CN112580267A - Aero-engine surge prediction method based on multi-branch feature fusion network - Google Patents
Aero-engine surge prediction method based on multi-branch feature fusion network Download PDFInfo
- Publication number
- CN112580267A CN112580267A CN202110040087.5A CN202110040087A CN112580267A CN 112580267 A CN112580267 A CN 112580267A CN 202110040087 A CN202110040087 A CN 202110040087A CN 112580267 A CN112580267 A CN 112580267A
- Authority
- CN
- China
- Prior art keywords
- engine
- network
- surge
- time
- sample
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 52
- 230000004927 fusion Effects 0.000 title claims abstract description 30
- 238000012549 training Methods 0.000 claims abstract description 23
- 238000012360 testing method Methods 0.000 claims abstract description 18
- 238000012544 monitoring process Methods 0.000 claims abstract description 13
- 238000007781 pre-processing Methods 0.000 claims abstract description 12
- 238000001228 spectrum Methods 0.000 claims abstract description 9
- 230000003068 static effect Effects 0.000 claims description 30
- 239000002243 precursor Substances 0.000 claims description 15
- 238000005070 sampling Methods 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 9
- 230000001364 causal effect Effects 0.000 claims description 8
- 230000001687 destabilization Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 5
- 238000013528 artificial neural network Methods 0.000 claims description 3
- 238000000354 decomposition reaction Methods 0.000 claims description 3
- 238000001914 filtration Methods 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000005457 optimization Methods 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000013135 deep learning Methods 0.000 abstract description 2
- 230000006870 function Effects 0.000 description 18
- 238000004458 analytical method Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 4
- 238000011156 evaluation Methods 0.000 description 4
- 238000000605 extraction Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 230000008030 elimination Effects 0.000 description 3
- 238000003379 elimination reaction Methods 0.000 description 3
- 208000012661 Dyskinesia Diseases 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 230000010349 pulsation Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000032683 aging Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000368 destabilizing effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000000295 fuel oil Substances 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000000638 solvent extraction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F02—COMBUSTION ENGINES; HOT-GAS OR COMBUSTION-PRODUCT ENGINE PLANTS
- F02C—GAS-TURBINE PLANTS; AIR INTAKES FOR JET-PROPULSION PLANTS; CONTROLLING FUEL SUPPLY IN AIR-BREATHING JET-PROPULSION PLANTS
- F02C9/00—Controlling gas-turbine plants; Controlling fuel supply in air- breathing jet-propulsion plants
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Combustion & Propulsion (AREA)
- Chemical & Material Sciences (AREA)
- Evolutionary Computation (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Medical Informatics (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Mechanical Engineering (AREA)
- Control Of Positive-Displacement Air Blowers (AREA)
Abstract
The invention provides an aircraft engine surge prediction method based on a multi-branch feature fusion network, and belongs to the field of aircraft engine stability active control. The method comprises the following steps: (1) collecting operation data of an engine surge suppressing test; (2) preprocessing engine operation data and dividing samples by using a sliding window; (3) monitoring power change of a specific frequency signal component by using short-time Fourier transform, determining the surge state of a sample, setting a label, and making a data set; (4) preliminarily extracting time domain statistical characteristics, time frequency spectrum characteristics and complete machine parameter change characteristics of the sample; (5) constructing and training a multi-branch feature fusion network based on FC-VGG 16-TCN; (6) and inputting the real-time running parameters of the engine into the trained network, and detecting the instability foreboding on line. The invention is based on deep learning and feature fusion algorithm, accurately detects instability foreboding by extracting and classifying deep features of the whole engine operation data, makes surge early warning in advance and provides sufficient response time for active control.
Description
Technical Field
The invention relates to an aircraft engine surge prediction method based on a multi-branch feature fusion network, and belongs to the field of aircraft engine stability active control.
Background
Surge, which is a typical aerodynamically unstable flow regime of an aircraft engine, manifests as low frequency pulsations in the engine shaft direction over time in both the flow through the system and the compressor outlet pressure. The low-frequency pulsation greatly reduces the efficiency and the performance of the engine, the vibration of unstable airflow micro-clusters adds extra vibration load to the blades of the compressor, and the temperature rise in front of the turbine caused by pneumatic instability adds extra heat load to the blades of the engine. The vibration load and the heat load greatly accelerate the aging of the engine, shorten the service life of the engine and increase the maintenance cost of the engine. In severe cases, engine failure will also result, leading to catastrophic consequences.
In order to ensure the operation safety of the aero-engine and avoid entering surge, the research on the active stability control of the aero-engine is gradually started, and the main idea is that by detecting a destabilization precursor before surge and stall occur, the feedback control system is used for deflating, changing the angle of a stationary blade, adjusting the flow of fuel oil, injecting air at the position close to a blade tip and the like so as to avoid the occurrence of surge and stall and enable the engine to work in the optimal state all the time. Therefore, stall and surge prediction is a necessary prerequisite for active engine stability control. At present, the prediction of the aircraft engine surge is mainly based on the analysis of a pressure signal at an outlet of a gas compressor in an engine surge test, and the state characteristics of the pressure signal are extracted by using analysis methods such as a time domain analysis method, a frequency domain analysis method, a time frequency analysis method and the like so as to judge the generation of instability precursors.
The patent with publication number CN107165850A discloses a frequency domain hump recognition-based axial flow compressor rotating stall early warning method, which extracts a characteristic hump discrimination factor by performing fast Fourier transform on dynamic pressure in a rotating period of the compressor, judges that rotating stall is approached if the characteristic hump discrimination factor is greater than a set detection threshold value, and sends out early warning; the patent with the publication number of CN110610026A discloses an online identification method of the surge precursors of an aero-engine, wherein 9 monitoring frequency peak values, throttle opening degrees and rotor rotating speeds after pressure ratio signal S conversion are used as characteristic quantities of training samples, and then the characteristic quantities are substituted into a constructed deep network DNN model for training, and finally the surge precursors of an air compressor are identified; the patent with the publication number of CN111737910A discloses an axial flow compressor stall surge prediction method based on deep learning, which includes the steps of constructing a training set and a testing set through sample segmentation and time domain statistical feature extraction of pressure data of a compressor measuring point, then constructing an LR-WaveNet prediction model and conducting training, and predicting the sample surge probability by using the model.
The methods are only limited to the pressure signal of the compressor, and do not consider the attempt of extracting instability precursor characteristics from other state parameters in the process of generating engine surge. In addition, only one method of a frequency domain or a time domain is considered in the preliminary artificial feature design of the method, the design features are not comprehensive enough, and certain limitations are realized.
Disclosure of Invention
In order to avoid the defects in the prior art, the invention provides the method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network, which can detect modal wave type disturbance and spike wave type disturbance which can be directly observed in a time domain, and can also detect other destabilizing precursors such as statistical characteristic abnormal motion hidden in measurable parameters of the engine, thereby further improving the precision of predicting the surge.
The technical solution of the invention is as follows:
the method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network comprises the following specific steps:
1) collecting relevant operation data of an engine surge suppressing test;
2) preprocessing engine operation data and dividing samples by using a sliding window;
3) monitoring power change of a specific frequency signal component by using short-time Fourier transform, determining the surge state of a sample, setting a label, and making a data set;
4) preliminarily extracting time domain statistical characteristics, time frequency spectrum characteristics and complete machine parameter change characteristics of the sample;
5) constructing and training a multi-branch feature fusion network based on FC-VGG 16-TCN;
6) and inputting the real-time running parameters of the engine into the trained network, and detecting the instability foreboding on line.
Further, the engine operating parameters collected in step 1) include: the method comprises the following steps of (1) setting engine speed N, total compressor inlet pressure Pt1, compressor outlet static pressure Ps3, total compressor inlet temperature Tt1, total compressor outlet temperature Tt3 and exhaust temperature EGT; because the most significant characteristic of surge is embodied as the change of the static pressure at the outlet of the compressor, the static pressure Ps3 at the outlet of the compressor is called as a main characteristic parameter, and the rest parameters are called as engine state parameters;
further, the step 2) of preprocessing the engine operation data and dividing the sample by using a sliding window specifically includes the following steps:
and 2-1, removing high-frequency components in the static pressure Ps3 at the outlet of the compressor by using a Butterworth low-pass filter, and preventing aliasing interference of high-frequency signals.
And 2-2, considering that the monitoring surge frequency component is 5-30Hz, down-sampling the Ps3 data after high frequency filtering to 300Hz according to the sampling frequency of the pressure sensor so as to meet the Nyquist sampling theorem.
And 2-3, setting a sliding window to segment the compressor outlet static pressure Ps3 by considering the correlation of the control period of the aircraft engine and the time sequence data. The control period is T, a window with the length of 10T is set to slide on time domain data, the moving step length is T, a compressor outlet static pressure sequence Ps3 '(N), an engine speed N' (N), a compressor inlet total pressure Pt1 '(N), a compressor inlet total temperature Tt 1' (N), a compressor outlet total temperature Tt3 '(N) and an exhaust temperature EGT' (N) which fall in the window are cut out to be used as characteristic parameters of a sample, wherein N is 1, 2, 3, … … steps, and steps are the length of an engine state parameter sequence corresponding to 10 control periods;
and 2-4, finally forming a sample X ═ { X1, X2, X3 … … Xn }, wherein Xn represents a sample corresponding to the time of the nth control period after the start of the engine operation. Wherein each sample comprises: the method comprises the following steps that a compressor outlet static pressure sequence Ps3 '(N), an engine speed N' (N), a compressor inlet total pressure Pt1 '(N), a compressor inlet total temperature Tt 1' (N), a compressor outlet total temperature Tt3 '(N) and an exhaust temperature EGT' (N) in 10 control periods are counted at the current time and historical time;
further, the monitoring of the power change of the specific frequency signal component by using the short-time fourier transform in the step 3), determining the surge state of the sample, setting a label, and making a data set specifically includes the following steps:
step 3-1, performing short-time Fourier transform on the static pressure Ps3 of the compressor in the whole surge approaching test to obtain a frequency spectrum matrix of the static pressure Ps 3;
step 3-2, monitoring signal component power corresponding to the surge frequency of 0-30 Hz, and considering that the power enters a destabilization propagation state when the power suddenly increases; entering a full surge condition when Ps3 suddenly drops; when the power of the signal component of 0-30 Hz suddenly drops to below 0, the engine is considered to be recovered to a stable running state; determining the engine running state corresponding to each sample according to the data;
3-3, setting a label for each sample by adopting a single-hot coding mode: the stable operation state is [1, 0, 0 ]; the destabilization propagation state is [0, 1, 0 ]; surge condition is [0, 0, 1 ];
3-4, dividing the training data set into a training set and a test set according to the ratio of 7: 3;
further, the preliminary extraction of the time domain statistical characteristics, the time frequency spectrum characteristics and the complete machine parameter variation characteristics of the sample in the step 4) specifically includes the following steps:
and 4-1, extracting the time domain statistical characteristics F1 of the samples obtained in the step 2-5. The statistical parameter calculation is carried out on the compressor outlet static pressure sequence Ps3 '(n) of the sample, and 9 statistical parameters of the mean value, the variance, the root mean square error, the peak factor, the pulse factor, the margin factor, the kurtosis factor, the wave factor and the skewness factor of the Ps 3' (n) are respectively obtained and used as the first type characteristics of the sample, and the first type characteristics are also input into a first branch network. The specific calculation formula is as follows:
mean value:
variance:
standard deviation:
root mean square error:
crest factor:
pulse factor:
margin factor:
kurtosis factor:
form factor:
skewness factor:
wherein, x (n) is the time domain sequence of the signal, which is Ps 3' (n) in the present invention.
And 4-2, extracting the time-frequency characteristics F2 of the sample obtained in the step 2-5. The method comprises the steps of solving a maximum value point x (t) of an input 20ms sample signal x (t) by using Hilbert-Huang transformi) And minimum value point x (t)j) Constructing an upper envelope line x and a lower envelope line x by adopting cubic spline function interpolation on a maximum value and minimum value pointu(t) and xl(t) and calculating the mean function m1(t) of (d). Obtaining a first component h1(t)＝x(t)-m1(t) checking whether the condition of the modal component is satisfied, and if so, obtaining the modal component c satisfying the IMF1(t) of (d). Subtracting the first modal component from the original signal to obtain a signal r1(t)＝x(t)-c1(t) adding r1(t) repeating the above operations as a new "original signal" until the screening condition is reached
And when the value is less than the preset value, ending the empirical mode decomposition. The original signal is thus divided into empirical mode components and a residual signal:
the original signal can be expressed into a three-dimensional time-frequency characteristic matrix of time, frequency and power as the input of the second branch network by performing Hilbert transform on the signal.
And 4-3, taking 5 engine state parameter sequences of the engine speed N ' (N), the compressor inlet total pressure Pt1 ' (N), the compressor inlet total temperature Tt1 ' (N), the compressor outlet total temperature Tt3 ' (N) and the exhaust temperature EGT ' (N) in 10 control cycles as the state parameter characteristics F3 of the corresponding time samples obtained in the step 2-5, and taking the state parameter characteristics F3 as the input of the third branch network.
Further, step 5) constructing and training a multi-branch feature fusion network based on the FC-VGG 16-TCN;
step 5-1, establishing a branch network 1, wherein the branch network is a fully Connected (Full Connected) neural network and comprises a 1-layer input layer, a 1-layer hidden layer and a 1-layer output layer, and all the layers are Connected in sequence in a fully Connected mode;
the input of the branch network is a time domain statistical characteristic F1 of the sample, and the dimensionality is (9, 1);
the output of the branching network is a high-dimensional feature HF1 with a dimension of (m1, 1);
step 5-2, establishing a branch network 2, wherein the branch network adopts a VGG16 structure and consists of 13 layers (convolutional layer + pooling layer + ReLU activation function) and 3 layers (full connection layer + ReLU activation function), wherein the ReLU function: relu (x) max { ax, x }, (0 < a < 1);
the input of the branch network is a time-frequency characteristic F2 of a sample, the dimension of the time-frequency characteristic is (steps, steps, 3) through reconstruction, and the steps are the lengths of engine state parameter sequences corresponding to 10 control cycles;
the output of the branching network is a high-dimensional feature HF2 with a dimension of (m2, 1);
step 5-3, establishing a branch Network 3, wherein the branch Network is a Time Convolutional Network (Time Convolutional Network) and comprises a causal Convolutional sum and residual error module; wherein the causal convolution is yt＝w1·xt-2+w2·xt-1+w1·xtThe residual module comprises two layers of convolution and ReLU nonlinear mapping, and Weightnorm and Dropout are added into each layer to normalize the network;
the input of the branch network is a state parameter characteristic F3 of a sample, the dimensionality is (5, steps), and the steps are the length of an engine state parameter sequence corresponding to 10 control cycles;
the output of the branching network is a high-dimensional feature HF3, whose dimension is (m3, 1).
Step 5-4, splicing three high-dimensional features output by three branch networks by adopting a Stacking algorithm, building a Stacking fusion module, inputting two full convolution layers and one full connection layer, and finally outputting through a Softmax activation function to obtain an engine state S;
and 5-5, selecting a loss function as a multi-classification cross entropy function L, selecting an optimization algorithm as an Adam algorithm, training by using samples, and updating network parameters.
The cross entropy function is defined as follows:
wherein: m is the number of classifications; y isicIs an indicator variable (take 0 or 1), is 1 if the class is the same as that of sample i, and is 0 otherwise; p is a radical oficIs the predicted probability that the observation sample i belongs to class c.
The concrete process of updating the network parameters by the Adam algorithm is as follows:
defining: step size ε (default 0.001)
Defining: exponential decay rate of moment estimation, ρ1And ρ2Within the interval [0, 1). (Default 0.9 and 0.999 respectively)
Defining: small constant delta for numerical stability (default 10)-8)
Defining: initial parameter theta of network
Initializing the first and second moment variables s-0 and r-0
Initialization time step t is 0
If the training times are not reached, the following processes are circulated:
(1) sampling from training set containing m samples { x(1)，……，x(m)Small batch with the corresponding target y(i)。
(2) Calculating the gradient:
(3)t＝t+ 1
(4) updating biased first moment estimates: where is ρ1s+(1-ρ1)g
(5) Updating the biased second moment estimation:
(6) correcting the deviation of the first moment:
(7) correcting the deviation of the second moment:
(8) and (3) calculating and updating:
(9) application updating: theta + delta theta
Step 5-6, inputting the test set into the model obtained in the step 5-5, and performing performance evaluation on the model; the invention selects F1-score as the evaluation index of the model:
wherein TP is a real example, FN is a false negative example, FP is a false positive example; precision is the Precision rate, that is, the classification result is the proportion of true positive examples in the positive example samples; recall is Recall rate, namely the proportion of the prediction result in the sample which is really a positive example to the sample which is a positive example;
further, step 6) inputting the real-time running parameters of the engine into the trained network, and detecting the instability precursor on line, wherein the method specifically comprises the following steps:
step 6-1, preprocessing the real-time running data of the engine in the latest 10 control periods according to the preprocessing method in the step 2) to obtain a sample Xt corresponding to the current moment;
step 6-2, substituting Xt into the FC-VGG 16-TCN-based multi-branch feature fusion network trained in the step 5), and outputting the current state St of the engine;
and 6-3, judging the engine running state according to St: and when St is [0, 1, 0], judging that the engine has instability precursors, and giving a surge early warning.
Advantageous effects
The method is based on a characteristic fusion method, not only considers a time domain analysis and a time-frequency analysis method of a typical compressor outlet static pressure signal, but also combines the abnormal movement characteristics hidden in engine state parameters when the instability precursor appears, and extracts and fuses the time domain characteristics, the time-frequency characteristics and the abnormal movement characteristics of the instability precursor by utilizing the strong characteristic extraction capability of a CNN (CNN) to a three-dimensional matrix and a TCN (TCN network) to multi-dimensional time sequence data, thereby effectively detecting the instability precursor. Compared with the method for performing the surge elimination measure after the engine enters the surge in the traditional surge elimination control, the method can judge the state of the engine on line, identify the instability foreboding before the surge occurs, send out early warning in advance, perform the surge elimination measure, avoid damage to the engine caused by entering the surge, and guarantee the safe operation of the engine. In addition, the method is an automatic feature extraction method based on data driving, so that the method can be applied to various types of aircraft engines, does not need to artificially design features and threshold values of engine instability, and has certain universality.
Drawings
FIG. 1 is a flow chart of an aircraft engine surge prediction method based on a multi-branch feature fusion network
FIG. 2 is a schematic diagram of variation of static pressure Ps3 at the outlet of the compressor in the surge-forcing test
FIG. 3 is a schematic diagram of engine operating state division
FIG. 4 is a diagram of causal convolution in a TCN branch network
FIG. 5 is a schematic diagram of residual error modules in a TCN branch network
FIG. 6 is a diagram of a multi-branch feature fusion network architecture
FIG. 7 is a schematic diagram of an on-line prediction result of engine surge
Detailed Description
The technical solution of the present invention is further explained with reference to the accompanying drawings and specific embodiments.
The flow chart of the aircraft engine surge prediction method based on the multi-branch feature fusion network is shown in fig. 1, and specifically comprises the following steps:
1) the surge test time of an engine at a certain time is 1007 seconds, and relevant operation data including the engine speed N, the total pressure Pt1 at the inlet of the compressor, the static pressure Ps3 at the outlet of the compressor, the total temperature Tt1 at the inlet of the compressor, the total temperature Tt3 at the outlet of the compressor and the exhaust temperature EGT are collected at the sampling frequency of 1000 Hz; the static pressure at the outlet of the compressor in the test process is shown in figure 2;
2) preprocessing the engine operating parameters collected in the step 1), and dividing samples by using a sliding window, wherein the method comprises the following steps:
and 2-1, setting the cut-off frequency to be 150Hz by using a Butterworth low-pass filter, removing high-frequency components in the static pressure Ps3 at the outlet of the compressor, and preventing aliasing interference of high-frequency signals.
And 2-2, considering that the monitoring surge frequency component is 5-30Hz, down-sampling the Ps3 data after high frequency filtering to 300Hz according to the sampling frequency of the pressure sensor so as to meet the Nyquist sampling theorem.
And 2-3, setting a sliding window to segment the compressor outlet static pressure Ps3 by considering the correlation of the control period of the aircraft engine and the time sequence data. The control period is 20ms, a window with the length of 200ms is set to slide on time domain data, the moving step length is 20ms, and a compressor outlet static pressure sequence Ps3 '(N), an engine speed N' (N), a compressor inlet total pressure Pt1 '(N), a compressor inlet total temperature Tt 1' (N), a compressor outlet total temperature Tt3 '(N) and an exhaust temperature EGT' (N) which fall in the window are cut out to be used as characteristic parameters of a sample, wherein N is 1, 2, 3 and … … 200;
step 2-4, 50364 samples X ═ X1, X2, X3 … … X50364 are finally formed, where each sample comprises: the method comprises the following steps that a compressor outlet static pressure sequence Ps3 '(N), an engine speed N' (N), a compressor inlet total pressure Pt1 '(N), a compressor inlet total temperature Tt 1' (N), a compressor outlet total temperature Tt3 '(N) and an exhaust temperature EGT' (N) in 10 control periods are counted at the current time and historical time;
3) monitoring the power change of a specific frequency signal component by using short-time Fourier transform, determining the surge state of a sample, setting a label, and making a data set, wherein the method specifically comprises the following steps:
step 3-1, performing short-time Fourier transform on the static pressure Ps3 of the compressor in the whole surge approaching test to obtain a frequency spectrum matrix of the static pressure Ps 3;
step 3-2, monitoring signal component power corresponding to the surge frequency of 0-30 Hz, and considering that the power enters a destabilization propagation state when the power suddenly increases; entering a full surge condition when Ps3 suddenly drops; when the power of the signal component of 0-30 Hz suddenly drops to below 0, the engine is considered to be recovered to a stable running state; from this, the engine operating state corresponding to each sample can be determined, and the engine state partitioning results are shown in FIG. 3.
3-3, setting a label for each sample by adopting a single-hot coding mode: the stable operation state is [1, 0, 0 ]; the destabilization propagation state is [0, 1, 0 ]; surge condition is [0, 0, 1 ];
3-4, dividing the training data set into a training set and a test set according to the ratio of 7: 3;
4) the method for preliminarily extracting the time domain statistical characteristic F1, the time-frequency spectrum characteristic F2 and the complete machine parameter change characteristic F3 of the sample comprises the following steps:
and 4-1, extracting the time domain statistical characteristics F1 of the samples obtained in the step 2-5. The statistical parameter calculation is carried out on the compressor outlet static pressure sequence Ps3 '(n) of the sample, and 9 statistical parameters of the mean value, the variance, the root mean square error, the peak factor, the pulse factor, the margin factor, the kurtosis factor, the wave factor and the skewness factor of the Ps 3' (n) are respectively obtained and used as the first type characteristics of the sample, and the first type characteristics are also input into a first branch network. The specific calculation formula is as follows:
mean value:
variance:
standard deviation:
root mean square error:
the peaks are due to:
pulse factor:
margin factor:
kurtosis factor:
form factor:
skewness factor:
wherein, x (n) is the time domain sequence of the signal, which is Ps 3' (n) in the present invention.
And 4-2, extracting the time-frequency characteristics F2 of the sample obtained in the step 2-5. The Hilbert-Huang transform is adopted, a compressor outlet static pressure sequence Ps 3' (n) of a sample is used as a sample signal x (t), and a maximum value point x (t) of the sample is obtainedi) And minimum value point x (t)j) Constructing an upper envelope line x and a lower envelope line x by adopting cubic spline function interpolation on a maximum value and minimum value pointu(t) and xl(t) and calculating the mean function m1(t); obtaining a first component h1(t)＝x(t)-m1(t) checking whether the condition of the modal component is satisfied, and if so, obtaining the modal component c satisfying the IMF1(t) of (d). Subtracting the first modal component from the original signal to obtain a signal r1(t)＝x(t)-c1(t) adding r1(t) repeating the above operations as a new "original signal" until the screening condition is reached
And when the value is less than the preset value, ending the empirical mode decomposition. The original signal is thus divided into empirical mode components and a residual signal:
the original signal can be expressed into a three-dimensional time-frequency characteristic matrix of time, frequency and power as the input of the second branch network by performing Hilbert transform on the signal.
And 4-3, taking 5 engine state parameter sequences of the engine speed N ' (N), the compressor inlet total pressure Pt1 ' (N), the compressor inlet total temperature Tt1 ' (N), the compressor outlet total temperature Tt3 ' (N) and the exhaust temperature EGT ' (N) in 10 control cycles as the state parameter characteristics F3 of the corresponding time samples obtained in the step 2-5, and taking the state parameter characteristics F3 as the input of the third branch network.
5) The method comprises the following steps of constructing and training a multi-branch feature fusion network based on FC-VGG16-TCN, wherein the network structure is shown in FIG. 4, and the method specifically comprises the following steps:
step 5-1, establishing a branch network 1, wherein the branch network is a fully Connected (Full Connected) neural network and comprises a 1-layer input layer, a 1-layer hidden layer and a 1-layer output layer, and all the layers are Connected in sequence in a fully Connected mode;
the input of the branch network is a time domain statistical characteristic F1 of the sample, and the dimensionality is (9, 1);
the output of the branching network is a high-dimensional feature HF1 with dimensions (3, 1);
step 5-2, establishing a branch network 2, wherein the branch network adopts a VGG16 structure and consists of 13 layers (convolutional layer + pooling layer + ReLU activation function) and 3 layers (full connection layer + ReLU activation function), wherein the ReLU function: relu (x) max { ax, x), (0 < a < 1);
the input of the branch network is the time-frequency characteristic F2 of the sample, and the dimension of the reconstructed time-frequency characteristic is (200, 200, 3);
the output of the branching network is a high-dimensional feature HF2, whose dimension is (16, 1);
step 5-3, establishing a branch Network 3, wherein the branch Network is a Time Convolutional Network (Time Convolutional Network) and comprises causal Convolutional and residual modules, which are respectively shown in fig. 5 and fig. 6;
wherein the causal convolution comprises an input layer, 2 hidden layers and an output layer, each layer taking into account the causal relationship of the historical input x (i) to the current output y (t):
yt＝w1·x(t-n)+w2·x(t-n+1)+…+wn+1·x(t)
the residual module comprises two layers of convolution and ReLU nonlinear mapping, and Weightnorm and Dropout are added into each layer to regularize the network;
the input of the branch network is a state parameter characteristic F3 of a sample, the dimensionality is (5, steps), and the steps are the length of an engine state parameter sequence corresponding to 10 control cycles;
the output of the branching network is a high-dimensional feature HF3, whose dimension is (9, 1).
Step 5-4, splicing three high-dimensional features output by three branch networks by adopting a Stacking algorithm, building a Stacking fusion module, inputting two full convolution layers and one full connection layer, and finally outputting through a Softmax activation function to obtain an engine state S corresponding to a sample;
and 5-5, selecting a loss function as a multi-classification cross entropy function L, selecting an optimization algorithm as an Adam algorithm, training by using samples, and updating network parameters.
The cross entropy function is defined as follows:
wherein: m is 3; y isicIs an indicator variable (take 0 or 1), is 1 if the class is the same as that of sample i, and is 0 otherwise; p is a radical oficIs the predicted probability that the observation sample i belongs to class c.
The concrete process of updating the network parameters by the Adam algorithm is as follows:
defining: step size epsilon is 0.001
Defining: exponential decay rate of moment estimation, ρ1＝0.9，ρ2＝0.999
Defining: small constant δ 10 for numerical stability-8
Defining: initial network parameter θ
Initializing the first and second moment variables s-0 and r-0
Initialization time step t is 0
Setting the training times to be 1000, circulating the following processes for 1000 times:
(1) sampling from training set containing m samples { x(1)，……，x(m)Small batch of }, pairShould be targeted to y(i)。
(2) Calculating the gradient:
(3)t＝t+ 1
(4) updating biased first moment estimates: where is ρ1s+(1-ρ1)g
(5) Updating the biased second moment estimation:
(6) correcting the deviation of the first moment:
(7) correcting the deviation of the second moment:
(8) and (3) calculating and updating:
(9) application updating: theta + delta theta
Step 5-6, inputting the test set into the model obtained in the step 5-5, and performing performance evaluation on the model; the invention selects F1-score as the evaluation index of the model:
wherein TP is a real example, FN is a false negative example, FP is a false positive example; precision is the Precision rate, that is, the classification result is the proportion of true positive examples in the positive example samples; recall is Recall rate, namely the proportion of the prediction result in the sample which is really a positive example to the sample which is a positive example;
6) the method provided by the invention is applied to online real-time prediction of surge by using another set of engine surge operation parameters, and the result is shown in FIG. 7, and the method specifically comprises the following steps:
step 6-1, preprocessing the real-time running data of the engine in the latest 10 control periods according to the preprocessing method in the step 2) to obtain a sample Xt corresponding to the current moment;
step 6-2, substituting Xt into the FC-VGG 16-TCN-based multi-branch feature fusion network trained in the step 5), and outputting the current state St of the engine;
and 6-3, judging the engine running state according to St: and when St is [0, 1, 0], judging that the engine has instability precursors, and giving a surge early warning.
As shown in FIG. 5, in the test, the engine enters a full surge state from a stable operation state at 1.95s, the invention can effectively detect the instability precursors in the group of engine operation data at 1s, and the early warning of surge is made 1s ahead, thereby verifying the effectiveness of the method for predicting the aircraft engine surge based on the multi-branch feature fusion network provided by the invention.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any person skilled in the art should be considered to be within the technical scope of the present invention, and the technical solutions and the inventive concepts thereof according to the present invention should be equivalent or changed within the scope of the present invention.
Claims (7)
1. The method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network is characterized by comprising the following steps of:
1) collecting relevant operation data of an engine surge suppressing test;
2) preprocessing engine operation data and dividing samples by using a sliding window;
3) monitoring power change of a specific frequency signal component by using short-time Fourier transform, determining the surge state of a sample, setting a label, and making a data set;
4) preliminarily extracting time domain statistical characteristics, time frequency spectrum characteristics and complete machine parameter change characteristics of the sample;
5) constructing and training a multi-branch feature fusion network based on FC-VGG 16-TCN;
6) and inputting the real-time running parameters of the engine into the trained network, and detecting the instability foreboding on line.
2. The method for predicting the surge of the aero-engine based on the multi-branch feature fusion network as claimed in claim 1, wherein the step 1) collects relevant operation data of an engine surge-approaching test, and the relevant operation data specifically comprises engine speed N, compressor inlet total pressure Pt1, compressor outlet static pressure Ps3, compressor inlet total temperature Tt1, compressor outlet total temperature Tt3 and exhaust temperature EGT.
3. The method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network as claimed in claim 1, wherein the step 2) of preprocessing the engine operation data and dividing samples by using a sliding window comprises the following steps:
and 2-1, removing high-frequency components in the static pressure Ps3 at the outlet of the compressor by using a Butterworth low-pass filter, and preventing aliasing interference of high-frequency signals.
And 2-2, considering that the monitoring surge frequency component is 5-30Hz, down-sampling the Ps3 data after high frequency filtering to 300Hz according to the sampling frequency of the pressure sensor so as to meet the Nyquist sampling theorem.
And 2-3, setting a sliding window to segment the compressor outlet static pressure Ps3 by considering the correlation of the control period of the aircraft engine and the time sequence data. The control period is T, a window with the length of 10T is set to slide on time domain data, the moving step length is T, a compressor outlet static pressure sequence Ps3 '(N), an engine speed N' (N), a compressor inlet total pressure Pt1 '(N), a compressor inlet total temperature Tt 1' (N), a compressor outlet total temperature Tt3 '(N) and an exhaust temperature EGT' (N) which fall in the window are cut out to be used as characteristic parameters of a sample, wherein N is 1, 2, 3, … … steps, and steps are the length of an engine state parameter sequence corresponding to 10 control periods;
and 2-4, finally forming a sample X ═ { X1, X2, X3 … … Xn }, wherein Xn represents a sample corresponding to the time of the nth control period after the start of the engine operation. Wherein each sample comprises: the method comprises the following steps of counting 10 control periods of compressor outlet static pressure sequences Ps3 '(N), engine speed N' (N), compressor inlet total pressure Pt1 '(N), compressor inlet total temperature Tt 1' (N), compressor outlet total temperature Tt3 '(N) and exhaust temperature EGT' (N) at the current time and historical time.
4. The method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network as claimed in claim 1, wherein the step 3) of monitoring the power change of the specific frequency signal component by using short-time Fourier transform, determining the surge state of the sample, setting a label, and making a data set specifically comprises the following steps:
step 3-1, performing short-time Fourier transform on the static pressure Ps3 of the compressor in the whole surge approaching test to obtain a frequency spectrum matrix of the static pressure Ps 3;
step 3-2, monitoring signal component power corresponding to the surge frequency of 0-30 Hz, and considering that the power enters a destabilization propagation state when the power suddenly increases; entering a full surge condition when Ps3 suddenly drops; when the power of the signal component of 0-30 Hz suddenly drops to below 0, the engine is considered to be recovered to a stable running state; determining the engine running state corresponding to each sample according to the data;
3-3, setting a label for each sample by adopting a single-hot coding mode: the stable operation state is [1, 0, 0 ]; the destabilization propagation state is [0, 1, 0 ]; surge condition is [0, 0, 1 ];
and 3-4, dividing the training data set into a training set and a test set according to the ratio of 7: 3.
5. The method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network as claimed in claim 1, wherein the step 4) of preliminarily extracting the time-domain statistical feature F1, the time-frequency spectrum feature F2 and the complete machine parameter variation feature F3 of the sample specifically comprises the following steps:
and 4-1, extracting the time domain statistical characteristics F1 of the samples obtained in the step 2-5. Calculating statistical parameters of a compressor outlet static pressure sequence Ps3 '(n) of a sample, and respectively obtaining 9 statistical parameters of a mean value, a variance, a root mean square error, a peak value factor, a pulse factor, a margin factor, a kurtosis factor, a waveform factor and a skewness factor of the Ps 3' (n) as time domain statistical characteristics of the sample;
and 4-2, extracting the time-frequency characteristics F2 of the sample obtained in the step 2-5. The method comprises the steps of solving a maximum value point x (t) of an input 20ms sample signal x (t) by using Hilbert-Huang transformi) And minimum value point x (t)j) Constructing an upper envelope line x and a lower envelope line x by adopting cubic spline function interpolation on a maximum value and minimum value pointu(t) and xl(t) and calculating the mean function m1(t) of (d). Obtaining a first component h1(t)＝x(t)-m1(t) checking whether the condition of the modal component is satisfied, and if so, obtaining the modal component c satisfying the IMF1(t) of (d). Subtracting the first modal component from the original signal to obtain a signal r1(t)＝x(t)-c1(t) adding r1(t) repeating the above operations as a new "original signal" until the screening condition is reached
And when the value is less than the preset value, ending the empirical mode decomposition. The original signal is thus divided into empirical mode components and a residual signal:
performing Hilbert transform on the signal, namely expressing the original signal as a three-dimensional time-frequency characteristic matrix of time, frequency and power;
and 4-3, obtaining the state parameter characteristics F3 of the corresponding time samples in the step 2-5 by using 5 engine state parameter sequences of the engine speed N ' (N), the compressor inlet total pressure Pt1 ' (N), the compressor inlet total temperature Tt1 ' (N), the compressor outlet total temperature Tt3 ' (N) and the exhaust temperature EGT ' (N) in 10 control periods.
6. The method for predicting the aircraft engine surge based on the multi-branch feature fusion network according to claim 1, wherein the step 5) is implemented by constructing and training the multi-branch feature fusion network based on the FC-VGG16-TCN, and specifically comprises the following processes:
step 5-1, establishing a branch network 1, wherein the branch network is a fully Connected (Full Connected) neural network and comprises a 1-layer input layer, a 1-layer hidden layer and a 1-layer output layer, and all the layers are Connected in sequence in a fully Connected mode; the input of the branch network is a time domain statistical characteristic F1 of the sample, and the dimensionality is (9, 1); the output of the branching network is a high-dimensional feature HF1 with a dimension of (m1, 1);
step 5-2, establishing a branch network 2, wherein the branch network adopts a VGG16 structure and consists of 13 layers (convolutional layer + pooling layer + ReLU activation function) and 3 layers (full connection layer + ReLU activation function), wherein the ReLU function: relu (x) max { ax, x }, (0 < a < 1);
the input of the branch network is a time-frequency characteristic F2 of a sample, the dimension of the time-frequency characteristic is (steps, steps, 3) through reconstruction, and the steps are the lengths of engine state parameter sequences corresponding to 10 control cycles; the output of the branching network is a high-dimensional feature HF2 with a dimension of (m2, 1);
step 5-3, establishing a branch Network 3, wherein the branch Network is a Time Convolutional Network (Time Convolutional Network) and comprises a causal Convolutional sum and residual error module; wherein the causal convolution is yt＝w1·xt-2+w2·xt-1+w1·xtThe residual module comprises two layers of convolution and ReLU nonlinear mapping, and Weightnorm and Dropout are added into each layer to normalize the network; the input of the branch network is a state parameter characteristic F3 of a sample, the dimensionality is (5, steps), and the steps are the length of an engine state parameter sequence corresponding to 10 control cycles; the output of the branching network is a high-dimensional feature HF3, whose dimension is (m3, 1).
Step 5-4, splicing three high-dimensional features output by three branch networks by adopting a Stacking algorithm, building a Stacking fusion module, inputting two full convolution layers and one full connection layer, and finally outputting through a Softmax activation function to obtain an engine state S;
and 5-5, selecting a loss function as a multi-classification cross entropy function L, selecting an optimization algorithm as an Adam algorithm, training by using samples, and updating network parameters.
7. The method for predicting the surge of the aircraft engine based on the multi-branch feature fusion network as claimed in claim 1, wherein the step 6) of inputting the real-time operation parameters of the engine into the trained network to detect the instability precursor on line so as to realize the prediction of the surge specifically comprises the following steps:
step 6-1, preprocessing the real-time running data of the engine in the latest 10 control periods according to the preprocessing method in the step 2) to obtain a sample Xt corresponding to the current moment;
step 6-2, substituting Xt into the FC-VGG 16-TCN-based multi-branch feature fusion network trained in the step 5), and outputting the current state St of the engine;
and 6-3, judging the engine running state according to St: and when St is [0, 1, 0], judging that the engine has instability precursors, and giving a surge early warning.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110040087.5A CN112580267A (en) | 2021-01-13 | 2021-01-13 | Aero-engine surge prediction method based on multi-branch feature fusion network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110040087.5A CN112580267A (en) | 2021-01-13 | 2021-01-13 | Aero-engine surge prediction method based on multi-branch feature fusion network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN112580267A true CN112580267A (en) | 2021-03-30 |
Family
ID=75145378
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110040087.5A Pending CN112580267A (en) | 2021-01-13 | 2021-01-13 | Aero-engine surge prediction method based on multi-branch feature fusion network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112580267A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114037026A (en) * | 2022-01-11 | 2022-02-11 | 江苏南通元辰钢结构制造有限公司 | Air compressor pressurized starting self-checking method based on industrial big data and artificial intelligence |
| CN114548555A (en) * | 2022-02-22 | 2022-05-27 | 大连理工大学 | Axial flow compressor stall surge prediction method based on deep autoregressive network |
| CN114564996A (en) * | 2022-03-03 | 2022-05-31 | 南京航空航天大学 | Method and device for online detection of surge precursors of aero-engine |
| CN114857062A (en) * | 2022-04-13 | 2022-08-05 | 大连理工大学 | Aircraft engine unstability state detecting system based on TCN |
| WO2023010658A1 (en) * | 2021-08-06 | 2023-02-09 | 大连理工大学 | Time dilated convolutional network-based method for alerting of rotating stall in compressor |
| CN115795282A (en) * | 2023-01-30 | 2023-03-14 | 武汉工程大学 | Shock tube dynamic pressure reconstruction method and device, electronic equipment and storage medium |
| EP4216013A1 (en) * | 2022-01-24 | 2023-07-26 | Hamilton Sundstrand Corporation | Incipient compressor surge detection using artificial intelligence |
| WO2023159336A1 (en) * | 2022-02-22 | 2023-08-31 | 大连理工大学 | Deep autoregressive network based prediction method for stalling and surging of axial-flow compressor |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610026A (en) * | 2019-08-13 | 2019-12-24 | 清华大学 | Online recognition method for surge precursors of aero-engine |
| CN112001128A (en) * | 2020-06-10 | 2020-11-27 | 大连理工大学 | Axial flow compressor stall surge prediction method based on deep learning |
-
2021
- 2021-01-13 CN CN202110040087.5A patent/CN112580267A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110610026A (en) * | 2019-08-13 | 2019-12-24 | 清华大学 | Online recognition method for surge precursors of aero-engine |
| CN112001128A (en) * | 2020-06-10 | 2020-11-27 | 大连理工大学 | Axial flow compressor stall surge prediction method based on deep learning |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2023010658A1 (en) * | 2021-08-06 | 2023-02-09 | 大连理工大学 | Time dilated convolutional network-based method for alerting of rotating stall in compressor |
| CN114037026A (en) * | 2022-01-11 | 2022-02-11 | 江苏南通元辰钢结构制造有限公司 | Air compressor pressurized starting self-checking method based on industrial big data and artificial intelligence |
| CN114037026B (en) * | 2022-01-11 | 2022-03-25 | 江苏南通元辰钢结构制造有限公司 | Air compressor pressurized starting self-checking method based on industrial big data and artificial intelligence |
| EP4216013A1 (en) * | 2022-01-24 | 2023-07-26 | Hamilton Sundstrand Corporation | Incipient compressor surge detection using artificial intelligence |
| US11859626B2 (en) | 2022-01-24 | 2024-01-02 | Hamilton Sundstrand Corporation | Incipient compressor surge detection using artificial intelligence |
| CN114548555A (en) * | 2022-02-22 | 2022-05-27 | 大连理工大学 | Axial flow compressor stall surge prediction method based on deep autoregressive network |
| WO2023159336A1 (en) * | 2022-02-22 | 2023-08-31 | 大连理工大学 | Deep autoregressive network based prediction method for stalling and surging of axial-flow compressor |
| CN114564996A (en) * | 2022-03-03 | 2022-05-31 | 南京航空航天大学 | Method and device for online detection of surge precursors of aero-engine |
| CN114857062A (en) * | 2022-04-13 | 2022-08-05 | 大连理工大学 | Aircraft engine unstability state detecting system based on TCN |
| CN114857062B (en) * | 2022-04-13 | 2023-03-24 | 大连理工大学 | Aircraft engine unstability state detecting system based on TCN |
| CN115795282A (en) * | 2023-01-30 | 2023-03-14 | 武汉工程大学 | Shock tube dynamic pressure reconstruction method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112580267A (en) | Aero-engine surge prediction method based on multi-branch feature fusion network | |
| CN111598438B (en) | Civil aviation engine gas circuit anomaly detection method based on piecewise fitting analysis and evaluation | |
| CN112001128B (en) | Axial flow compressor stall surge prediction method based on deep learning | |
| CN110131193B (en) | Method and system for monitoring surge fault of aircraft engine | |
| CN109211548A (en) | A kind of mechanical failure diagnostic method | |
| CN113569338B (en) | Gas compressor rotating stall early warning method based on time expansion convolutional network | |
| CN111474919B (en) | Aeroengine control system sensor fault diagnosis method based on AANN network group | |
| CN111400961B (en) | Wind generating set blade fault judging method and device | |
| CN111506049B (en) | Multiple fault diagnosis method for aero-engine control system based on AANN network system | |
| CN108801641A (en) | The fault diagnosis of exhaust-driven turbo-charger exhaust-gas turbo charger and Reliability Prediction Method and its system | |
| CN111415070A (en) | Wind turbine generator gearbox oil temperature over-temperature fault early warning method based on SCADA data | |
| CN103925155A (en) | Self-adaptive detection method for abnormal wind turbine output power | |
| WO2024045246A1 (en) | Spike echo state network model for aero engine fault prediction | |
| CN111931851A (en) | Fan blade icing fault diagnosis method based on one-dimensional residual error neural network | |
| CN114708885A (en) | Fan fault early warning method based on sound signals | |
| CN114741945B (en) | On-line fault diagnosis method for aero-engine | |
| CN114151320B (en) | Identification algorithm for instability of compressor flow system | |
| Feng et al. | Gas turbine blade fracturing fault diagnosis based on broadband casing vibration | |
| CN109872511B (en) | Self-adaptive two-stage alarm method for monitoring axial displacement sudden change | |
| CN114169718A (en) | Method for improving reliability of wind turbine generator based on state evaluation of wind turbine generator | |
| CN110826600B (en) | Engine surge prediction method based on adaptive resonance network online incremental learning | |
| Quan et al. | Detection of Rotating Stall Inception of Axial Compressors Based on Deep Dilated Causal Convolutional Neural Networks | |
| CN117272753A (en) | Power transmission line icing galloping early warning method based on process parameters | |
| CN112418306A (en) | Gas turbine compressor fault early warning method based on LSTM-SVM | |
| CN116840631A (en) | Transformer partial discharge monitoring and positioning method based on joint diagnosis model |
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 |