CN112613226A - Feature enhancement method for residual life prediction - Google Patents
Feature enhancement method for residual life prediction Download PDFInfo
- Publication number
- CN112613226A CN112613226A CN202011432003.4A CN202011432003A CN112613226A CN 112613226 A CN112613226 A CN 112613226A CN 202011432003 A CN202011432003 A CN 202011432003A CN 112613226 A CN112613226 A CN 112613226A
- Authority
- CN
- China
- Prior art keywords
- data
- test
- sample
- training
- rul
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/15—Vehicle, aircraft or watercraft design
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Geometry (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Computer Hardware Design (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Automation & Control Theory (AREA)
- Aviation & Aerospace Engineering (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Pure & Applied Mathematics (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
A characteristic enhancement method for residual life prediction belongs to the field of aircraft engine fault prediction and health management. First, the engine sensor data is normalized and the remaining life values of the training set data and the test set data are calculated. Secondly, the training data set and the testing data set are subjected to dimensionality reduction, and a training data set sample and a testing data set sample are extracted in a sliding window mode. And thirdly, performing data characteristic enhancement on the training set sample and the test and sample, and determining the sample after the data characteristic enhancement. And finally, setting a long-term and short-term memory structure of the deep learning neural network, adding a sample for training, and predicting test data by using the neural network model. The deep neural network model is established based on a data-driven form, is irrelevant to the actual engine model, can be migrated to engines of different models for use by training different data sets, and has certain universality.
Description
Technical Field
The invention provides a novel Tensorflow deep learning framework-based data enhancement method for predicting the residual life of an aircraft engine system or a component, and belongs to the field of aircraft engine fault Prediction and Health Management (PHM).
Background
The subject of the invention is the turbofan engine degradation simulation dataset disclosed by NASA. The data set was used for engine degradation simulation using C-MAPSS, four different sets were simulated under different combinations of conditions and failure modes, and multiple sensor channel data were recorded to characterize the failure evolution.
The data set is divided into four sub-data sets, each of which is divided into a training data set, a test data set, and a true system or component remaining life value (RUL). For each subdata set, the training data set and the test data set are composed of a plurality of parameters, the parameters include engine number, time series, 3 engine operation setting parameters and 21 sensor parameter data, and the parameters are time series parameters.
The engine operates normally at the beginning of each time series and then a fault occurs at some point in the time series. In the training set, the fault continues to increase until the system fails. In the test set, the time series ends at a time prior to the system failure. What we need to do is to predict the number of operating cycles remaining before the test set fails, i.e., the number of operating cycles that the engine will continue to operate after the last operating cycle, based on the complete training data set given. The RUL file for each subdata set also provides the actual Remaining Useful Life (RUL) value for the engine to which the test data corresponds.
For a time series turbofan engine degradation simulation data set, most of the current methods adopt a Recurrent Neural Network (RNN) to train and predict the data set, and a long-short term memory network (LSTM) is also an RNN network model which proves to be very successful and is widely applied to the aspects of speech recognition, machine translation and the like. LSTM solves the problem of gradient disappearance or gradient explosion that typical RNNs can exhibit, using input, forgetting and output gates to control the flow of information, and the inherent sequential nature of sensor data for this turbofan engine also means that it is suitable for use with LSTM for RUL prediction.
Disclosure of Invention
The invention provides a feature extraction method based on data increase, aiming at the problem that the features and information of an original data set cannot be fully learned and mined by adding data directly using standard normalization into an LSTM network for training.
The technical scheme of the invention is as follows:
a novel feature enhancement method for remaining life prediction comprises the following steps:
s1, normalization processing is carried out on engine sensor data, and the method comprises the following steps:
s1.1 reading public NASA turbofan engine training set data train _ FD001, test set data test _ FD001, and remaining life data set RUL _ FD001 (the public data set includes four divided sub-data sets, each of which is divided into a training data set, a test data set, and a remaining life data set, here the first sub-data set is taken as an example);
s1.2, the sensor parameter data of the training set data train _ FD001 and the test set data test _ FD001 are time series data, the sensor data are normalized by adopting a maximum and minimum normalization method, and normalized training set data train _01_ ru and test set data test _01_ ru are obtained, wherein the normalization formula is as follows:
wherein x is sensor parameter data before normalization, and x' isNormalized sensor parameter data, xminIs the minimum value of the sensor parameter, xmaxIs the maximum value of the sensor parameter;
s2, calculating the residual life RUL values of the training set data and the test set data according to the training set data train _01_ ru, the test set data test _01_ ru and the residual life data set RUL _ FD001 obtained in the step S1, and the method comprises the following steps:
s2.1 finding the maximum engine cycle number cycle corresponding to each engine number id of the training data setmax_train(the first and second columns of parameters in the training dataset data are engine number and engine cycle number, respectively), cycle is usedmax_trainMinus the number of engine cycles cyclet_trainObtaining the corresponding residual service life RUL under the engine cycle numbert_trainA value;
s2.2 finding the maximum engine cycle number cycle corresponding to each engine number id of the test data settest(the first and second columns of parameters in the test data set are the engine number and the engine cycle number respectively), and the engine cycle number is added with the RUL value in the remaining life data set (the data set only contains the remaining life RUL value corresponding to each engine number), so as to obtain the maximum cycle number cycle of the enginemax_testUsing cyclemax_testMinus the number of engine cycles cyclet_testObtaining the corresponding residual service life RUL under the engine cycle numbert_testA value;
s2.3 setting the maximum residual Life RULmaxWith a value of 125, recalculating the residual lifetime RUL of the training set datat_trainResidual lifetime RUL of value and test set datat_testThe value, the calculation formula is as follows:
in the formula, RULtResidual Life RUL for training set datat_trainRemaining lifetime RUL of value or test set datat_testA value;
s3, performing dimension reduction processing on the normalized training data set and the test data set obtained in the step S2, namely removing the sensor parameter data with parameter values unchanged from the data set in the sensor parameter data, and taking the rest sensor parameter data as original sample data;
s4, obtaining the residual life RUL of the training set data according to the training data set train _01_ ru and the test data set test _01_ ru which are subjected to dimension reduction processing in S3 and the training set data obtained in S2t_trainResidual lifetime RUL of value and test set datat_testValues, extracting training data set samples and testing data set samples in a sliding window manner, comprising the steps of:
s4.1 finding the maximum engine cycle number cycle corresponding to each engine number id of the training data setmax_trainAnd taking out the minimum cycle thereofmax_train_minSetting the sliding window size winSize close to and not exceeding the minimum cyclemax_train_min；
S4.2, taking the size of the sliding window winSize as a row, taking the dimension reduction sensor parameter data FeaSIze as a column, extracting samples of a training data set train _01_ ru and a test data set test _01_ ru, taking the samples as a 2-dimensional matrix with the winSize as a row and the FeaSIze as a column, and taking the output of a corresponding network as the residual life of the engine (sampling the RUL corresponding to the last time sequence of the samplet) At this time, the input sample of the training data set is train x _ New, the input sample of the testing data set is testX _ New, the output sample of the training data set is train y, and the output sample of the testing data set is testY;
s5, performing data feature enhancement on the training set sample train X _ New and the test and sample testX _ New acquired in the step S4 in a manual feature extraction mode, adding four feature values of ridge regression weight, sample mean value, sample maximum value and sample minimum value to each sample, expanding sample dimensionality, and determining a sample after data feature enhancement, wherein the method comprises the following steps:
s5.1, respectively calculating a ridge regression weight value coef, a mean value mean, a maximum value max and a minimum value min of each sensor parameter for the training set sample train X _ New and the test and sample testX _ New obtained in the step S4;
s5.2 for the four characteristic values obtained, winSize is added to enhanceObtaining a final data feature enhanced sample from the samples, wherein the sample is a 2-dimensional matrix with winSize +4 as rows and FeaSeze as columns, the final training set sample is defined as train X _ reNew, the final testing set sample is defined as testX _ New, and the corresponding output engine residual life RUL istThe values are unchanged, i.e. trainY and testY are unchanged;
s6, setting a deep learning neural network long-short term memory (LSTM) structure, adding a sample for training, and predicting test data by using a trained neural network model, wherein the method comprises the following steps:
s6.1, establishing a Sequential model structure which is a linear stack of a network layer;
s6.2, adding an LSTM layer and an NN layer, wherein the LSTM layer structure is generally set to be a 3-4-layer neural network structure, the NN layer is generally set to be a 1-2-layer neural network structure, a Dropout function is used for preventing overfitting of neural network model learning, and a Reynolds function (RELU) is adopted as an activation function;
s6.3, selecting an optimizer of the network, a target loss function loss, a batch size and a training time epochs, and adding an input sample train X _ reNew and an output sample train Y into the network for training to obtain a prediction model;
s6.4, predicting the testset testX _ reenw by the trained prediction model, carrying out model prediction comparison by combining testY to obtain prediction results rmse and score,
wherein rmse is the root mean square error and the calculation mode is
In the above formula, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrue(i.e., the RUL _ FD001 data RUL value).
score is a scoring function calculated by
Likewise, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrueThe error between.
The invention has the beneficial effects that: compared with a deep LSTM network which directly performs normalized training on data, the root mean square error rmse and score of a target function of the deep LSTM network are reduced to a certain degree, the prediction accuracy of the network model is improved to a certain degree, and in addition, the results of multiple tests also show that the prediction stability of the network model is improved to a certain degree. The method establishes the deep neural network model based on a data-driven form, and is irrelevant to the actual engine model, so that the model can be conveniently transferred to engines of different models for use by training different data sets, and the method has certain universality.
Drawings
FIG. 1 is a flow diagram of a CMPASS dataset system or component remaining life prediction based on deep learning;
FIG. 2 is a flow chart of data processing in three different ways (wherein the third method is the method provided by the present invention);
FIG. 3 is a diagram of a neural network model architecture in accordance with the present disclosure.
Detailed Description
The new feature enhancement method for remaining life prediction according to the present invention will be further described with reference to the accompanying drawings.
The present invention relies on the turbine fan engine degradation simulation dataset disclosed in the background by NASA, and is based on a new feature enhancement method flow for residual life prediction as shown in FIG. 1.
Fig. 2 is a flow chart of three methods for processing data and training a network respectively, the first two methods in fig. 2 are conventional methods, the first method is to determine samples in a sliding window manner after only performing normalization processing on original data to perform network training for predicting RUL, the second method is similar to the first method except that feature values are obtained from the original data to perform subsequent network training and prediction operations instead of the original data, the method is the third method, and the main steps include:
s1, determining a data set as a degradation simulation data set of a turbofan engine disclosed by NASA, wherein the data set comprises four sub data sets, each sub data set comprises a training data set, a testing data set and an RUL data set, and the purpose is to train a deep learning neural network according to the training data set for testing the data set and compare the difference between a prediction result and actual RUL data;
s2, carrying out normalization processing on the data of the engine sensor, and comprising the following steps:
s2.1 reading public NASA turbofan engine training set data train _ FD001, test set data test _ FD001, and remaining life data set RUL _ FD001 (the public data set includes four divided sub-data sets, each of which is divided into a training data set, a test data set, and a remaining life data set, here the first sub-data set is taken as an example);
s2.2, the sensor parameter data of the training set data train _ FD001 and the test set data test _ FD001 which are read are time series data, normalization processing is carried out on the sensor data by adopting a maximum and minimum normalization method, normalized training set data train _01_ ru and normalized test set data test _01_ ru are obtained, and the normalization formula is as follows:
wherein x is sensor parameter data before normalization, x' is sensor parameter data after normalization, and xminIs the minimum value of the sensor parameter, xmaxIs the maximum value of the sensor parameter;
s3, calculating the residual life RUL values of the training set data and the test set data according to the training set data train _01_ ru, the test set data test _01_ ru and the residual life data set RUL _ FD001 obtained in the step S2, and the method comprises the following steps:
s3.1 finding the maximum engine cycle number cycle corresponding to each engine number id of the training data setmax_train(the first and second columns of parameters in the training dataset data are engine number and engine cycle number, respectively), cycle is usedmax_trainMinus the number of engine cycles cyclet_trainObtaining the corresponding residual service life RUL under the engine cycle numbert_trainA value;
s3.2 finding the maximum engine cycle number cycle corresponding to each engine number id of the test data settest(the first and second columns of parameters in the test data set are the engine number and the engine cycle number respectively), and the engine cycle number is added with the RUL value in the remaining life data set (the data set only contains the remaining life RUL value corresponding to each engine number), so as to obtain the maximum cycle number cycle of the enginemax_testUsing cyclemax_testMinus the number of engine cycles cyclet_testObtaining the corresponding residual service life RUL under the engine cycle numbert_testA value;
s3.3 setting the maximum residual Life RULmaxWith a value of 125, recalculating the residual lifetime RUL of the training set datat_trainResidual lifetime RUL of value and test set datat_testThe value, the calculation formula is as follows:
in the formula, RULtResidual Life RUL for training set datat_trainRemaining lifetime RUL of value or test set datat_testA value;
s4, performing dimensionality reduction on the normalized training data set and the test data set obtained in the S3, namely removing sensor parameter data with parameter values unchanged from the data set (the data set can be observed to find that 7 sensor parameter values in the sensor parameter data keep constant values in a given time sequence, and if the sensor parameters are S1, S2 and so on, S1, S5, S6, S10, S16, S18 and S19 parameter values are constant respectively, and removing the data to achieve a dimensionality reduction effect), and taking the rest sensor parameter data as original sample data;
s5, obtaining the residual life RUL of the training set data according to the training data set train _01_ ru and the test data set test _01_ ru which are subjected to dimension reduction processing in S4 and the training set data obtained in S3t_trainResidual lifetime RUL of value and test set datat_testValues, extracting training data set samples and testing data set samples in a sliding window manner, comprising the steps of:
s5.1 finding the maximum engine cycle number cycle corresponding to each engine number id of the training data setmax_trainAnd taking out the minimum cycle thereofmax_train_min(the values corresponding to the four subdata sets are found to be 31,21,38 and 19 respectively in the experiment), and the size of the sliding window winSize is set to be close to and not exceed the minimum value cyclemax_train_min(setting the sliding window sizes winSize to 30,20,38,19, respectively);
s5.2, taking the size of the sliding window winSize as a row, taking the dimension reduction sensor parameter data FeaSIze as a column, extracting samples of a training data set train _01_ ru and a test data set test _01_ ru, taking the samples as a 2-dimensional matrix with the winSize as a row and the FeaSIze as a column, and taking the output of a corresponding network as the residual life of the engine (sampling the RUL corresponding to the last time sequence of the samplet) At this time, the input sample of the training data set is train x _ New, the input sample of the testing data set is testX _ New, the output sample of the training data set is train y, and the output sample of the testing data set is testY;
s6, performing data feature enhancement on the training set sample train X _ New and the test and sample testX _ New acquired in the step S5 in a manual feature extraction mode, adding four feature values of ridge regression weight, sample mean value, sample maximum value and sample minimum value to each sample, expanding sample dimensionality, and determining a sample after data feature enhancement, wherein the method comprises the following steps:
s6.1, respectively calculating a ridge regression weight value coef, a mean value mean, a maximum value max and a minimum value min of each sensor parameter for the training set sample train X _ New and the test and sample testX _ New obtained in S5, wherein the ridge regression weight value coef can be obtained by a RidgeCV () function in a linear _ model of a sketch kit, and the mean value, the maximum value and the minimum value of the sensor parameters can be obtained by a corresponding function mean (), max () and min () of the numpy kit;
s6.2, adding the obtained four characteristic values into a sample in a winSize enhancing mode to obtain a final data characteristic enhanced sample, wherein the sample is a 2-dimensional matrix with winSize +4 as rows and FeaSize as columns, the final training set sample is defined as train X _ reNew, the final test set sample is testX _ New, and the corresponding output engine residual life RUL istThe values are unchanged, i.e. trainY and testY are unchanged;
s7, setting a deep learning neural network long-short term memory (LSTM) structure, adding a sample for training, and predicting test data by using a trained neural network model, wherein the method comprises the following steps:
s7.1, establishing a Sequential model structure which is a linear stack of a network layer;
s7.2, adding an LSTM layer and an NN layer, wherein the LSTM layer structure is generally set to be a 3-4-layer neural network structure, the NN layer is generally set to be a 1-2-layer neural network structure, a Dropout function is used for preventing overfitting of neural network model learning, a Reynolds function (RELU) is adopted as an activation function, the network structure is shown in figure 3, namely, the 4-layer LSTM network structure is adopted, the number of memory units is respectively 64,64,8 and 8, a two-layer NN structure is adopted, and the number of neuron nodes is respectively 16 and 1;
the expression of the RELU function is as follows
f(x)＝max(0,x)
Where x is the functional input, for a given input to a neuron, the above equation becomes
f(x)＝max(0,x)＝max(0,WTx+b)
In the formula WTIs the weight of the neuron, and b is the bias of the neuron;
s7.3, selecting an optimizer of the network as rmsprop, a target loss function loss as mean square error mse, a batch size of 512 and training times epochs of 100, and adding an input sample train X _ reNew and an output sample train Y into the network for training to obtain a prediction model;
s7.4, predicting the testset testX _ reenw by the trained prediction model, carrying out model prediction comparison by combining testY to obtain prediction results rmse and score,
wherein rmse is the root mean square error and the calculation mode is
In the above formula, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrue(i.e., the RUL _ FD001 data RUL value).
score is a scoring function calculated by
Likewise, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrueThe error between.
Tables 1,2 and 3 are the results of the simulation of the sub-data set 1 of the turbofan engine degradation simulation data set disclosed by NASA according to the three different data processing methods in fig. 2. The prediction results of the LSTM network model obtained by the common direct normalization processing method have the RMSE mean value mean of 13.965, std of 0.520, score of 341.269 and std of 107.875, the second model prediction result of LSTM network training by using characteristic values instead of original data has the RMSE mean value mean of 13.904, std of 0.527, score of 297.373 and std of 36.840, the results obtained by the data enhancement method are that the RMSE mean value mean is 12.479, std of 0.348, score of 279.212 and std of 21.889, and it can be seen that the root mean square errors obtained by the first method and the second method are not greatly different, the scoring function obtained by the second method is effectively reduced compared with the first method, the standard deviation is also reduced, and the stability is better. The data enhancement method in the present invention can obtain better prediction performance, further decrease the root mean square error and the scoring function, and show better stability compared with the former two methods.
TABLE 1 prediction result table of LSTM trained based on common normalization method of subdata set 1
| | Score | ||
| 1 | 13.192 | 327.142 | |
| 2 | 13.263 | 234.108 | |
| 3 | 13.814 | 261.377 | |
| 4 | 14.474 | 604.091 | |
| 5 | 13.955 | 428.477 | |
| 6 | 14.52 | 312.76 | |
| 7 | 14.477 | 307.913 | |
| 8 | 13.82 | 251.531 | |
| 9 | 14.665 | 274.354 | |
| 10 | 13.469 | 344.024 | |
| mean | 13.965 | 341.269 | |
| std | 0.520 | 107.875 |
TABLE 2 prediction results Table for LSTM trained based on the feature data of subdata set 1
| | Score | ||
| 1 | 13.634 | 293.456 | |
| 2 | 13.519 | 263.007 | |
| 3 | 13.788 | 296.983 | |
| 4 | 14.646 | 274.509 | |
| 5 | 13.497 | 271.426 | |
| 6 | 13.566 | 287.183 | |
| 7 | 14.483 | 379.878 | |
| 8 | 14.694 | 354.434 | |
| 9 | 14.143 | 284.636 | |
| 10 | 13.073 | 268.22 | |
| mean | 13.904 | 297.373 | |
| std | 0.527 | 36.840 |
TABLE 3 prediction results Table for the present method based on subdata set 1
Table 4, table 5 and table 6 the results obtained from the simulation of the subset of turbofan engine degradation simulation data sets 2,3,4 disclosed by NASA according to the method herein.
TABLE 4 prediction results Table for the present methods based on subdata set 2
| | Score | ||
| 1 | 17.443 | 1409.015 | |
| 2 | 16.709 | 1308.895 | |
| 3 | 17.076 | 1342.558 | |
| 4 | 17.384 | 1511.902 | |
| 5 | 17.656 | 1809.671 | |
| 6 | 17.081 | 1292.520 | |
| 7 | 17.041 | 1383.842 | |
| 8 | 17.239 | 1323.760 | |
| 9 | 17.143 | 1260.521 | |
| 10 | 17.314 | 1587.507 | |
| mean | 17.209 | 1423.019 | |
| std | 0.248 | 160.927 |
TABLE 5 prediction results Table for the present methods based on subdata set 3
| | Score | ||
| 1 | 12.175 | 253.251 | |
| 2 | 11.935 | 255.896 | |
| 3 | 12.332 | 259.733 | |
| 4 | 12.492 | 271.425 | |
| 5 | 12.687 | 343.225 | |
| 6 | 12.332 | 255.477 | |
| 7 | 12.254 | 267.069 | |
| 8 | 12.213 | 231.005 | |
| 9 | 11.837 | 234.359 | |
| 10 | 12.414 | 259.910 | |
| mean | 12.267 | 263.135 | |
| std | 0.237 | 29.308 |
Table 6 prediction results table for the present method based on subdata set 4
In conclusion, the novel feature enhancement method for predicting the residual life is feasible, the prediction performance can be obviously improved, the prediction capability is stable, the root mean square error and the scoring function of the feature enhancement method are effectively reduced, and the results are reasonable and are not accidental results generated by network fluctuation.
The above-mentioned embodiments only express the embodiments of the present invention, but not should be understood as the limitation of the scope of the invention patent, it should be noted that, for those skilled in the art, many variations and modifications can be made without departing from the concept of the present invention, and these all fall into the protection scope of the present invention.
Claims (4)
1. A feature enhancement method for remaining life prediction, comprising the steps of:
s1, normalization processing is carried out on engine sensor data, and the method comprises the following steps:
s1.1, reading training set data, test set data and residual life data set data of the turbofan engine;
s1.2, the sensor parameter data of the training set data train _ FD001 and the test set data test _ FD001 which are read are time series data, and the sensor data are normalized by adopting a maximum and minimum normalization method to obtain normalized training set data and normalized test set data;
s2, calculating the residual life RUL values of the training set data and the test set data according to the residual life data set obtained in the step S1 and the training set data and the test set data after normalization processing, and comprising the following steps of:
s2.1 according to the maximum engine cycle number cycle corresponding to each engine number of the training data setmax_trainUsing cyclemax_trainMinus the number of engine cycles cyclet_trainObtaining the corresponding residual service life RUL under the engine cycle numbert_trainA value;
s2.2 according to the test data set, the maximum engine cycle number cycle corresponding to each engine numbertestAdding the RUL value in the remaining life data set to the number of engine cycles to obtain the maximum number of engine cycles cyclemax_testUsing cyclemax_testMinus the number of engine cycles cyclet_testObtaining the corresponding residual service life RUL under the engine cycle numbert_testA value;
s2.3 setting the maximum residual Life RULmaxValue, recalculation of training set data residual Life RULt_trainResidual lifetime RUL of value and test set datat_testThe value, the calculation formula is as follows:
in the formula, RULtResidual Life RUL for training set datat_trainRemaining lifetime RUL of value or test set datat_testA value;
s3, performing dimension reduction on the normalized training data set and the test data set obtained in the step S2, removing the sensor parameter data with parameter values unchanged from the data set in the sensor parameter data, and taking the rest sensor parameter data as original sample data;
s4, according to the training data set and the test data set subjected to the dimensionality reduction processing in the step S3 and the residual life RUL of the training set data obtained in the step S2t_trainResidual lifetime RUL of value and test set datat_testValues, extracting training data set samples and testing data set samples in a sliding window manner, comprising the steps of:
s4.1 finding the maximum engine cycle number cycle corresponding to each engine number of the training data setmax_trainAnd taking out the minimum cycle thereofmax_train_minSetting the sliding window size winSize close to and not exceeding the minimum cyclemax_train_min；
S4.2, taking the size of a sliding window as a row and the parameter data of the dimension reduction sensor as a column, extracting samples of a training data set and a testing data set, wherein the output of a corresponding network is the residual life of the engine, the input sample of the training data set at the moment is train X _ New, the input sample of the testing data set is testX _ New, the output sample of the training data set is train Y, and the output sample of the testing data set is testY;
s5, performing data feature enhancement on the training set sample train X _ New and the test and sample testX _ New acquired in the step S4 by using a manual feature extraction mode, adding four feature values of ridge regression weight, sample mean value, sample maximum value and sample minimum value to each sample, expanding sample dimensionality, and determining a sample after data feature enhancement, wherein the method comprises the following steps:
s5.1, respectively calculating a ridge regression weight value coef, a mean value mean, a maximum value max and a minimum value min of each sensor parameter for the training set sample train X _ New and the test and sample testX _ New obtained in the step S4;
s5.2, adding the obtained four characteristic values into a sample in a winSize enhancement mode to obtain a final data characteristic enhanced sample; defining the final training set sample as train X _ reNew, and the final testing set sample as testX _ New, and corresponding to the residual service life RUL of the output enginetThe values are unchanged, i.e. trainY and testY are unchanged;
s6, setting a deep learning neural network long-term and short-term memory LSTM structure, adding a sample for training, and predicting test data by using a trained neural network model, wherein the method comprises the following steps:
s6.1, establishing a Sequential model structure which is a linear stack of a network layer;
s6.2, adding an LSTM layer and an NN layer, and preventing overfitting of neural network model learning by adopting an activation function;
s6.3, selecting an optimizer of the network, a target loss function loss, a batch size and a training time epochs, and adding an input sample train X _ reNew and an output sample train Y into the network for training to obtain a prediction model;
s6.4, predicting the testset testX _ reenw by the trained prediction model, and performing model prediction comparison by combining testY to obtain a prediction result: root mean square error rmse and scoring function score;
the calculation mode of the root mean square error is as follows:
in the above formula, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrueThe error between;
the calculation mode of the scoring function is as follows:
likewise, n is the number of samples, hiPredicting RUL for remaining lifepreAnd actual remaining life value RULtrueThe error between.
2. The method as claimed in claim 1, wherein the step S2.3 of setting the maximum remaining lifetime RUL is performed by using a method of predicting the remaining lifetimemaxThe value is 125.
3. A method for feature enhancement for residual life prediction according to claim 1, characterized in that in step S6.2: the LSTM layer structure is generally set to be a 3-4 layer neural network structure, and the NN layer is generally set to be a 1-2 layer neural network structure.
4. A feature enhancement method for predicting remaining life according to claim 1, wherein the activation function in step S6.2 is a reynolds function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011432003.4A CN112613226B (en) | 2020-12-10 | 2020-12-10 | Feature enhancement method for residual life prediction |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011432003.4A CN112613226B (en) | 2020-12-10 | 2020-12-10 | Feature enhancement method for residual life prediction |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112613226A true CN112613226A (en) | 2021-04-06 |
| CN112613226B CN112613226B (en) | 2022-11-18 |
Family
ID=75229510
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011432003.4A Active CN112613226B (en) | 2020-12-10 | 2020-12-10 | Feature enhancement method for residual life prediction |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112613226B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113449463A (en) * | 2021-06-09 | 2021-09-28 | 重庆锦禹云能源科技有限公司 | LSTM-DNN-based equipment life prediction method and device |
| CN113722989A (en) * | 2021-08-23 | 2021-11-30 | 南京航空航天大学 | CPS-DP model-based aircraft engine life prediction method |
| CN113722833A (en) * | 2021-09-09 | 2021-11-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on dual-channel long-short time memory network |
| CN114186337A (en) * | 2021-11-30 | 2022-03-15 | 大连理工大学 | Gas compressor rotating stall prediction method based on multi-source data fusion |
| CN114944057A (en) * | 2022-04-21 | 2022-08-26 | 中山大学 | Road network traffic flow data restoration method and system |
| WO2024020960A1 (en) * | 2022-07-28 | 2024-02-01 | 西门子股份公司 | Method for predicting remaining useful life of circuit breaker, and electronic device and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555230A (en) * | 2019-07-12 | 2019-12-10 | 北京交通大学 | rotary machine residual life prediction method based on integrated GMDH framework |
| CN110807257A (en) * | 2019-11-04 | 2020-02-18 | 中国人民解放军国防科技大学 | Method for predicting residual life of aircraft engine |
-
2020
- 2020-12-10 CN CN202011432003.4A patent/CN112613226B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555230A (en) * | 2019-07-12 | 2019-12-10 | 北京交通大学 | rotary machine residual life prediction method based on integrated GMDH framework |
| CN110807257A (en) * | 2019-11-04 | 2020-02-18 | 中国人民解放军国防科技大学 | Method for predicting residual life of aircraft engine |
Non-Patent Citations (1)
| Title |
|---|
| 王旭等: "基于CAE与LSTM的航空发动机剩余寿命预测", 《北京信息科技大学学报(自然科学版)》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113449463A (en) * | 2021-06-09 | 2021-09-28 | 重庆锦禹云能源科技有限公司 | LSTM-DNN-based equipment life prediction method and device |
| CN113722989A (en) * | 2021-08-23 | 2021-11-30 | 南京航空航天大学 | CPS-DP model-based aircraft engine life prediction method |
| CN113722989B (en) * | 2021-08-23 | 2023-04-28 | 南京航空航天大学 | CPS-DP model-based aeroengine service life prediction method |
| CN113722833A (en) * | 2021-09-09 | 2021-11-30 | 湖南工业大学 | Turbofan engine residual service life prediction method based on dual-channel long-short time memory network |
| CN114186337A (en) * | 2021-11-30 | 2022-03-15 | 大连理工大学 | Gas compressor rotating stall prediction method based on multi-source data fusion |
| WO2023097705A1 (en) * | 2021-11-30 | 2023-06-08 | 大连理工大学 | Air compressor rotation stall prediction method based on multi-source data fusion |
| CN114944057A (en) * | 2022-04-21 | 2022-08-26 | 中山大学 | Road network traffic flow data restoration method and system |
| CN114944057B (en) * | 2022-04-21 | 2023-07-25 | 中山大学 | Road network traffic flow data restoration method and system |
| WO2024020960A1 (en) * | 2022-07-28 | 2024-02-01 | 西门子股份公司 | Method for predicting remaining useful life of circuit breaker, and electronic device and storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112613226B (en) | 2022-11-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112613226B (en) | Feature enhancement method for residual life prediction | |
| CN108805185B (en) | Face recognition method and device, storage medium and computer equipment | |
| CN108399201B (en) | Web user access path prediction method based on recurrent neural network | |
| CN112149316B (en) | Aero-engine residual life prediction method based on improved CNN model | |
| CN110298264B (en) | Human body daily behavior activity recognition optimization method based on stacked noise reduction self-encoder | |
| CN106203534A (en) | A kind of cost-sensitive Software Defects Predict Methods based on Boosting | |
| CN110688288A (en) | Automatic testing method, device, equipment and storage medium based on artificial intelligence | |
| CN111460728A (en) | Method and device for predicting residual life of industrial equipment, storage medium and equipment | |
| Li et al. | Domain adaptation remaining useful life prediction method based on AdaBN-DCNN | |
| CN112149373A (en) | Complex analog circuit fault identification and estimation method and system | |
| CN110210495A (en) | The XGBoost soft-measuring modeling method extracted based on parallel LSTM self-encoding encoder behavioral characteristics | |
| CN112001110A (en) | Structural damage identification monitoring method based on vibration signal space real-time recursive graph convolutional neural network | |
| CN114297918A (en) | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning | |
| CN113268833A (en) | Migration fault diagnosis method based on deep joint distribution alignment | |
| Sadoughi et al. | A deep learning approach for failure prognostics of rolling element bearings | |
| CN111079348A (en) | Method and device for detecting slowly-varying signal | |
| CN114462459A (en) | Hydraulic machine fault diagnosis method based on 1DCNN-LSTM network model | |
| JP2019095600A (en) | Acoustic model learning device, speech recognition device, and method and program for them | |
| CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
| CN112185423B (en) | Voice emotion recognition method based on multi-head attention mechanism | |
| CN109493975B (en) | Chronic disease recurrence prediction method, device and computer equipment based on xgboost model | |
| CN116720079A (en) | Wind driven generator fault mode identification method and system based on multi-feature fusion | |
| Arifin et al. | Automatic essay scoring for Indonesian short answers using siamese Manhattan long short-term memory | |
| CN113052060B (en) | Bearing residual life prediction method and device based on data enhancement and electronic equipment | |
| CN113962431A (en) | Bus load prediction method for two-stage feature processing |
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 |