained at the same time, the sampling time t is recorded, and the format is { Y }t-Mt-Dt-Ht-MMt-St}nRespectively obtaining the original data Cn,i,j,tAnd Nn,i,j,t(ii) a Since the temperature and pressure changes of the cylinder in the operating state are small, the measurement result is collected every 5 seconds, and the sampling frequency is 1/5 Hz.
A further development of the invention is that in step 2), the raw data set C of the cylinder measurement datan,i,j,tAnd Nn,i,j,tThe length of the individual data is T, i.e. the total number of measurement time points, measurement data for each cylinder { Ci,j,t}nAnd { Ni,j,t}nRandomly selecting M starting time points T between the time points 0-T-L0Extending backward L time points, t0≤t≤t0The data in + L is a short data, resulting in N × M cylinders of L length measuring short data set C'n×m,i,j,tAnd N'n×m,i,j,t;
Wherein, the value range of L is 15-30 min, namely 160-360 time points.
A further development of the invention is that in step 2), the creep is propagatedData set in State { C'i,j,t}n×mData of (4) is exemplified, data of { N 'in a normal operation state'i,j×k,t}n×mSimilarly, n, m and i are fixed, and the normalization method is 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×mThe data before the normalization is represented by,data after normalization is shown, { ui}n×mAverage of all data representing the properties of the ith data, { Maxi}n×mMaximum of all data representing the property of the ith data, { Mini}n×mMax _ value represents the upper limit of the normalized range, here the RGB pixel value upper limit 255; min _ value represents the lower limit of the normalized range, here the RGB pixel value lower limit 0; for convenience of description and representation, normalized data is taken as { C }i,j,t}n×mAnd (4) showing.
In a further development of the invention, in step 3), the operating life is the time corresponding to the last time point of the cylinder measurement data section, i.e. the operating life is measured
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。
The invention is further improved in that, in step 5), the creep detection and life prediction network based on DENSENET performs a part of fine tuning on the traditional DENSENET network structure according to the characteristics of the cylinder measurement data, the original DENSENET network performs 5 times of size reduction on the picture, the width and height of the picture are reduced by half each time by using the step size in the convolution and pooling steps, and in order to enable the cylinder measurement data to be used in the DENSENET network, the height J and width L converted into the picture according to the cylinder measurement data are used as the adjustment standard to adjust the network:
A) according to the total number of the measurement areas of the cylinder measurement data, firstly, whether the pooling layer needs to be adjusted is judged, and as shown in table 1, the step length selection of the pooling layer is divided into three cases:
when the total number J of the cylinder measuring areas is more than or equal to 32, setting the step length in the uniform pooling of the convolution 1, the pooling 1 and the transition layers 2-4 of the network as [1,2,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 16 and J is less than 32, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the 3-4 of the transition layer of the network is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2 is set to be [1,1,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 8 and J is less than 16, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the transition layer 4 is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2-3 is set to be [1,1,2,1 ];
wherein step [1,2,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 in both the picture height and width directions, and [1,1,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 only in the picture width direction and a step size 1 in the picture height direction;
B) the pooling kernel size of pooling 2 is determined according to picture size and network structure:
C) for the creep detection network, in the last full connection layer, the output is set to be 2, and a softmax classifier is connected; and in the creep life prediction network, in the last full-connection layer, the output is set to be 1, namely the residual life of the cylinder.
The further improvement of the invention is that in the step 6), in the process of training the network, firstly, an optimizer is set as Adam, the initial learning rate is set to be 0.01, and 20 steps of training are carried out; the optimizer is then set to SGD and then the learning rate is reduced to 1/10 as per steps 100, 80, 50, 30 of training.
The further improvement of the invention is that in the step 8), the network parameters finished by the previous training are used as a pre-training model, the optimizer adopts an SGD gradient descent algorithm in the training process, the initial learning rate is set to be 0.001, and then the learning rate is attenuated to 1/10 of the original learning rate every 20 steps.
The invention has the following beneficial technical effects:
the invention provides a cylinder creep detection and life prediction method based on DENSENET, which is used for detecting whether a cylinder creeps or not, if creep expansion occurs, stopping for inspection; if the normal operation state is still in, predicting the residual creep life and reasonably arranging a repair plan according to the residual life. According to the invention, the temperature, pressure and strain information of the cylinder are collected and converted into pictures, and a DENSENET network is established to realize classification and prediction tasks of input data, so that common operators can timely and effectively detect the creep condition of the cylinder and predict the creep life while ensuring that the cylinder body is not damaged, a unit operation plan is reasonably arranged, the service life of the cylinder is prolonged, and safe and effective operation of the unit is ensured.
Drawings
FIG. 1 is a flow chart of a DENSENET-based cylinder creep detection and life prediction method of the present invention;
FIG. 2 is a flow chart of creep detection and life prediction according to the present invention.
Detailed Description
The present invention will be described in further detail with reference to examples in accordance with the summary of the invention. The following description is one application of the present invention, but is not limited thereto, and the practitioner may modify the parameters thereof as appropriate.
Assuming that creep deformation and service life prediction of a certain power plant cylinder are required, the power plant has N medium-high pressure cylinders which need to be detected, and the specific implementation steps are as follows:
according to the shape and operation state of the cylinder, firstly, respectively arranging 3 measuring regions in the middle part of the steam inlet section of the cylinder of the steam turbine according to the circumferential direction, uniformly arranging 3 measuring regions at the outlet part of the steam inlet section, secondly, uniformly distributing 3 measuring regions on the middle surface of the cylinder connection because the traditional cylinder adopts bolts to connect the upper cylinder and the lower cylinder together, and finally, respectively uniformly arranging 3 measuring regions at the two ends of the cylinder according to the shape of the cylinder.
The data that needs to be collected include measurements under normal operating conditions as well as under creep-extension conditions. Marking the measurement results of the normal operation state and the creep expansion state, Nn,i,j,tMeasured data for normal operation of a certain cylinder, Cn,i,j,tMeasured data for creep expansion of the cylinder, Cn,i,j,tAnd Nn,i,j,tOne to one correspondence, the total creep life of the cylinder is T L ifenIn the format of { Y-M-D-H-MM-S }nI.e. { year-month-day-hour-minute-second }n. Where n 1,2,3, 4, i 1,2,3 represent temperature data, pressure data and strain data, respectively, j 1,2,3, 18,j is the total number of measurement regions of a single cylinder, T is 1,2,3, T is the total number of measurement time points, and each T corresponds to a specific time, using { Y }t-Mt-Dt-Ht-MMt-St}nRepresentation, i.e. { yearst-monthtDay (c)tWhen istIs divided intotSecond(s) (-)t}nAll temperature, pressure and strain data are measurements at the same time.
And secondly, preprocessing the cylinder measurement data, and respectively carrying out normalization processing on the temperature, pressure and strain measurement data. Collecting cylinder measurement data Cn,i,j,tAnd Nn,i,j,tM cylinder measurement short data sets C 'with the length of 240 are randomly cut out respectively'n×m,i,j,tAnd N'n×m,i,j,tData { N 'is measured for all cylinders'i,j,t}n×mAnd { C'i,j,t}n×mAnd carrying out normalization processing according to the following formula, wherein the temperature, the pressure and the strain of the three data correspond to the RGB pixel values of the picture, and converting the data into the picture according to the RGB pixel values.
{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)]
For convenience of description and representation, the normalized data still adopts { C }i,j,t}n×mAnd (4) showing.
And thirdly, setting a cylinder measurement data label. Measurement data of Normal operating State { N'i,j,t}n×mHas two labels, respectively Class label Classn×m0, indicating normal operation, life label L ifen×m=TLifen-OLifen×mRepresenting the remaining life of the cylinder at that moment, where T L ifenIndicating the total creep life of the corresponding cylinder measured, O L ifen×mIndicating the corresponding cylinder operating time. Measurement data of creep extension State { C'i,j,t}n×m Class label Class n×m1, indicating that the cylinder has been in a state of creep expansion, life label L ifen×m=TLifen-OLifen×mAnd indicates the remaining life of the cylinder at that time. The specific life calculation formula 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
and fourthly, dividing a training set and a verification set aiming at the cylinder measurement data. For a creep detection network, the input is X2n×m={N′i,j,t}n×m+{C′i,j,t}n×mWill input X2n×mRandom scrambling, according to 4: 1 ratio of X2n×mDivided into training sets (X)2n×m×0.8)trainAnd verification set (X)2n×m×0.2)validateThe corresponding label is (Class)2n×m×0.8)trainAnd (Class)2n×m×0.2)validate. For the prediction network, the input processing method is the same as that of the creep detection network, X'2n×m={N′i,j,t}n×m+{C′i,j,t}n×mX 'will be input'2n×mRandom scrambling, according to 4: 1 into a creep life prediction network training set (X'2n×m×0.8)trainAnd verification set (X'2n×m×0.2)validateThe corresponding label is the cylinder remaining life (L ife)2n×m×0.8)trainAnd (L ife)2n×m×0.2)validate。
The fifth step is to respectivelyAnd constructing a cylinder creep detection network and a creep life prediction network based on DENSENET. Firstly, adjusting a DENSENET network structure according to parameters of a pooling layer of a picture size classification layer, wherein step length in uniform pooling steps of convolution 1, pooling 1 and 3-4 of a transition layer of the network is set to be [1,2,2,1]]And the step size in the uniform pooling step of the transition layer 2 is set to [1,1,2,1]]The size of the pooling nucleus of the pooling 2 is set asSecondly, modifying the structure of a full connection layer of the cylinder creep detection and life prediction network, modifying the creep detection network, wherein the number of output categories of the full connection layer is 2, and adopting a softmax classifier, wherein the two categories are respectively a normal operation state and a creep failure state; for the creep life prediction network, the output class number of the full connection layer is modified to be 1, namely the residual life.
The creep detection and life prediction network based on DENSENET carries out partial fine adjustment on the traditional DENSENET network structure according to the characteristics of cylinder measurement data, the original DENSENET network carries out 5 times of size reduction on pictures, the width and the height of the pictures are reduced by half each time by using the step length in the convolution and pooling steps, and in order to enable the cylinder measurement data to be used in the DENSENET network, the height J and the width L which are converted into the pictures according to the cylinder measurement data are used as adjustment standards to adjust the network:
A) according to the total number of the measurement areas of the cylinder measurement data, firstly, whether the pooling layer needs to be adjusted is judged, and as shown in table 1, the step length selection of the pooling layer is divided into three cases:
when the total number J of the cylinder measuring areas is more than or equal to 32, setting the step length in the uniform pooling of the convolution 1, the pooling 1 and the transition layers 2-4 of the network as [1,2,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 16 and J is less than 32, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the 3-4 of the transition layer of the network is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2 is set to be [1,1,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 8 and J is less than 16, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the transition layer 4 is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2-3 is set to be [1,1,2,1 ];
wherein step [1,2,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 in both the picture height and width directions, and [1,1,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 only in the picture width direction and a step size 1 in the picture height direction;
B) the pooling kernel size of pooling 2 is determined based on the picture size and the network structure.
C) For the creep detection network, in the last full connection layer, the output is set to be 2, and a softmax classifier is connected; and in the creep life prediction network, in the last full-connection layer, the output is set to be 1, namely the residual life of the cylinder.
And sixthly, training a creep detection and creep life prediction model. In the process of training a creep detection and creep life prediction network, firstly setting an optimizer Adam, setting an initial learning rate to be 0.01, and training for about 20 steps; the optimizer is then set to SGD and then the learning rate is reduced to 1/10 as per steps 100, 80, 50, 30 of training.
And seventhly, creep detection and life prediction. And (3) processing the cylinder measurement data according to the step 1), the step 2) and the step 3) respectively to obtain normalized cylinder test data, and converting the normalized cylinder test data into a picture form. Taking a cylinder measurement picture as input, firstly adopting a creep detection network to judge whether the cylinder of the steam turbine is in a normal operation state at the moment, if the output result of the classification network is 0, the cylinder normally operates, otherwise, the cylinder is in a creep expansion state at the moment, the cylinder needs to be stopped for inspection immediately, and the operation state of the cylinder is closely observed to prevent major safety accidents. And secondly, obtaining the residual life of the cylinder by adopting a creep life prediction network, arranging the repair and the inspection of the unit according to the residual life, and adjusting the operation mode to prolong the life.
And eighthly, maintaining the algorithm. In the process of practical application, if the arrangement effective measurement area is less than the expected arrangement measurement area or the measurement area is increased according to the requirement, the measurement data is processed according to the steps 1), 2) and 3), and the network parameters which are trained before are used as a pre-training model to restart the training. In the training process, the optimizer adopts an SGD gradient descent algorithm, the initial learning rate is set to be 0.001, and then the learning rate is attenuated to 1/10 of the original learning rate every 20 steps.
Table 1 shows the creep detection and life prediction network fine tuning structure based on denseneet of the present invention.
Claims (9)
1. A cylinder creep detection and life prediction method based on DENSENET is characterized by comprising the following steps:
1) arranging a measuring area at the inner side of the cylinder, and arranging a temperature sensor, a pressure sensor and a strain sensor in the measuring area respectively to obtain the data of the temperature, the pressure and the strain of a single cylinder along with the time change;
the data to be collected comprises the measurement results in the normal operation state and the creep expansion state, and the measurement results in the normal operation state and the creep expansion state are marked, Nn,i,j,tMeasured data for normal operation of a certain cylinder, Cn,i,j,tMeasured data for creep expansion of the cylinder, Cn,i,j,tAnd Nn,i,j,tOne to one correspondence, the total creep life of the cylinder is T L ifenIn the format of { Y-M-D-H-MM-S }nI.e. { year-month-day-hour-minute-second }n(ii) a Where N is 1,2,3, N is the total number of cylinders measured, i is 1,2,3, which represents temperature data, pressure data and strain data, respectively, and j is1,2,3, J being the total number of measurement regions of a single cylinder, T being the total number of measurement time points, T being the total number of measurement time points, and each T corresponding to a specific time point, using { Y }t-Mt-Dt-Ht-MMt-St}nRepresentation, i.e. { yearst-monthtDay (c)tWhen istIs divided intotSecond(s) (-)t}nAll the temperature, pressure and strain data are measurement results at the same time;
2) preprocessing the measurement data of the cylinder, and respectively carrying out normalization processing on the measurement data of temperature, pressure and strain;
collecting cylinder measurement data Cn,i,j,tAnd Nn,i,j,tM cylinder measurement short data sets C 'with the length of L are randomly cut out respectively'n×m,i,j,tAnd N'n×m,i,j,tWhere M1, 2,3, M is the total number of short data for a single cylinder, t 1,2,3, L is the length of the short data, i.e. the total number of time points, i and j remaining unchanged, representing the nature of the measured data, temperature, pressure and strain and the total number of measured areas of the cylinder, respectively, the measured data for all cylinders { N'i,j,t}n×mAnd { C'i,j,t}n×mCarrying out normalization processing, wherein the temperature, the pressure and the strain of the three data correspond to the RGB pixel values of the picture, and converting the data into the picture according to the RGB pixel values;
3) setting a cylinder measurement data label;
measurement data of Normal operating State { N'i,j,t}n×mHas two labels, respectively Class label Classn×m0, indicating normal operation, life label L ifen×m=TLifen-OLifen×mRepresenting the remaining life of the cylinder at that moment, where T L ifenIndicating the total creep life of the corresponding cylinder measured, O L ifen×mRepresenting the corresponding cylinder operating time; measurement data of creep extension State { C'i,j,t}n×mClass label Classn×m1, indicating that the cylinder has been in a state of creep expansion, life label L ifen×m=TLifen-OLifen×mIndicates the cylinder remaining at that timeResidual life;
4) dividing a training set and a verification set aiming at cylinder measurement data;
for a creep detection network, the input is X2n×m={N′i,j,t}n×m+{C′i,j,t}n×mWill input X2n×mRandom scrambling, according to 4: 1 ratio of X2n×mDivided into training sets (X)2n×m×0.8)trainAnd verification set (X)2n×m×0.2)validateThe corresponding label is (Class)2n×m×0.8)trainAnd (Class)2n×m×0.2)validate(ii) a For the prediction network, the input processing method is the same as that of the creep detection network, X'2n×m={N′i,j,t}n×m+{C′i,j,t}n×mX 'will be input'2n×mRandom scrambling, according to 4: 1 into a creep life prediction network training set (X'2n×m×0.8)trainAnd verification set (X'2n×m×0.2)validateThe correspondence label is the cylinder remaining life (L ife)2n×m×0.8)trainAnd (L ife)2n×m×0.2)validate;
5) Respectively building a cylinder creep detection network and a creep life prediction network based on DENSENET;
firstly, adjusting a DENSENET network structure according to parameters of a pooling layer of a picture size classification layer; secondly, modifying the structure of a full connection layer of the cylinder creep detection and life prediction network, wherein for the creep detection network, the number of output types of the full connection layer is modified to be 2, and the output types are respectively in a normal operation state and a creep failure state; for a creep life prediction network, modifying the number of output categories of a full connection layer to be 1, namely obtaining the residual life;
6) training a DENSENET-based creep detection network and a creep life prediction network;
7) creep detection and life prediction;
processing the cylinder measurement data according to the steps 1), 2) and 3) respectively to obtain normalized cylinder test data, and converting the normalized cylinder test data into a picture form; taking a cylinder measurement picture as input, firstly adopting a creep detection network to judge whether the cylinder of the steam turbine is in a normal operation state at the moment, if the output result of the classification network is 0, the cylinder normally operates, otherwise, the cylinder is in a creep expansion state at the moment, the cylinder needs to be stopped for inspection immediately, and the operation state of the cylinder is closely observed to prevent major safety accidents; and secondly, obtaining the residual life of the cylinder at the moment by adopting a creep life prediction network, arranging the repair and the maintenance of the unit according to the residual life, and adjusting the operation.
2. The DENSENET-based cylinder creep detection and life prediction method according to claim 1, further comprising the steps of:
8) maintaining an algorithm;
in the process of practical application, if the arrangement effective measurement area is less than the expected arrangement measurement area or the measurement area is increased according to the requirement, the measurement data is processed according to the steps 1), 2) and 3), and the trained DENSENET network is used as a pre-training model, so that the training of the whole neural network is restarted on the basis.
3. The DENSENET-based cylinder creep detection and life prediction method according to claim 1 or 2, wherein in step 1), J is firstly respectively arranged in the middle part of the steam inlet section of the steam turbine cylinder according to the shape and the operation state of the cylinder and is arranged according to the circumferential direction1A measuring region, wherein J is uniformly arranged at the outlet part of the steam inlet section2A measurement area; secondly, because the traditional cylinder adopts the bolt to connect the upper cylinder and the lower cylinder together, J is evenly distributed on the middle section of the cylinder connection3A measurement area; finally, J are uniformly arranged at two ends of the cylinder respectively according to the shape of the cylinder4A measurement area; according to the analysis result of the finite element method, a part with the highest cylinder temperature, a part with the highest stress and a part with the highest strain are respectively found as numerical analysis measurement areas, namely the temperature sensor, the pressure sensor and the strain sensor of each cylinder are distributed, and J is equal to J1+J2+J3+J4× 2+3 assaysA measurement region;
the normal operation state is the stage from the beginning of the service of the cylinder to the occurrence of the first macro-engineering crack of the cylinder; the creep expansion state is a stage from the occurrence of a first macro-engineering crack to the occurrence of a first critical crack; the macro-engineering cracks are cracks with the length of 0.3-0.5 mm and the depth of 0.1-0.15 mm; the critical crack is a crack with the depth of 5mm, and the cylinder is extremely likely to break at the moment;
the temperature sensor, the pressure sensor and the strain sensor are sampled at the same sampling frequency f, all data are guaranteed to be obtained at the same time, the sampling time t is recorded, and the format is { Y }t-Mt-Dt-Ht-MMt-St}nRespectively obtaining the original data Cn,i,j,tAnd Nn,i,j,t(ii) a Since the temperature and pressure changes of the cylinder in the operating state are small, the measurement result is collected every 5 seconds, and the sampling frequency is 1/5 Hz.
4. The DENSENET-based cylinder creep detection and life prediction method according to claim 3, wherein in step 2), the original data set C of the cylinder measurement datan,i,j,tAnd Nn,i,j,tThe length of the individual data is T, i.e. the total number of measurement time points, measurement data for each cylinder { Ci,j,t}nAnd { Ni,j,t}nRandomly selecting M starting time points T between the time points 0-T-L0Extending backward L time points, t0≤t≤t0The data in + L is a short data, resulting in N × M cylinders of L length measuring short data set C'n×m,i,j,tAnd N'n×m,i,j,t;
Wherein the value range of L is 15-30 min, namely 180-360 time points.
5. The DENSENET-based cylinder creep detection and life prediction method according to claim 4, wherein in the step 2), the data set { C 'in a creep expansion state is adopted'i,j,t}n×mTake the example ofData in Normal running State { N'i,j,t}n×mSimilarly, n, m and i are fixed, and the normalization method is 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×mThe data before the normalization is represented by,data after normalization is shown, { ui}n×mAverage of all data representing the properties of the ith data, { Maxi}n×mMaximum of all data representing the property of the ith data, { Mini}n×mMax _ value represents the upper limit of the normalized range, here the RGB pixel value upper limit 255; min _ value represents the lower limit of the normalized range, here the RGB pixel value lower limit 0; for convenience of description and representation, normalized data is taken as { C }i,j,t}n×mAnd (4) showing.
6. The DENSENET-based cylinder creep detection and life prediction method according to claim 5, wherein in the step 3), the operation life is the time corresponding to the last time point of the cylinder measurement data segment, namely the time point
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. The method as claimed in claim 6, wherein in step 5), the DENSENET-based creep detection and life prediction network performs a part of fine tuning on a traditional DENSENET network structure according to the characteristics of cylinder measurement data, the original DENSENET network performs 5 times of size reduction on the picture, and the width and height of the picture are reduced by half each time by using the step size in the convolution and pooling steps, so as to enable the cylinder measurement data to be used in the DENSENET network, the height J and width L converted into the picture according to the cylinder measurement data are used as adjustment criteria to adjust the network:
A) according to the total number of the measurement areas of the cylinder measurement data, firstly, whether the pooling layer needs to be adjusted is judged, and as shown in table 1, the step length selection of the pooling layer is divided into three cases:
when the total number J of the cylinder measuring areas is more than or equal to 32, setting the step length in the uniform pooling of the convolution 1, the pooling 1 and the transition layers 2-4 of the network as [1,2,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 16 and J is less than 32, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the 3-4 of the transition layer of the network is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2 is set to be [1,1,2,1 ];
when the total number of the cylinder measuring areas is more than or equal to 8 and J is less than 16, the step length in the uniform pooling step of the convolution 1, the pooling 1 and the transition layer 4 is set to be [1,2,2,1], and the step length in the uniform pooling step of the transition layer 2-3 is set to be [1,1,2,1 ];
wherein step [1,2,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 in both the picture height and width directions, and [1,1,2,1] indicates that the convolution or pooling kernel performs an operation of step size 2 only in the picture width direction and a step size 1 in the picture height direction;
B) the pooling kernel size of pooling 2 is determined according to picture size and network structure:
C) for the creep detection network, in the last full connection layer, the output is set to be 2, and a softmax classifier is connected; and in the creep life prediction network, in the last full-connection layer, the output is set to be 1, namely the residual life of the cylinder.
8. The method for detecting creep and predicting life of a cylinder based on DENSENET as claimed in claim 7, wherein in the step 6), in the process of training the network, firstly setting an optimizer Adam, setting an initial learning rate to 0.01, and training for 20 steps; the optimizer is then set to SGD and then the learning rate is reduced to 1/10 as per steps 100, 80, 50, 30 of training.
9. The method for detecting cylinder creep and predicting cylinder life based on DENSENET as claimed in claim 8, wherein in step 8), the network parameters finished by previous training are used as a pre-training model, the optimizer adopts SGD gradient descent algorithm in the training process, the initial learning rate is set to 0.001, and then the learning rate is attenuated to 1/10 of the original learning rate every 20 steps.
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 CN109781395A (en) | 2019-05-21 |
CN109781395B true 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 (3)
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 |
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 |
Family Cites Families (1)
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 |
-
2018
- 2018-12-27 CN CN201811615955.2A patent/CN109781395B/en active Active
Patent Citations (3)
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 |
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 |
---|---|
CN109781395A (en) | 2019-05-21 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106870298B (en) | Blade root bolt fracture fault detection method based on machine learning | |
CN111855816B (en) | Fan blade fault identification method for comprehensive probability model and cnn network | |
CN111648992B (en) | Gas turbine compressor fault identification early warning method | |
CN115329812B (en) | Bridge infrastructure anomaly monitoring method based on artificial intelligence | |
CN110046409B (en) | ResNet-based steam turbine component health state evaluation method | |
CN116308304B (en) | New energy intelligent operation and maintenance method and system based on meta learning concept drift detection | |
CN113011763A (en) | Bridge damage identification method based on space-time diagram convolution attention | |
CN111950201B (en) | Full life cycle monitoring system and method for sealing device for pump | |
CN111312414B (en) | Fatigue evaluation method of nuclear power plant fatigue monitoring and life evaluation system | |
CN113361686A (en) | Multilayer heterogeneous multi-mode convolutional neural network integrated robot inspection method | |
CN116720048A (en) | Power station auxiliary machine fault diagnosis method and system based on machine learning model | |
CN117371337A (en) | Water conservancy model construction method and system based on digital twin | |
CN109781395B (en) | Cylinder creep detection and life prediction method based on DENSENET | |
CN117368644A (en) | Sensor cable detection method | |
CN112200251A (en) | Monitoring system data evaluation method based on multi-scale convolutional neural network | |
CN113392576B (en) | Method for evaluating and early warning tension state of main mooring rope of tethered balloon | |
CN108573119A (en) | In-service hoisting machinery overall structure remaining life real-time evaluation system | |
CN111914353B (en) | Rotor low-cycle fatigue loss detection method, device and computer equipment | |
KR102406994B1 (en) | Bigdata based system and method for diagnose and analyze abnormalities in facility | |
CN117975372B (en) | Construction site safety detection system and method based on YOLOv and transducer encoder | |
CN113551157B (en) | Oil-gas data acquisition system and method based on sensor network | |
CN112485018B (en) | Mechanical equipment energy consumption abnormity detection method based on mechanism data fusion | |
CN114784974B (en) | Real-time early warning system of smart power grids | |
CN117291351B (en) | Intelligent production management system for cables | |
CN110940506B (en) | High-temperature fastener performance monitoring system and monitoring method |
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 |