CN115510740A - Aero-engine residual life prediction method based on deep learning - Google Patents
Aero-engine residual life prediction method based on deep learning Download PDFInfo
- Publication number
- CN115510740A CN115510740A CN202211072225.9A CN202211072225A CN115510740A CN 115510740 A CN115510740 A CN 115510740A CN 202211072225 A CN202211072225 A CN 202211072225A CN 115510740 A CN115510740 A CN 115510740A
- Authority
- CN
- China
- Prior art keywords
- data
- model
- feature
- residual
- engine
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
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
- 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
Abstract
The invention relates to a method for predicting the residual life of an aircraft engine based on deep learning, which comprises the following steps: acquiring data reflecting the full life cycle of the aero-engine, and obtaining the predicted residual life of the engine through a trained residual life prediction model; the life prediction model is constructed based on deep learning, and the training process of the prediction model comprises the following steps: s1, acquiring data reflecting the full life cycle of the aero-engine; s2, preprocessing the data; s3, selecting characteristics of the data based on a random forest model; s4, performing feature extraction on the data subjected to feature selection based on a Transformer model; and S5, training the LSTM model by using the data after feature extraction. Compared with the prior art, the method has the advantages of improving the training speed, enhancing the stability of the algorithm and the like.
Description
Technical Field
The invention relates to the field of prediction of the service life of an aircraft engine, in particular to a method for predicting the residual service life of the aircraft engine based on deep learning.
Background
The aero-engine is used as a key part for normal operation of the aircraft, and real-time monitoring of the health condition and accurate prediction of the residual life of the aero-engine are beneficial to ensuring safe use of the aircraft and making a proper maintenance replacement strategy. Practice shows that the prediction and health management technology is used as a new system state prediction and health management technology, and the application of the prediction and health management technology can realize early warning of faults, prevent catastrophic accidents and reduce maintenance cost, so that the prediction and health management technology is widely applied to the industrial fields of various countries and achieves remarkable effect.
The prediction of the RUL for aircraft engines, as the most challenging direction in PHMs, is of widespread interest to the industry researchers. With the development of sensor technology, it becomes easier to acquire performance degradation information of engine components. This drives the development of data-driven based prediction methods. The method utilizes the monitoring data to extract the internal rule of performance degradation to predict the performance degradation trend of components in a certain period in the future, and has low prediction cost and high accuracy, thereby becoming a research hotspot at present.
Meanwhile, a large amount of time series data has the problems of low unit data value, high timeliness, difficult feature extraction and the like, and the utilization value of the data is greatly limited. How to screen out the most relevant and effective characteristics from massive data becomes the difficult problem of prediction of the RUL of the aircraft engine. Therefore, selecting a proper method to perform feature selection and feature extraction on high-dimensional, multi-parameter and large-scale data has a great influence on the prediction effect of the RUL of the aircraft engine.
Chinese patent CN202210447896.2 discloses a method for predicting the remaining life of an aircraft engine based on deep learning. The method adopts a residual life prediction model consisting of a self-attention mechanism and a bidirectional long and short term memory network, selects key features in time sequence data by using the model and endows the key features with corresponding weights, then inputs the key features into the bidirectional long and short term memory network layer to mine internal relation, and finally obtains a residual life prediction result of the aircraft engine through a mapping relation formed by two fully connected layers.
In the prior art, parameters are not screened, model training time is increased due to redundant characteristics and noise data, and prediction accuracy is reduced. The feature extraction model of the self-attention mechanism in the prior art adopts the self-attention mechanism only, and has the problems of gradient disappearance and data feature distribution stability reduction along with network deepening, and the convergence speed is slow.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a method for predicting the residual life of an aircraft engine based on deep learning.
The purpose of the invention can be realized by the following technical scheme:
the invention provides a method for predicting the residual life of an aircraft engine based on deep learning, which comprises the following steps: acquiring data reflecting the full life cycle of the aero-engine, and obtaining the predicted residual life of the engine through a trained residual life prediction model;
the life prediction model is constructed based on LSTM, and the training process of the life prediction model comprises the following steps:
s1, acquiring data reflecting the full life cycle of the aircraft engine;
s2, preprocessing the data;
s3, selecting characteristics of the data based on a random forest model;
s4, performing feature extraction on the data subjected to feature selection based on a Transformer model;
and S5, training the LSTM model by using the data after feature extraction.
Preferably, the data reflecting the full life cycle of the aircraft engine comprises an engine id number, an operating cycle time, operation settings and sensor data.
Preferably, the preprocessing the data includes:
carrying out filtering and noise reduction processing on the data;
carrying out normalization processing on the data;
and eliminating data which has no influence on the life prediction.
As a preferred scheme, the step of selecting features based on the random forest model comprises the following steps:
s31, selecting corresponding data outside the bag for each decision tree, and calculating a first data error outside the bag;
s32, noise interference is added to the characteristics of all samples of the out-of-bag data at random, and a second out-of-bag data error is calculated;
s33, solving the importance of the features through the first out-of-bag data error and the second out-of-bag data error;
based on the importance of each feature, an optimal feature subset is found by adopting a recursive feature elimination method.
As a preferred scheme, the specific process of performing feature screening by using the recursive feature elimination method is as follows:
s34, sorting the features according to feature importance;
s35, determining the rejection proportion, rejecting the features with the corresponding proportion, and obtaining a new feature set;
s36, judging whether the new feature set has the set number of remaining features, if not, returning to the step S34, and if so, entering the step S37;
s37 outputs the feature set.
Preferably, the Transformer model includes a multi-head attention mechanism layer based on an attention mechanism, the multi-head attention mechanism layer calculates an attention value of each data through the attention mechanism, and performs data feature extraction based on the attention value of the data, and the attention value calculation process includes the following steps:
respectively generating three initialization matrixes Query, key and Value according to input data, wherein the matrix Query represents a vector to be subsequently inquired and calculated, the matrix Key represents a vector to be subsequently inquired and calculated, and the matrix Value represents the current actual characteristic weight;
calculating the similarity degree of the input data relative to the matrix Query and the matrix Key;
performing Softmax operation on the obtained similarity, and performing normalization processing;
multiplying the matrix Value by the Value obtained by performing the Softmax operation, and summing the products to obtain the attention Value under the current data.
As a preferred scheme, the Transformer model further comprises a residual connecting layer, and the residual connecting layer solves the problem of gradient disappearance caused by network deepening by establishing a residual network and avoids the problem of stability reduction of data feature distribution caused by network deepening;
the formula of the residual block in the residual network is expressed as:
x i+1 ＝h(x i )+F(x i )
wherein x is i Input for the ith multi-head attention layer, x i+1 For the ith multi-head attention layer output, F (x) i ) Is the residual part, h (x) i ) A convolutional neural network of 1 x 1.
As a preferred scheme, the LSTM model training specifically includes the following steps:
and inputting the data subjected to the characteristic extraction of the Transformer model and an LSTM training label into an LSTM model for model training, wherein the training label is the residual life, namely the time sequence of each engine is in a reverse order.
As a preferred scheme, the LSTM model training improves the model prediction effect by adjusting the number of memory units and the number of LSTM layers.
Preferably, the LSTM model training adopts a Dropout mechanism and a mode of dynamically adjusting a learning rate to improve the generalization ability and the training speed of the model.
Compared with the prior art, the invention has the following beneficial effects:
1) The invention adopts random forest to extract features, adopts error rate outside the bag as a measurement standard to calculate the importance of the features, and then sorts and eliminates the features in corresponding proportion according to the importance of the features, thereby effectively eliminating redundant features and noise data. The characteristics of strong resolving power are reserved, the training speed is improved, the stability of the algorithm is guaranteed, and overfitting is avoided.
2) The method adopts a transform model with a residual error layer to extract features, uses a multi-head attention mechanism layer based on an attention mechanism to extract features, extracts different features through a plurality of groups of matrixes, and then splices the extracted feature vectors to obtain final output feature vectors, and aims to extract the features of a plurality of groups of sensors at the same time so as to improve the feature extraction efficiency. And then, the problem that the gradient disappears along with the deepening of the network is solved by using the residual error layer, the problem that the stability of data feature distribution is reduced along with the deepening of the network is avoided, the convergence speed of the model is increased, and the extraction efficiency of the model features is improved.
Drawings
FIG. 1 is a flow chart of the LSTM model training of the present invention;
FIG. 2 is a schematic representation of the engine life cycle data variation of the present invention;
FIG. 3 is a flow chart of feature selection based on a random forest model according to the present invention;
FIG. 4 is a graph of feature importance ranking results for a data set of the present invention; a) is an FD001 data set characteristic importance ranking result, b) is an FD002 data set characteristic importance ranking result, c) is an FD003 data set characteristic importance ranking result, and d) is an FD004 data set characteristic importance ranking result;
FIG. 5 is a graph of data set life prediction results of the present invention; a) is the predicted result for the FD001 dataset, b) is the predicted result for the FD002 dataset, c) is the predicted result for the FD003 dataset, and d) is the predicted result for the FD004 dataset.
Detailed Description
The invention is described in detail below with reference to the figures and the specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.
The embodiment is used as one implementation mode of the prediction method of the residual life of the aircraft engine based on deep learning, and the operation process comprises the following parts:
s1, firstly, selecting a plurality of groups of sensor data reflecting the whole life cycle of the aero-engine from an aero-engine data set, wherein the sensor data comprises an engine id number, operation cycle time, 3 operation settings and 21 sensor data.
TABLE 1C introduction of MAPSS data set
S2, carrying out data preprocessing on the selected data set;
the embodiment selects data to be processed on the python platform. Firstly, because the original data contains a large amount of random noise, the noise reduction processing is carried out by using a filter function, and the window width is set to be 10 so as to improve the data smoothness.
Because the monitoring data returned by a plurality of sensors of the engine represent different physical characteristics, namely different dimensions and magnitude levels, in order to eliminate the influence of data non-specification on the prediction effect and improve the prediction precision, the original data is subjected to normalization processing.
The data normalization process limits the original data to a range of [0,1], and the specific formula is
In the formula, x i,j (t) represents the data value, max (x), monitored at time t by the jth sensor of the ith engine ：,j ) Represents the maximum value, min (x), of all samples of the jth sensor ：,j ) Represents the minimum value, x 'of all samples of the j-th sensor' i,j (t) denotes x i,j (t) normalized value.
Considering that the data dimension of the data set is too much, 3 operation settings and 21 sensor data in the data set are preliminarily analyzed in order to eliminate irrelevant data attributes, reduce the dimension, reduce the model training time and improve the prediction accuracy.
As shown in fig. 2, 7 attribute data are removed without change in the engine life cycle and without influence on RUL prediction.
And S3, carrying out feature selection on the data aiming at the RF feature selection model established on the basis of the python platform, and carrying out parameter setting on the model.
Feature selection for RF we can consider the process of ordering multiple features according to the importance of each feature. Out-of-bag error rate (Out-of-bag error) is used herein as a measure of feature importance. The out-of-bag error rate represents the effect of the feature on model accuracy with and without training. The method for calculating the importance of a certain feature x in a random forest comprises the following steps:
s31, for each decision tree, selecting corresponding Out-of-bag data (Out of bag, OOB) to calculate Out-of-bag data error, and recording as errOOB1.
S32, noise interference is added to the characteristics X of all samples of the out-of-bag data OOB randomly, and the out-of-bag data error is calculated again and recorded as errOOB2.
S33 if there are N trees in the forest, the importance of the feature x is
On the basis of obtaining the importance of each Feature through random forest, an optimal Feature subset is found by adopting a Recursive Feature Elimination (Recursive Feature Elimination) method. The specific process of RFE feature screening is as follows:
s34, sorting the features according to feature importance;
s35, determining the rejection proportion, and rejecting the features of the corresponding proportion to obtain a new feature set;
s36, judging whether a set number of characteristics are left in the new characteristic set, if not, returning to the step S34, and if so, entering the step S37;
s37 outputs a feature set.
An RF feature selection model is built by a python platform by using a sklern library, and the built RF feature selection model is in a shape of RandomForestRegessor (n _ estimators =20, max _ defects = 2). Where n _ estimators is the number of decision trees and max _ features is the number of features that are partitioned when selecting the best attribute that cannot exceed this value. The present embodiment sets the initial parameters to n _ estimators =20, max _lives =2.
As shown in fig. 4, the features of the data sets F001-FD004 are ranked in importance, and the sensor data is screened and selected according to the ranking result of importance.
S4, the specific steps of carrying out feature extraction on the data subjected to feature selection based on the Transformer model comprise:
and inputting the sensor data subjected to RF characteristic selection into a Transformer model for characteristic extraction, and taking the extracted characteristics as the input of subsequent LSTM model training.
Compared with the traditional deep learning network, the Transformer model introduces a Self-Attention mechanism. Due to the introduction of the mechanism, the Transformer model can more easily extract long-distance interdependent features in data, and can better extract global information. Therefore, in consideration of the high-dimensional, multi-parameter and large-scale data characteristics of the aircraft engine, the present embodiment adopts a Transformer model for feature extraction. The transform feature extraction process mainly calculates the relevance among data through a Self-orientation mechanism, and extracts data features according to the Attention scores of the data.
The transform model of the embodiment mainly includes a multi-head attention mechanism layer and a residual connecting layer based on the self-attention mechanism.
The multi-head attention mechanism layer is composed of a plurality of groups of Query, key and Value matrixes. Through different characteristics of multiunit Query, key, value matrix extraction, the eigenvector concatenation of will extracting again obtains ultimate output eigenvector for the characteristic of extracting multiunit sensor simultaneously for improve the characteristic extraction efficiency, its flow mainly includes:
and generating three matrixes, namely Query, key and Value, according to input data, wherein Query vectors represent vectors to be subjected to Query calculation subsequently, key vectors represent vectors to be subjected to Query calculation subsequently, and Value vectors represent current actual characteristic weights. The three matrices are obtained by multiplying the input data with a randomly initialized matrix.
For the input data, the similarity degree of the data relative to other data is calculated and is expressed by f (Q, K) i ) I =1,2, 3.., m.
(1) And performing Softmax operation on the obtained similarity, and performing normalization processing:
(2) Multiplying and adding values obtained by the Value matrix and Softmax to obtain a result, namely a self-attention mechanism, and obtaining the attention Value under the current data, namely the attention degree of the data so as to achieve the purpose of feature extraction.
The problem that gradient disappears caused by network deepening is solved by the residual connecting layer through establishing a residual network, and meanwhile the problem that stability of data feature distribution is reduced caused by the network deepening is avoided, so that the convergence speed of the model is increased, and the model feature extraction efficiency is improved.
The formula of the residual block in the residual network is expressed as:
x i+1 ＝h(x i )+F(x i )
wherein x i Inputting, x, for the ith multi-head attention layer i+1 For the ith multi-head attention layer output, F (x) i ) Is the residual part, h (x) i ) A convolutional neural network of 1 x 1.
In this embodiment, a transform model is built based on a python platform, and the model entity, such as the orientation (multiple heads =20, head_dim = 10), sets the multiple heads as the number of multi-head Attention heads, and head _ dim as the dimension of the head. In this embodiment, the parameters are set to multiheads =20, head_dim =10
S5, predicting the residual life of the aircraft engine by using the data subjected to feature extraction based on the LSTM model, specifically comprising the following steps of:
and inputting the sensor data subjected to the characteristic extraction of the Transformer model and the set LSTM training label into the LSTM model for model training. The training labels are the remaining life of the engines, i.e. the time series of each engine is reversed. The test set data is then input into a trained model for RUL prediction. The prediction effect of the model is improved by adjusting the number of memory units and the number of LSTM layers, and the generalization capability and the training speed of the model are improved by introducing a Dropout mechanism and a dynamic learning rate adjusting mode. The number of LSTM network layers is set to one layer, the number of memory cells is 20, and Dropout is set to 0.5. The effect of the model is judged by comparing the accuracy of the residual life prediction under the test set.
In this embodiment, root Mean Square Error (RMSE), mean Absolute Percentage Error (MAPE), and Mean Absolute Error (MAE) are selected as evaluation indexes to measure the accuracy of model prediction. The evaluation index is mathematically expressed as:
in the formula: n is the number of the engines,
the actual remaining life and the predicted remaining life of the ith engine are respectively. RMSE is used as an error analysis comprehensive index to reflect the prediction precision, MAPE evaluates the fluctuation degree of model prediction errors and reflects the robustness and stability of the model, and MAE can reflect the actual situation of the model prediction errors.
As shown in FIG. 5, the FD001-FD004 test sets were input into the model for RUL prediction, respectively. As can be seen from the figure, the model adopted in the embodiment has good performance under four data sets, which shows that the model has good universality. In order to further compare the performance of the model of the embodiment, the embodiment compares the model with the CNN-LSTM model, the LSTM model and the SVR model under four data sets FD001-FD 004. The parameters of the LSTM in the other models are consistent with the model parameters herein. Model performance was compared by evaluating the indices RMSE, MAPE, MAE. The RMSE value can reflect the prediction precision, and the smaller the value is, the higher the prediction precision is; MAPE (mapping adaptive image optimization) evaluates the fluctuation degree of the prediction error of the model, reflects the robustness and stability of the model, and indicates that the robustness and stability of the model are better when the numerical value is smaller; the magnitude of the MAE value can reflect the actual situation of the prediction error of the model, and the smaller the value is, the smaller the prediction error of the model is.
TABLE 2 comparison of evaluation indexes under FD001
TABLE 3 comparison of evaluation indexes under FD002
TABLE 4 comparison of evaluation indexes at FD003
TABLE 5 comparison of evaluation indexes under FD004
From tables 2 to 5, it can be seen from the transverse comparison that, in the four sets of data, the RMSE, MAPE and MAE of the model of the present embodiment are all smaller than those of the other models, so that the prediction accuracy, stability and robustness of the model of the present embodiment are superior to those of the other comparison models. And the comparison with the SVR traditional machine learning model shows that the other three deep learning models have good performance in RMSE, MAPE and MAE and have more advantages. The comparison in the deep learning model shows that the prediction performance of the LSTM model after feature extraction is further improved compared with that of the LSTM model no matter CNN-LSTM or Transformer-LSTM, especially the latter LSTM model.
From tables 2 to 5, it can be seen from the longitudinal comparison that, as the number of engine states and the number of failure modes increase in the data set, the RMSE, MAPE and MAE performances of each model decrease from FD001 to FD004, but the model adopted in the embodiment is more stable in prediction performance and smaller in variation amplitude. Therefore, the model adopted by the embodiment can be more suitable for predicting the RUL of the aircraft engine under complex conditions, and still has excellent performance under the conditions of multiple working conditions and multiple fault modes.
In consideration of the characteristics of complex working environment, multiple fault modes, high dimensionality, multiple parameters and large scale of data of the aircraft engine, the embodiment provides an aircraft engine RUL prediction method based on an RF-Transformer-LSTM model. The method is based on an RF model, a Transformer model and an LSTM model. Aiming at the characteristics of high dimensionality, large scale, multiple parameters and the like of a prediction data set, a good mapping relation from high dimension to low dimension is established through an RF model and a Transformer model, and then key features are extracted for RUL prediction. The extracted features are then input into an LSTM model, and the overall logical characteristics of the time series are reflected by the LSTM model to perform RUL prediction. Experimental results show that the method provided by the embodiment is feasible and effective, and has better prediction accuracy, stability and robustness than other 3 methods.
The foregoing detailed description of the preferred embodiments of the invention has been presented. It should be understood that numerous modifications and variations could be devised by those skilled in the art in light of the present teachings without departing from the inventive concepts. Therefore, the technical solutions available to those skilled in the art through logic analysis, reasoning and limited experiments based on the prior art according to the concept of the present invention should be within the scope of protection defined by the claims.
Claims (10)
1. A method for predicting the residual life of an aircraft engine based on deep learning is characterized by comprising the following steps: acquiring data reflecting the full life cycle of the aircraft engine, and obtaining the predicted residual life of the engine through a trained residual life prediction model;
the life prediction model is constructed based on LSTM, and the training process of the prediction model comprises the following steps:
s1, acquiring data reflecting the full life cycle of the aircraft engine;
s2, preprocessing the data;
s3, selecting characteristics of the data based on a random forest model;
s4, performing feature extraction on the data subjected to feature selection based on a Transformer model;
and S5, training the LSTM model by using the data after feature extraction.
2. The method of claim 1, wherein the data reflecting the full life cycle of the aircraft engine comprises an engine id number, an operating cycle time, an operating setting, and sensor data of various engine parts.
3. The deep learning-based prediction method for the remaining life of an aircraft engine according to claim 1, wherein the preprocessing of the data comprises:
carrying out filtering and noise reduction processing on the data;
carrying out normalization processing on the data;
and eliminating data which has no influence on the life prediction.
4. The method for predicting the residual life of the aero-engine based on the deep learning as claimed in claim 1, wherein the step of selecting the features based on the random forest model comprises the following steps:
s31, selecting corresponding data outside the bag for each decision tree, and calculating a first data error outside the bag;
s32, noise interference is added to the characteristics of all samples of the out-of-bag data at random, and a second out-of-bag data error is calculated;
s33, solving the importance of the features through the first out-of-bag data error and the second out-of-bag data error;
based on the importance of each feature, an optimal feature subset is found by adopting a recursive feature elimination method.
5. The method for predicting the residual life of the aircraft engine based on the deep learning as claimed in claim 4, wherein the specific process of performing the feature screening by using the recursive feature elimination method comprises the following steps:
s34, sorting the features according to feature importance;
s35, determining the rejection proportion, rejecting the features with the corresponding proportion, and obtaining a new feature set;
s36, judging whether the new feature set has the set number of remaining features, if not, returning to the step S34, and if so, entering the step S37;
s37 outputs the feature set.
6. The method for predicting the remaining life of the aircraft engine based on deep learning of claim 1, wherein the Transformer model comprises a multi-head attention mechanism layer based on a self-attention mechanism, the multi-head attention mechanism layer calculates an attention value of each data through the self-attention mechanism, and performs data feature extraction based on the attention value of each data, and the attention value calculation process comprises the following steps:
respectively generating three initialization matrixes Query, key and Value according to input data, wherein the matrix Query represents a vector to be subsequently inquired and calculated, the matrix Key represents a vector to be subsequently inquired and calculated, and the matrix Value represents the current actual characteristic weight;
calculating the similarity degree of the input data relative to the matrix Query and the matrix Key;
performing Softmax operation on the obtained similarity, and performing normalization processing;
the matrix Value and the Value obtained by performing the Softmax operation are multiplied, and the products are summed to obtain the attention Value under the current data.
7. The method for predicting the remaining life of the aero-engine based on the deep learning as claimed in claim 6, wherein the transform model further comprises a residual connecting layer, the residual connecting layer solves the problem of gradient disappearance caused by network deepening by establishing a residual network and avoids the problem of stability reduction of data feature distribution caused by network deepening;
the formula of the residual block in the residual network is expressed as:
x i+1 ＝h(x i )+F(x i )
wherein x is i Input for the ith multi-head attention layer, x i+1 Output for the ith multi-head attention layer, F (x) i ) Is the residual part, h (x) i ) A convolutional neural network of 1 x 1.
8. The deep learning-based prediction method for the residual life of the aircraft engine as claimed in claim 1, wherein the LSTM model training specifically comprises the following steps:
and inputting the data subjected to the characteristic extraction of the Transformer model and an LSTM training label into an LSTM model for model training, wherein the training label is the residual life, namely the time sequence of each engine is in a reverse order.
9. The method as claimed in claim 8, wherein the LSTM model training is used for improving the model prediction effect by adjusting the number of memory units and the number of LSTM layers.
10. The method as claimed in claim 8, wherein the LSTM model training employs Dropout mechanism and dynamically adjusting learning rate to improve the generalization capability and training speed of the model.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211072225.9A CN115510740A (en) | 2022-09-02 | 2022-09-02 | Aero-engine residual life prediction method based on deep learning |
| PCT/CN2022/135013 WO2024045377A1 (en) | 2022-09-02 | 2022-11-29 | Deep-learning-based method for predicting residual life of aero-engine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211072225.9A CN115510740A (en) | 2022-09-02 | 2022-09-02 | Aero-engine residual life prediction method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115510740A true CN115510740A (en) | 2022-12-23 |
Family
ID=84501788
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211072225.9A Pending CN115510740A (en) | 2022-09-02 | 2022-09-02 | Aero-engine residual life prediction method based on deep learning |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN115510740A (en) |
| WO (1) | WO2024045377A1 (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117494071A (en) * | 2023-12-29 | 2024-02-02 | 深圳市科沃电气技术有限公司 | Life prediction method based on motor rotation speed monitoring and related device |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| 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 |
| CN112712209B (en) * | 2020-12-31 | 2021-11-05 | 润联智慧科技(西安)有限公司 | Reservoir warehousing flow prediction method and device, computer equipment and storage medium |
| CN113139278A (en) * | 2021-04-08 | 2021-07-20 | 上海工程技术大学 | Deep learning-based bearing residual life prediction method and system |
| CN114297918A (en) * | 2021-12-22 | 2022-04-08 | 大连理工大学 | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning |
-
2022
- 2022-09-02 CN CN202211072225.9A patent/CN115510740A/en active Pending
- 2022-11-29 WO PCT/CN2022/135013 patent/WO2024045377A1/en unknown
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117494071A (en) * | 2023-12-29 | 2024-02-02 | 深圳市科沃电气技术有限公司 | Life prediction method based on motor rotation speed monitoring and related device |
| CN117494071B (en) * | 2023-12-29 | 2024-04-16 | 深圳市科沃电气技术有限公司 | Life prediction method based on motor rotation speed monitoring and related device |
Also Published As
| Publication number | Publication date |
|---|---|
| WO2024045377A1 (en) | 2024-03-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105117602B (en) | A kind of metering device running status method for early warning | |
| CN113887616B (en) | Real-time abnormality detection method for EPG connection number | |
| CN110245745B (en) | Equipment residual service life prediction method based on integrated bidirectional cyclic neural network | |
| CN111722046A (en) | Transformer fault diagnosis method based on deep forest model | |
| CN111768000A (en) | Industrial process data modeling method for online adaptive fine-tuning deep learning | |
| Liao et al. | Electricity theft detection using Euclidean and graph convolutional neural networks | |
| CN112836604A (en) | Rolling bearing fault diagnosis and classification method, system and equipment based on VMD-SSAE and storage medium thereof | |
| CN111652271A (en) | Nonlinear feature selection method based on neural network | |
| CN110879377A (en) | Metering device fault tracing method based on deep belief network | |
| CN114297918A (en) | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning | |
| CN113869721A (en) | Substation equipment health state classification method and apparatus | |
| CN111273125A (en) | RST-CNN-based power cable channel fault diagnosis method | |
| CN111860775A (en) | Ship fault real-time diagnosis method based on CNN and RNN fusion | |
| CN110705859A (en) | PCA-self-organizing neural network-based method for evaluating running state of medium and low voltage distribution network | |
| CN115510740A (en) | Aero-engine residual life prediction method based on deep learning | |
| CN114819315A (en) | Bearing degradation trend prediction method based on multi-parameter fusion health factor and time convolution neural network | |
| CN115375026A (en) | Method for predicting service life of aircraft engine in multiple fault modes | |
| CN113177357B (en) | Transient stability assessment method for power system | |
| CN113157732A (en) | Underground scraper fault diagnosis method based on PSO-BP neural network | |
| CN116384224A (en) | Aero-engine life prediction method based on conditional parameter dynamic convolutional neural network | |
| CN115879018A (en) | Cabin equipment state perception and fault diagnosis method based on K-means algorithm and BP neural network | |
| CN115979660A (en) | Filter fault diagnosis method for internal combustion engine based on MPA optimization | |
| CN113935413A (en) | Distribution network wave recording file waveform identification method based on convolutional neural network | |
| CN114926702A (en) | Small sample image classification method based on depth attention measurement | |
| CN113112166A (en) | Equipment state variable selection method and equipment based on gray fuzzy hierarchical analysis |
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 |