CN109000930B - Turbine engine performance degradation evaluation method based on stacking denoising autoencoder - Google Patents
Turbine engine performance degradation evaluation method based on stacking denoising autoencoder Download PDFInfo
- Publication number
- CN109000930B CN109000930B CN201810565712.6A CN201810565712A CN109000930B CN 109000930 B CN109000930 B CN 109000930B CN 201810565712 A CN201810565712 A CN 201810565712A CN 109000930 B CN109000930 B CN 109000930B
- Authority
- CN
- China
- Prior art keywords
- data
- network
- denoising
- self
- health factor
- 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
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M15/00—Testing of engines
- G01M15/14—Testing gas-turbine engines or jet-propulsion engines
Landscapes
- Chemical & Material Sciences (AREA)
- Engineering & Computer Science (AREA)
- Combustion & Propulsion (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Testing Of Engines (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
A turbine engine performance degradation evaluation method based on a stacked denoising autoencoder is used for the technical field of engine performance degradation evaluation. The invention solves the problems that the traditional multi-sensor data selection needs to rely on a complex information evaluation criterion, the extraction of degradation characteristics during HI construction needs to rely on a large number of signal processing technologies and expert experiences, the label selection in a supervised training mode needs to depend on manual participation, and the method is not universal enough. The 4 denoising self-coding machines construct a stacked denoising self-coder to extract a single node value of input data, a training set data is used for fine tuning parameters by using a BP algorithm after a network is pre-trained, the extracted single node value is used as a health factor value of each cycle, and an HI curve of the training set is constructed; inputting the test set into a trained stacked denoising self-encoder to obtain a health factor value of each cycle, and constructing an HI curve; and respectively carrying out smoothing treatment on the HI curves of the training set and the testing set, and evaluating the HI curves after the smoothing treatment.
Description
Technical Field
The invention belongs to the technical field of engine performance degradation evaluation, and particularly relates to a turbine engine performance degradation evaluation method based on a stacking denoising autoencoder.
Background
The turbine engine is one of the common and important aircraft components, and the guarantee of the reliable operation state of the turbine engine has very important practical significance for the stable operation of the aircraft and the reduction of the maintenance cost. The health factor of a turbine engine, as a characteristic quantity for evaluating its health level, may be indicative of the state or degree of degradation of the engine's health level. The method is obtained by mapping different degradation performance variables in the operation process, the constructed health factor curve characterizes the degradation process of the engine performance as a monotone ascending or descending trend, and the residual life of equipment can be obtained by predicting whether the change of the health factor curve reaches a degradation failure threshold value, so that the well constructed health factor curve has important significance for monitoring the operation condition of the turbine engine and predicting the residual life.
The method based on data driving benefits from the development of sensor technology and storage technology, can utilize a large number of sensors to monitor data and mine degradation information, reduces the dependence on expert experience, and thus gradually becomes a mainstream mode. According to the characteristics of the health state of the target object reflected by the processing data, namely health factors representing the system degradation behavior and the service life of the object, data-driven methods can be divided into two main categories, namely direct prediction and indirect prediction. Because the direct prediction method directly uses the original data as the health factor of the tested object, the general situation is difficult to satisfy better trend, and the monitoring process of the turbine engine is often completed by cooperation of a plurality of sensors, and the original monitoring data cannot be directly used as the health factor, the indirect turbine engine HI construction method is widely researched by domestic and foreign scholars.
In the indirect HI construction process, beneficial information contained in each sensor is often evaluated according to a certain evaluation index, and monitoring data of the sensor is selected; and then, a feature extraction process is carried out to obtain a feature set so as to carry out more advanced characterization on the original data, feature selection is carried out on the basis of the feature set to remove redundant features, feature fusion is carried out continuously under necessary conditions, and the health state of the equipment is reflected by combining a plurality of features. The sensor selection method such as manually observing data trend, principal component analysis, calculating sensor data trend according to the permutation entropy, and the like, and the feature extraction method is taken as a key step, and mainly comprises a method based on a traditional signal processing technology and a machine learning method. For example, Liyongxiang et al firstly performs dimensionality reduction on multi-sensor data by using principal component analysis, and then obtains an HI curve by means of a weighted Euclidean distance and a regression algorithm; after screening sensor information, RachaKhelif et al obtain a HI curve by linear regression and curve fitting; in addition, there are cluster analysis, wiener process, vector machine model, and the like.
The deep learning obtains remarkable results in the fields of image processing, voice recognition and the like due to the powerful self-adaptive feature extraction capability and the nonlinear function representation capability. Since tamiselvan et al first applied deep learning to the field of fault diagnosis in 2013, typical deep learning algorithms such as a deep neural network and a convolutional neural network have been increasingly widely applied to the field of health management of devices. For example, Pankaj Malhotra et al first performs principal component analysis on original sensor information to perform optimization selection, then encodes and decodes the information by using a long-time and short-time memory network, and then maps a reconstruction error into an HI curve through a linear regression model. Although the traditional data-driven approach has achieved significant success in the construction of turbine engine HI, the following problems remain: the original sensor information is often optimized and selected by means of a complex sensor information evaluation criterion, and a large amount of expert experience and a traditional signal processing method are still needed for extracting degradation characteristics; part of HI construction model training usually adopts a supervision mode, namely, a real output value corresponding to input needs to be provided as a label in the training process, and the label selection needs to depend on manual participation, so that time is consumed and no consistent standard exists; in order to obtain a comprehensive monotonic HI curve, multiple signal processing methods are often adopted for fusion and parameters are selected depending on manual experience aiming at specific prediction problems, and certain universality is lacked.
Disclosure of Invention
The invention aims to solve the problems that the traditional method for constructing the health factor (HI) curve of the turbine engine needs to optimize and select the original sensor information by means of a complex sensor information evaluation criterion, and the extraction of degradation characteristics still needs to rely on a large amount of expert experience; the HI curve building model is still usually trained in a supervised mode, and the selection of training labels is time-consuming and has no consistent standard; needs the fusion of a plurality of signal processing methods and lacks certain universality.
The technical scheme adopted by the invention for solving the technical problems is as follows:
a performance degradation evaluation method of a turbine engine based on a stacked denoising autoencoder comprises the following specific steps:
acquiring monitoring data of all sensors of the turbine engine by using N monitoring units; observing and screening out data x of sensor with monitoring data changedkK is 1,2, …, Q, and Q is the number of sensors with change of monitoring data;
step two, screening the data x of each sensor screened in the step onekNormalized to [0,1 respectively]Within the interval; taking the sensor data of one part of the screened monitoring units as training set data, and taking the sensor data of the other part of the screened monitoring units as test set data;
establishing a stacked denoising self-encoder network consisting of 4 denoising automatic encoders for performing feature extraction on training set data and test set data, wherein the first denoising self-encoder and the second denoising self-encoder form an encoding network of the stacked denoising self-encoder network, and the third denoising self-encoder and the fourth denoising self-encoder form a decoding network of the stacked denoising self-encoder network;
inputting the training set data determined in the second step into a coding network of the stacking denoising self-coder network, and enabling the training set data to pass through the unsupervised pre-training of a first denoising self-coder and a second denoising self-coder of the coding network in sequence to obtain a coding parameter theta 1 ═ W ═ of the first denoising self-coder of the coding network1,b1And coding parameters theta 2 ═ W of the second denoising self-coding machine2,b2}；
Setting the coding weight W of a third denoising self-coding machine of a decoding network3Is W2Transpose of (3), fourth denoising autoencoder's encoding weight W4Is W1Transposing; after the pre-training is finished, fine tuning is carried out on the network parameters of the stacking denoising self-encoder network by using a BP algorithm; obtaining a single node value output by the encoding network by using the network parameters after fine tuning;
step four, the single node value extracted in the step three is used as a health factor value at the turbine engine cycle, and a health factor curve on a training set is constructed;
inputting the test set data into the stacked self-encoder network trained in the step three, and extracting a single node value through a plurality of hidden layers; constructing a health factor curve on the test set by using the method of the fourth step;
step six, smoothing the health factor curves constructed in the step four and the step five respectively to obtain a training set health factor curve and a test set health factor curve after smoothing;
step seven, respectively calculating the time relevance and monotonicity of the training set health factor curve and the test set health factor curve; and evaluating the performance degradation condition of the turbine engine by using the smoothed test set health factor curve.
The invention has the beneficial effects that: the invention provides a turbine engine performance degradation evaluation method based on a stack denoising autoencoder, which can screen out data of a sensor with monitoring data changing by adopting a method of directly observing the data, does not need to analyze and process the sensor according to complex indexes, and simplifies the processing process; the method utilizes 4 denoising self-coding mechanisms to build a stacked denoising self-coding machine to extract a single node value at each cycle of the turbine engine, is different from the existing HI curve building model training which adopts a supervision mode, and the training set data can be trained in an unsupervised mode; compared with the traditional method, the method can screen out useful sensor data through observation, and the extracted sensor data can be subjected to unsupervised training of the stacking denoising self-coding machine to extract a single node value, namely, the degradation condition of the turbine engine is predicted without adopting various signal processing methods and without depending on manual work and expert experience, so that the method has good universality.
The health factor curve of the turbine engine constructed by the method can better depict the degradation trend of the health condition of the turbine engine in the whole life cycle, the local oscillation is smaller, the curve is smoother, and the time relevance of the health factor curve of the turbine engine is improved by about 82% and the monotonicity is improved by about 7.4 times compared with the curve of the existing method, so that the degradation condition of the performance of the turbine engine can be better estimated.
Drawings
FIG. 1 is a flow chart of a method for evaluating performance degradation of a turbine engine based on a stacked denoising autoencoder according to the present invention;
FIG. 2 is a block diagram of a stacked denoising autoencoder according to the present invention;
wherein: the parts shared within the dashed box of the encoding network and the decoding network are: the output layer of the coding network is also the input layer of the decoding network;
FIG. 3 is a schematic diagram of a first denoising self-coding machine according to the present invention;
h is a hidden layer;
FIG. 4 is a turbine engine simulation illustration of an embodiment of the present invention;
wherein: 1 for a Fan (Fan), 2 for a combustion chamber (Combustor), 3 for a high-pressure channel (N1), 4 for a low-pressure turbine (LPT), 5 for a low-pressure compressor (LPC), 6 for a high-pressure compressor (HPC), 7 for a low-pressure channel (N2), 8 for a high-pressure turbine (HPT), 9 for a Nozzle (Nozzle);
FIG. 5 is a diagram of various modules and interconnections of a turbine engine according to an embodiment of the present invention;
wherein: the external environment is Ambient, an Inlet is an Inlet, a flow divider is a Splitter, a Bypass channel is Bypass Path, a Bypass Nozzle is a Bypass Nozzle, Fuel is Fuel, a Burner is a Burner, and a Core Nozzle is a Core Nozzle.
FIG. 6 is a graph of degradation data on sensor # 2 for a monitoring unit according to an embodiment of the present invention;
FIG. 7 is a graph of degradation data for a monitoring unit on sensor # 7 in accordance with an embodiment of the present invention;
FIG. 8 is a graph of degradation data on sensor #14 for a monitoring unit in accordance with an embodiment of the present invention;
FIG. 9 is a training set health factor graph of a turbine engine after smoothing according to an embodiment of the present invention;
FIG. 10 is a test set health factor graph of a turbine engine after smoothing according to an embodiment of the present invention.
Detailed Description
The technical solution of the present invention is further described below with reference to the accompanying drawings, but not limited thereto, and any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention shall be covered by the protection scope of the present invention.
The first embodiment is as follows: the present embodiment will be described with reference to fig. 1 and 2. The embodiment of the invention discloses a turbine engine performance degradation evaluation method based on a stack denoising autoencoder, which comprises the following specific steps:
acquiring monitoring data of all sensors of the turbine engine by using N monitoring units; observing and screening out data x of sensor with monitoring data changedkK is 1,2, …, Q, and Q is the number of sensors with change of monitoring data;
step two, screening the data x of each sensor screened in the step onekNormalized to [0,1 respectively]Within the interval; taking the sensor data of one part of the screened monitoring units as training set data, and taking the sensor data of the other part of the screened monitoring units as test set data;
establishing a stacked denoising self-encoder network consisting of 4 denoising automatic encoders for performing feature extraction on training set data and test set data, wherein the first denoising self-encoder and the second denoising self-encoder form an encoding network of the stacked denoising self-encoder network, and the third denoising self-encoder and the fourth denoising self-encoder form a decoding network of the stacked denoising self-encoder network;
inputting the training set data determined in the second step into a coding network of the stacking denoising self-coder network, and enabling the training set data to pass through the unsupervised pre-training of a first denoising self-coder and a second denoising self-coder of the coding network in sequence to obtain a coding parameter theta 1 ═ W ═ of the first denoising self-coder of the coding network1,b1And coding parameters theta 2 ═ W of the second denoising self-coding machine2,b2}；
Setting the coding weight W of a third denoising self-coding machine of a decoding network3Is W2Transpose of (3), fourth denoising autoencoder's encoding weight W4Is W1Transposing; after the pre-training is finished, fine tuning is carried out on the network parameters of the stacking denoising self-encoder network by using a BP algorithm; obtaining a single node value output by the encoding network by using the network parameters after fine tuning;
step four, the single node value extracted in the step three is used as a health factor value at the turbine engine cycle, and a health factor curve on a training set is constructed;
inputting the test set data into the stacked self-encoder network trained in the step three, and extracting a single node value through a plurality of hidden layers; constructing a health factor curve on the test set by using the method of the fourth step;
step six, smoothing the health factor curves constructed in the step four and the step five respectively to obtain a training set health factor curve and a test set health factor curve after smoothing;
step seven, respectively calculating the time relevance and monotonicity of the training set health factor curve and the test set health factor curve; and evaluating the performance degradation condition of the turbine engine by using the smoothed test set health factor curve.
The stack denoising autoencoder network established in the embodiment can be composed of an even number of denoising autocoders of 2, 4, 6, 8 or more than 8, and the number of network nodes of each denoising autocoder can be specifically determined according to the actual condition of data; and the invention extracts a single node value at each cycle, and constructs an HI curve according to the extracted node values. In order to make the constructed HI curve have a strengthening tendency and retain local details, we select the smoothing processing times to be 15 times according to the actual curve results.
After the information of a plurality of sensors is collected, the method can remove unchanged values to directly obtain input data only by observation, and compared with the existing method, the method needs to analyze and process the sensors according to complex indexes.
By observing the smoothed test set health factor curve, the performance degradation condition of the current state of the turbine engine can be evaluated, and the future operating condition of the turbine engine can be predicted by further extrapolating the health factor curve.
The second embodiment is as follows: the embodiment further defines the method for evaluating the performance degradation of the turbine engine based on the stacked denoising self-encoder in the first embodiment, and the process of data normalization of each sensor in the second step is as follows:
normalized in the manner of xk *＝(xk-xk,min)/(xk,max-xk,min) Wherein x isk *Is each sensor data xkNormalized value, xk,maxAnd xk,minCorresponding to the maximum and minimum values of each sensor over each cycle of the turbine engine, respectively.
The third concrete implementation mode: this embodiment will be described with reference to fig. 3. The embodiment further defines the method for evaluating the performance degradation of the turbine engine based on the stacked denoising self-encoder in the second embodiment, and the working principle of the first denoising self-encoder is as follows:
taking training set data as input data of a first denoising self-coding machine of the stacked denoising self-coding network, wherein the first denoising self-coding machine performs a random mapping function qDFor input data xkDestroying to obtain data after adding noise
Through a coding process fθ1Generating output of a hidden layer
Output of hidden layer
Then go through decoding process gθ1'Generating reconstruction data z; input data xkThe difference with the reconstructed data z is taken as the reconstruction error LH(xkZ) for training;
encoding process fθ1The specific process is as follows:
where s is a sigmoid activation function, W1Is the coding weight of the first denoise self-coder, b1The method comprises the steps that (1) the coding bias of a first denoising automatic coding machine is obtained, and theta 1 is a pre-trained coding parameter of the first denoising automatic coding machine;
θ1＝{W1,b1} (2)
decoding process gθ1'The specific process is as follows:
wherein W1' is the decoding weight of the first denoised self-coder of the coding network, b1'is a decoding bias of a first denoising auto-encoder of the encoding network, and theta 1' is a decoding parameter of the first denoising auto-encoder of the encoding network;
θ1'＝{W1',b1'} (4)
reconstruction error LH(xk,z)＝||xk-z||2Wherein, | | · | | represents a 2 norm;
applying an objective function by using a gradient descent algorithm
Minimization, enhancement of de-noising from input data x of self-encoding machinekWhere n is the training setNumber of samples, xk (i)Is the ith sample data of the first image,
the data is the data of the ith sample data added with noise, i is 1,2, …, n;
the working principle of the second denoising self-coding machine, the third denoising self-coding machine and the fourth denoising self-coding machine is the same as that of the first denoising self-coding machine.
The fourth concrete implementation mode: the embodiment further defines the method for evaluating the performance degradation of the turbine engine based on the stacked denoising autoencoder, which is described in the third embodiment, and the third step specifically includes:
the structure of the stacked denoising self-encoder network comprises an input layer, a hidden layer of the encoding network, an output layer of the encoding network, a hidden layer of the decoding network and an output layer of the decoding network;
after the network parameters of the stacking denoising self-encoder are randomly initialized, the training set data sequentially passes through the unsupervised pre-training of a first denoising self-encoder and a second denoising self-encoder in the encoding network; output of a reserved hidden layer after pre-training of a first denoising self-coding machine is completed
And will imply the output of the layer
The second denoise self-encoder is used as the input of the second denoise self-encoder to complete the unsupervised pre-training of the training set data in the encoding network;
the specific process of using the original data of the training set as the label and utilizing the BP algorithm to finely adjust the network parameters of the stacking denoising self-encoder network comprises the following steps:
assume the raw data of the training set as
Wherein the content of the first and second substances,
the method comprises the steps of (1) training an mth original sample of original data of a training set, wherein the value range of M is 1-M; the hidden layer output of the first denoised self-coder of the stacked denoised self-coder network is
Outputting a first denoised hidden layer from an encoder
As the input of the second denoising autoencoder, the hidden layer output of the second denoising autoencoder is
By analogy, the output of the fourth denoising autoencoder of the stacked denoising autoencoder network is
Original sample of training set
As the label value, the error function phi (theta) is calculated
Wherein Θ ═ θ1,θ2,θ3,θ4The updating mode of the parameters is
Wherein α is the learning rate during parameter fine tuning;
and extracting a single node value output by the encoding network by using the network parameters after fine adjustment.
The structure of the stacked denoising self-encoder network in the embodiment is 14-7-1-7-14, wherein 14 is a 14-dimensional sensor value of turbine engine data, and 14-7-14 and 7-1-7 respectively form 2 denoising self-encoders corresponding to the input layer to the output layer of the encoding network.
The fifth concrete implementation mode: the embodiment further defines the method for evaluating the performance degradation of the turbine engine based on the stacked denoising autoencoder, which is described in the fourth embodiment, and the specific process of smoothing the health factor curve in the sixth step is as follows:
setting the length of a health factor curve as L1, the window width of a filter as L2 and L2 as an odd number; returning a vector with the same length as the health factor curve, and noting that the data points at the left end and the right end of the smooth point are the same during smoothing:
by the way of analogy, the method can be used,
wherein k is 1,2, …, L1;
yy(L1)＝y(L1)；
and y is the health factor data corresponding to each cycle of the turbine engine in the health factor curve, and yy is the health factor data corresponding to each cycle after the health factor curve is subjected to smoothing filtering.
For example, when the filter window width is 15, yy (1) is equal to y (1),
by the way of analogy, the method can be used,
the sixth specific implementation mode: the embodiment further defines the turbine engine performance degradation estimator based on the stacked denoising autoencoder described in the fifth embodiment, and the specific process of calculating the time correlation and monotonicity of the health factor curve in the seventh step is as follows:
the time relevance and monotonicity are two indexes for evaluating the health factor curve, the former represents the linear correlation degree of the health factor value and the running time, and the latter measures the monotonous change trend condition of the health factor curve;
the time correlation of the health factor curves for the jth monitoring unit of the training set is as follows:
wherein, yytjRepresenting the value of the health factor curve at the t-th cycle of the health factor curve of the jth monitoring unit of the training set, wherein j is 1,2, … and K, and K belongs to N and is the number of the monitoring units corresponding to the training set data; ltjNumber, T, representing the jth monitoring unit turbine engine cyclejFor the training set jth monitoring unit turbine engine health factor curve length,
the average value of the corresponding curve values at each cycle of the jth monitoring unit health factor curve of the training set is obtained,
average value of the cycle number of the turbine engine of the jth monitoring unit;
the monotonicity of the health factor curve for the jth monitoring unit of the training set is as follows:
wherein dFjIs the differential between the sequence values in the jth monitoring unit health factor curve; num of dFj>0 represents dF greater than 0jNumber of values, Num of dFj<0 represents less thandF of 0jThe number of values;
then the health factor curves of K monitoring units in the training set are evaluated in a whole manner, and the average value Corr of the time correlation of the K curvesavgComprises the following steps:
mean value Mon of monotonicity of K curvesavgComprises the following steps:
the time relevance and monotonicity calculation modes of the test set health factor curve and the training set health factor curve are the same.
Examples
The invention selects C-MAPSS which is an aircraft engine simulation state monitoring public data set provided by the NASA central of Excellence (PCoE).
An engine simulation diagram is shown in fig. 4, and the different modules and interconnections are shown in fig. 5. The number of the monitoring units for acquiring the monitoring data of all the sensors of the turbine engine is 200, namely the data of 100 monitoring units in 200 monitoring units is used as training set data, and the data of the other 100 monitoring units is used as test set data.
In the embodiment, the HI curve construction is performed by using files of "train _ FD 001" and "test _ FD 001" provided by a data set as a training set and a test set respectively, the data are 21 sensor parameters continuously monitored under a single working condition, and the physical meanings of the sensors are shown in table 1.
The time series of 200 monitoring units are different in length. The training and test sets comprise 20631 and 13096 cycles, respectively. During the experiment, all monitoring units started to degrade from slight loss states of different degrees.
Analysis of the data reveals that 14-dimensional information in 21-dimensional sensor information has different degrees of tendency (#2, #3, #4, #7, #8, #9, #11, #12, #13, #14, #15, #17, #20, #21), and the remaining 7-dimensional monitoring information has no change (#1, #5, #6, #10, #16, #18, #19), and has no useful information.
Fig. 6, 7 and 8 respectively show a specific degradation process schematic diagram by taking a sensor # 2 with all monitoring units having an ascending trend, a sensor # 7 with all monitoring units having a descending trend, and a sensor #14 with all monitoring units having inconsistent change trends as an example.
Inputting the monitoring signal data of the original sensor into the stacking denoising self-encoder network, and obtaining a training set health factor curve graph of the turbine engine after smoothing processing as shown in FIG. 9 by the method of the invention;
the test set health factor graph of the smoothed turbine engine, as shown in FIG. 10, may be used to evaluate the degradation of the turbine engine performance based on the obtained smoothed turbine engine health factor graph and based on the time that the turbine engine has been operating.
Through algorithm verification, compared with the traditional method, the method for constructing the health factor curve of the turbine engine has the advantages that the time relevance of the obtained health factor curve of the turbine engine is improved by about 82%, and the monotonicity is improved by about 7.4 times.
TABLE 1 aircraft Engine Multi-sensor information
Claims (5)
1. A performance degradation evaluation method of a turbine engine based on a stacked denoising autoencoder is characterized by comprising the following specific steps:
acquiring monitoring data of all sensors of the turbine engine by using N monitoring units; observing and screening out data x of sensor with monitoring data changedkK is 1,2, …, Q, and Q is the number of sensors with change of monitoring data;
step two, screening the data x of each sensor screened in the step onekAre respectively normalized to[0,1]Within the interval; taking the sensor data of one part of the screened monitoring units as training set data, and taking the sensor data of the other part of the screened monitoring units as test set data;
establishing a stacked denoising self-encoder network consisting of 4 denoising automatic encoders for performing feature extraction on training set data and test set data, wherein the first denoising self-encoder and the second denoising self-encoder form an encoding network of the stacked denoising self-encoder network, and the third denoising self-encoder and the fourth denoising self-encoder form a decoding network of the stacked denoising self-encoder network;
inputting the training set data determined in the second step into a coding network of the stacking denoising self-coder network, and enabling the training set data to pass through the unsupervised pre-training of a first denoising self-coder and a second denoising self-coder of the coding network in sequence to obtain a coding parameter theta 1 ═ W ═ of the first denoising self-coder of the coding network1,b1And coding parameters theta 2 ═ W of the second denoising self-coding machine2,b2}；
Setting the coding weight W of a third denoising self-coding machine of a decoding network3Is W2Transpose of (3), fourth denoising autoencoder's encoding weight W4Is W1Transposing; after the pre-training is finished, fine tuning is carried out on the network parameters of the stacking denoising self-encoder network by using a BP algorithm; obtaining a single node value output by the encoding network by using the network parameters after fine tuning;
the specific process of the third step is as follows:
the structure of the stacked denoising self-encoder network comprises an input layer, a hidden layer of the encoding network, an output layer of the encoding network, a hidden layer of the decoding network and an output layer of the decoding network;
after the network parameters of the stacking denoising self-encoder are randomly initialized, the training set data sequentially passes through the unsupervised pre-training of a first denoising self-encoder and a second denoising self-encoder in the encoding network; output of a reserved hidden layer after pre-training of a first denoising self-coding machine is completed
And will imply the output of the layer
The second denoise self-encoder is used as the input of the second denoise self-encoder to complete the unsupervised pre-training of the training set data in the encoding network;
the specific process of using the original data of the training set as the label and utilizing the BP algorithm to finely adjust the network parameters of the stacking denoising self-encoder network comprises the following steps:
assume the raw data of the training set as
Wherein the content of the first and second substances,
the method comprises the steps of (1) training an mth original sample of original data of a training set, wherein the value range of M is 1-M; the hidden layer output of the first denoised self-coder of the stacked denoised self-coder network is
Outputting a first denoised hidden layer from an encoder
As the input of the second denoising autoencoder, the hidden layer output of the second denoising autoencoder is
By analogy, the output of the fourth denoising autoencoder of the stacked denoising autoencoder network is
Original sample of training set
As the label value, the error function phi (theta) is calculated
Wherein Θ ═ θ1,θ2,θ3,θ4The updating mode of the parameters is
Wherein α is the learning rate during parameter fine tuning;
extracting a single node value output by the encoding network by using the network parameters after fine adjustment;
step four, the single node value extracted in the step three is used as a health factor value at the turbine engine cycle, and a health factor curve on a training set is constructed;
inputting the test set data into the stacked self-encoder network trained in the step three, and extracting a single node value through a plurality of hidden layers; constructing a health factor curve on the test set by using the method of the fourth step;
step six, smoothing the health factor curves constructed in the step four and the step five respectively to obtain a training set health factor curve and a test set health factor curve after smoothing;
step seven, respectively calculating the time relevance and monotonicity of the training set health factor curve and the test set health factor curve; and evaluating the performance degradation condition of the turbine engine by using the smoothed test set health factor curve.
2. The method for evaluating the performance degradation of a turbine engine based on a stacked denoising self-encoder as claimed in claim 1, wherein the data normalization process of each sensor in the second step is as follows:
normalized in the manner of xk *＝(xk-xk,min)/(xk,max-xk,min) Wherein x isk *Is each sensor data xkNormalized value, xk,maxAnd xk,minCorresponding to the maximum and minimum values of each sensor over each cycle of the turbine engine, respectively.
3. The method for evaluating the performance degradation of the turbine engine based on the stacked denoising self-encoder as claimed in claim 2, wherein the first denoising self-encoder works by:
taking training set data as input data of a first denoising self-coding machine of the stacked denoising self-coding network, wherein the first denoising self-coding machine performs a random mapping function qDFor input data xkDestroying to obtain data after adding noise
Through a coding process fθ1Generating output of a hidden layer
Output of hidden layer
Then go through decoding process gθ1'Generating reconstruction data z; input data xkThe difference with the reconstructed data z is taken as the reconstruction error LH(xkZ) for training;
encoding process fθ1The specific process is as follows:
where s is a sigmoid activation function, W1Is the coding weight of the first denoise self-coder, b1The method comprises the steps that (1) the coding bias of a first denoising automatic coding machine is obtained, and theta 1 is a pre-trained coding parameter of the first denoising automatic coding machine;
θ1＝{W1,b1} (2)
decoding process gθ1'The specific process is as follows:
wherein W1' is the decoding weight of the first denoised self-coder of the coding network, b1'is a decoding bias of a first denoising auto-encoder of the encoding network, and theta 1' is a decoding parameter of the first denoising auto-encoder of the encoding network;
θ1'＝{W1',b1'} (4)
reconstruction error LH(xk,z)＝||xk-z||2Wherein, in the step (A),
represents a 2 norm;
applying an objective function by using a gradient descent algorithm
Minimization, enhancement of de-noising from input data x of self-encoding machinekWhere n is the number of training set samples, xk (i)Is the ith sample data of the first image,
the data is the data of the ith sample data added with noise, i is 1,2, …, n;
the working principle of the second denoising self-coding machine, the third denoising self-coding machine and the fourth denoising self-coding machine is the same as that of the first denoising self-coding machine.
4. The method as claimed in claim 3, wherein the smoothing of the health factor curve in the sixth step is performed by:
setting the length of a health factor curve as L1, the window width of a filter as L2 and L2 as an odd number; returning vectors with the same length as the health factor curve, wherein the data points at the left end and the right end of the smoothing point are the same in number during smoothing treatment:
yy(1)＝y(1)，
by the way of analogy, the method can be used,
and y is the health factor data corresponding to each cycle of the turbine engine in the health factor curve, and yy is the health factor data corresponding to each cycle after the health factor curve is subjected to smoothing filtering.
5. The method for evaluating the performance degradation of the turbine engine based on the stacked denoising self-encoder as claimed in claim 4, wherein the specific process of calculating the time correlation and monotonicity of the health factor curve in the seventh step is as follows:
the time correlation of the health factor curves for the jth monitoring unit of the training set is as follows:
wherein, yytjRepresenting the value of the health factor curve at the t-th cycle of the health factor curve of the jth monitoring unit of the training set, wherein j is 1,2, … and K, and K belongs to N and is the number of the monitoring units corresponding to the training set data; ltjNumber, T, representing the jth monitoring unit turbine engine cyclejFor the training set jth monitoring unit turbine engine health factor curve length,
the average value of the corresponding curve values at each cycle of the jth monitoring unit health factor curve of the training set is obtained,
average value of the cycle number of the turbine engine of the jth monitoring unit;
the monotonicity of the health factor curve for the jth monitoring unit of the training set is as follows:
wherein dFjIs the differential between the sequence values in the jth monitoring unit health factor curve; num of dFj0 denotes dF > 0jNumber of values, Num of dFj< 0 denotes dF of less than 0jThe number of values;
then the health factor curves of K monitoring units in the training set are evaluated in a whole manner, and the average value Corr of the time correlation of the K curvesavgComprises the following steps:
mean value Mon of monotonicity of K curvesavgComprises the following steps:
the time relevance and monotonicity calculation modes of the test set health factor curve and the training set health factor curve are the same.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565712.6A CN109000930B (en) | 2018-06-04 | 2018-06-04 | Turbine engine performance degradation evaluation method based on stacking denoising autoencoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810565712.6A CN109000930B (en) | 2018-06-04 | 2018-06-04 | Turbine engine performance degradation evaluation method based on stacking denoising autoencoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109000930A CN109000930A (en) | 2018-12-14 |
| CN109000930B true CN109000930B (en) | 2020-06-16 |
Family
ID=64574283
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810565712.6A Active CN109000930B (en) | 2018-06-04 | 2018-06-04 | Turbine engine performance degradation evaluation method based on stacking denoising autoencoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109000930B (en) |
Families Citing this family (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111175054B (en) * | 2020-01-08 | 2021-12-03 | 沈阳航空航天大学 | Aeroengine fault diagnosis method based on data driving |
| CN111310830B (en) * | 2020-02-17 | 2023-10-10 | 湖北工业大学 | Blocking fault diagnosis system and method for combine harvester |
| CN111289250A (en) * | 2020-02-24 | 2020-06-16 | 湖南大学 | Method for predicting residual service life of rolling bearing of servo motor |
| CN112182854B (en) * | 2020-09-09 | 2022-08-16 | 中南大学 | Intelligent monitoring method and system for abnormal furnace conditions of blast furnace |
| CN112101659A (en) * | 2020-09-15 | 2020-12-18 | 安徽三禾一信息科技有限公司 | Complex equipment service life prediction method based on stacking denoising autoencoder |
| CN112232404B (en) * | 2020-10-13 | 2023-12-22 | 中国铁路上海局集团有限公司南京供电段 | Reliability calculation method and system based on historical abnormality and operation and maintenance information of railway power supply equipment |
| CN112633476B (en) * | 2020-12-24 | 2023-09-22 | 东北大学 | Engine state monitoring method, system, equipment and medium |
| CN112816191B (en) * | 2020-12-28 | 2022-07-29 | 哈尔滨工业大学 | Multi-feature health factor fusion method based on SDRSN |
| CN112597705B (en) * | 2020-12-28 | 2022-05-24 | 哈尔滨工业大学 | Multi-feature health factor fusion method based on SCVNN |
| CN113051809A (en) * | 2021-03-12 | 2021-06-29 | 中国人民解放军海军工程大学 | Virtual health factor construction method based on improved restricted Boltzmann machine |
| CN113111575B (en) * | 2021-03-30 | 2023-03-31 | 西安交通大学 | Combustion engine degradation evaluation method based on depth feature coding and Gaussian mixture model |
| CN113239979B (en) * | 2021-04-23 | 2024-01-09 | 广州市祺能电子科技有限公司 | Method and device for acquiring data of sensor of Internet of things |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1258618A2 (en) * | 2001-05-18 | 2002-11-20 | General Electric Company | System and method for monitoring engine performance in a gas turbine engine |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102789545B (en) * | 2012-07-12 | 2015-08-19 | 哈尔滨工业大学 | Based on the Forecasting Methodology of the turbine engine residual life of degradation model coupling |
| CN104159245B (en) * | 2014-08-22 | 2017-08-25 | 哈尔滨工业大学 | Towards the indirect health factor preparation method of radio data-transmission equipment |
| CN104748962B (en) * | 2015-04-03 | 2017-02-22 | 西安交通大学 | Planetary gear box intelligent diagnosis method based on stacking automatic encoding machine |
| US10330664B2 (en) * | 2015-06-18 | 2019-06-25 | Pratt & Whitney Canada Corp. | Evaluation of component condition through analysis of material interaction |
| CN105547705B (en) * | 2016-03-10 | 2018-07-20 | 西安工业大学 | Engine performance degradation trend prediction technique |
| CN106682688B (en) * | 2016-12-16 | 2020-07-28 | 华南理工大学 | Particle swarm optimization-based stacked noise reduction self-coding network bearing fault diagnosis method |
| CN107702922B (en) * | 2017-09-11 | 2020-10-09 | 南京信息工程大学 | Rolling bearing fault diagnosis method based on LCD and stacking automatic encoder |
| CN107766204A (en) * | 2017-10-10 | 2018-03-06 | 曙光信息产业(北京)有限公司 | A kind of method and system for checking cluster health status |
| CN107941537B (en) * | 2017-10-25 | 2019-08-27 | 南京航空航天大学 | A kind of mechanical equipment health state evaluation method |
| CN107957551A (en) * | 2017-12-12 | 2018-04-24 | 南京信息工程大学 | Stacking noise reduction own coding Method of Motor Fault Diagnosis based on vibration and current signal |
-
2018
- 2018-06-04 CN CN201810565712.6A patent/CN109000930B/en active Active
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP1258618A2 (en) * | 2001-05-18 | 2002-11-20 | General Electric Company | System and method for monitoring engine performance in a gas turbine engine |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109000930A (en) | 2018-12-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109000930B (en) | Turbine engine performance degradation evaluation method based on stacking denoising autoencoder | |
| CN108398268B (en) | Bearing performance degradation evaluation method | |
| CN107657250B (en) | Bearing fault detection and positioning method and detection and positioning model implementation system and method | |
| CN107122790A (en) | Non-intrusion type load recognizer based on hybrid neural networks and integrated study | |
| CN110245745B (en) | Equipment residual service life prediction method based on integrated bidirectional cyclic neural network | |
| CN110555230B (en) | Rotary machine residual life prediction method based on integrated GMDH framework | |
| CN108537259A (en) | Train control on board equipment failure modes and recognition methods based on Rough Sets Neural Networks model | |
| CN112163640A (en) | Planetary gearbox fault diagnosis method and system based on deep convolution migration learning | |
| CN116226646B (en) | Method, system, equipment and medium for predicting health state and residual life of bearing | |
| CN111241673A (en) | Health state prediction method for industrial equipment in noisy environment | |
| CN106656357B (en) | Power frequency communication channel state evaluation system and method | |
| CN114186379A (en) | Transformer state evaluation method based on echo network and deep residual error neural network | |
| CN111324110A (en) | Fermentation process fault monitoring method based on multiple shrinkage automatic encoders | |
| CN114580545A (en) | Wind turbine generator gearbox fault early warning method based on fusion model | |
| Sadoughi et al. | A deep learning approach for failure prognostics of rolling element bearings | |
| CN115587290A (en) | Aero-engine fault diagnosis method based on variational self-coding generation countermeasure network | |
| CN115345222A (en) | Fault classification method based on TimeGAN model | |
| CN114563671A (en) | High-voltage cable partial discharge diagnosis method based on CNN-LSTM-Attention neural network | |
| CN115356599B (en) | Multi-mode urban power grid fault diagnosis method and system | |
| CN112101659A (en) | Complex equipment service life prediction method based on stacking denoising autoencoder | |
| CN114235409B (en) | Rolling bearing multi-user cooperative intelligent fault diagnosis method for light weight communication | |
| CN111241629A (en) | Intelligent prediction method for performance change trend of airplane hydraulic pump based on data driving | |
| CN115618751A (en) | Steel plate mechanical property prediction method | |
| CN115733673A (en) | Data anomaly detection method based on multi-scale residual classifier | |
| CN115345192A (en) | Fault diagnosis method for rolling bearing under variable working conditions and complex noise environment |
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 |