CN110378052A - It is looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network - Google Patents
It is looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network Download PDFInfo
- Publication number
- CN110378052A CN110378052A CN201910676192.0A CN201910676192A CN110378052A CN 110378052 A CN110378052 A CN 110378052A CN 201910676192 A CN201910676192 A CN 201910676192A CN 110378052 A CN110378052 A CN 110378052A
- Authority
- CN
- China
- Prior art keywords
- operating condition
- lstm
- model
- group
- data
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/10—Geometric CAD
- G06F30/17—Mechanical parametric or variational design
-
- 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
-
- 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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/04—Manufacturing
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Abstract
It is looked to the future the equipment method for predicting residual useful life of operating condition the invention discloses a kind of based on Recognition with Recurrent Neural Network, it is the following steps are included: S1: obtaining multiple sensor monitoring data collection of reflection target device working condition, monitoring data including equipment from initial time to failure, and data are pre-processed；S2: modeling sample is obtained according to monitoring data using sliding time window method, and preset time window size is chosen according to experiment；S3: establishing depth LSTM model according to modeling sample, carries out preliminary predicting residual useful life to target device using the depth LSTM model after training；S4: according to tentative prediction result and the following floor data, establishing the multi input end model for the operating condition that looks to the future, and carries out predicting residual useful life to target device.This method is capable of the remaining life of accurately pre- measurement equipment, in the situation known to the following operating condition, it can be considered that influence of the following operating condition to remaining life, improves prediction accuracy.
Description
Technical field
The present invention relates to the method for predicting residual useful life fields of data-driven, are based on LSTM (Long more particularly to one kind
Short-Term Memory, i.e. shot and long term memory) Recognition with Recurrent Neural Network it can be considered that the equipment remaining life of the following operating condition is pre-
Survey method.
Background technique
Method for predicting residual useful life can be divided into following four classes, i.e., method based on physical model, based on statistical model
Method, artificial intelligence approach and mixed method.Accurate failure physical model is often difficult to set up for some complex products,
Artificial intelligence approach can predict remaining life based on history degraded data.It was mentioned in recent years there are many artificial intelligence approach
Out and achieve preferable prediction result.With the development of sensor technology, there is a large amount of degraded data that can obtain, while right
The processing and feature extraction of these data need a large amount of manual operation.Therefore, important side is commonly used as one kind of artificial intelligence
Method, deep learning can merge multivariate data, extract high dimensional feature, be a kind of strong and efficient predicting residual useful life side
Method.When there is a large amount of history degraded datas that can obtain, deep learning method is often better than traditional method for predicting residual useful life.
Have a large amount of depth learning technology and is used to prediction remaining life.However, work before only focuses on processing historical data, and
The factor that the following operating condition this pair of of predicting residual useful life has a major impact is not accounted for.By the operating condition that looks to the future, can obtain
More accurate predicting residual useful life for policymaker as a result, can provide information to adjustment future work plan, to extend simultaneously
Equipment life.
In order to combine historical sensor data and the following work information, needs to obtain the current health status of equipment, permitted
System health status is described by establishing health indicator in more researchs.In general, it is an one-dimensional change that health indicator, which is assumed to be,
Amount.However for there are many complex device of failure mode and degradation mechanism, one-dimensional health indicator can not be described comprehensively
The current health status of equipment needs to establish multidimensional health indicator to combine with the following work information.
Correlation when LSTM model can excavate long in list entries is the effective ways of Series Modeling prediction, still
The input of common LSTM model is the matrix of a fixed size, can not be by the different historical sensor data of dimension and the following work
Condition data input simultaneously.Still the two is not handled respectively in conjunction with the method to predict remaining life at present.
Summary of the invention
It is an object of the invention to propose a kind of equipment method for predicting residual useful life of operating condition that can look to the future.It needs thus
Model appropriate is established, historical sensor data and the following floor data are combined, predict remaining life jointly.
Method proposed by the present invention using multi input end model by with different dimensions historical sensor data and future
Floor data is effectively bonded together, and predicts remaining life jointly.Model can portray the following operating condition to the shadow of remaining life
It rings, can both obtain higher prediction accuracy, the work plan of equipment can also be adjusted, according to the prediction result of model to prolong
The service life of long equipment.
Specifically, provided by the invention a kind of to be looked to the future the equipment predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network
Method comprising following steps:
S1: the monitoring data slave initial time to failure of reflection target device working condition are obtained, and to monitoring data
It is pre-processed, wherein the monitoring data include the historical sensor data that multiple sensor measurements obtain；
S2: modeling sample is obtained according to monitoring data using sliding time window method, and preset according to experiment selection
Time window size；
S3: the modeling sample obtained in the S2 is divided into training group, validation group and test by building depth LSTM model
Group, and the depth LSTM model is trained using training group, the ginseng in depth LSTM model is updated using Adam algorithm
Number uses mean square error as loss function when training, stops training when the prediction error of validation group stops reducing, calculates
The prediction error of the depth LSTM model in test group, and target device is carried out using the depth LSTM model after training
Preliminary predicting residual useful life；And
S4: the result of the preliminary predicting residual useful life according to obtained in the S3 calculates the length of the following operating condition input data
Degree, establishes the multi input end model for the operating condition that looks to the future, and multi input end model includes first group of LSTM network and second group
LSTM network, by first group of LSTM network and second group of LSTM network respectively to historical sensor data and the target device
The following operating condition input data carry out feature extraction, use the depth LSTM model after being trained in the S3 as first group
The pre-training model of LSTM network, last LSTM layers of first group of LSTM network described in accurate adjustment, then will be from first group of LSTM
The characterization of multidimensional health indicator data and the following operating condition input data obtained from second group of LSTM network that network obtains
Data combine, and carry out predicting residual useful life to target device jointly.
It may be preferred that the monitoring data include the described of reflection target device health status, working environment or load
Multiple sensor historic monitoring data.
Further, include: to the pretreatment of the monitoring data
S11, the data of each sensor are screened, removes the sensor monitoring not changed under same operating
Data；
S12, data normalization are returned according to the following formula according to maximum value of each sensor under different operating conditions and minimum value
One changes,
Wherein, x(i, m)Indicate m-th of sensor in the original value at i-th of time point,Indicate x(i, m)After normalization
Value,WithIndicate maximum value and minimum value of m-th of sensor under c kind operating condition；Each x(i, m)All use phase
It answersWithIt is normalized；
S13, each time point remaining life label of setting are chosen most according to the variation tendency of the monitoring data of each sensor
Big remaining life label value, and the remaining life label value at each time point is set using piecewise linear function；And
If the operating condition of S14, target device experience can cluster, the operating condition after cluster is encoded with one-hot, is obtained anti-
Reflect the monitoring data of the code expression of operating condition.
It may be preferred that all historical sensor monitoring data within the scope of time window are common for each sample
Constitute the input matrix of model；For each input matrix, the corresponding remaining life of the last one time step is as the input
The target output value of matrix；By the way that time window is slided into last from first time step of historical sensor monitoring data
A time step can obtain the sample comprising input matrix and target output value for establishing model, wherein the time window
Size is determined according to the result of preliminary experiment.
It may be preferred that the depth LSTM model includes one layer of shielding layer and LSTM layers of several layers, the shielding layer is used
In the time step for skipping filling, described LSTM layers is used for the feature extraction of input data, according to the modeling sample to the mould
Type is trained, to obtain the depth LSTM model.
It may be preferred that being remained first with the depth LSTM model to target device before establishing multi input end model
The remaining service life is estimated, and prediction mean absolute error and discreet value the sum of of the depth LSTM model in test group are calculated,
And using calculated result as the length of the following operating condition input data of multi input end model.
It may be preferred that being obtained by the first group of LSTM network and second group of LSTM network of multi input end model more
After being combined for health indicator and the following operating condition input data characterization data, it is input in full articulamentum, by full articulamentum
It returns and calculates, obtain final predicting residual useful life as a result, the wherein full articulamentum and first group of LSTM network and the
Two groups of LSTM networks are trained jointly.
It may be preferred that the training of multi input end model based on the depth LSTM model, constructs more first
Input terminal model structure, then by the parameter assignment other than the output layer of depth LSTM model obtained in the S3 to described more
Parameter in first group of LSTM network of input terminal model, by the parameter other than last LSTM layers of first group of LSTM network
It remains unchanged, updates remaining all parameter of multi input end model simultaneously using Adam optimization algorithm.
The learning rate of optimization algorithm of the present invention is smaller, and the amplitude for guaranteeing that parameter updates is sufficiently small with tamper-proof pre-training net
The extracted feature of network.
Compared with prior art, the present invention has following innovative point:
(1) present invention proposes a kind of how defeated for remaining life is better anticipated in the situation known to the following operating condition
Enter to hold model, which can be effectively combined one with the following floor data for the historical sensor data with different dimensions
It rises, predicts remaining life jointly.Model can portray influence of the following operating condition to remaining life, and it is quasi- can both to have obtained higher prediction
Exactness can also adjust the work plan of equipment, according to the prediction result of model to extend the service life of equipment.
(2) present invention proposes that the first operating condition that do not look to the future estimates the remaining life of target device, accurate in conjunction with what is estimated
The length of the following operating condition list entries is calculated in degree, to avoid the loss or redundancy of information.
(3) present invention proposes a kind of accurate adjustment strategy of multi input end model, using depth LSTM model as multi input end mould
The pre-training model of type a part, then accurate adjustment is carried out, to reduce the training difficulty of model, improve the validity of model.
(4) present invention proposes the current health state that equipment is characterized using multidimensional health indicator.
Detailed description of the invention
Fig. 1 be look to the future the present invention is based on Recognition with Recurrent Neural Network operating condition equipment method for predicting residual useful life the step of stream
Cheng Tu；
Fig. 2 is the monitor value of a sensor life cycle management of an equipment in the embodiment of the present invention；
Fig. 3 is the internal structure of LSTM neuron of the present invention；
Fig. 4 is depth LSTM model structure of the present invention；
Fig. 5 is model training flow chart of the present invention；
Fig. 6 is FD002 test set prediction result of the present invention；
Fig. 7 is FD004 test set prediction result of the present invention；
Fig. 8 is multi input end of the present invention model structure；And
Fig. 9 is the following operating condition list entries length of the invention and multi input end model fine adjusting method schematic diagram.
Specific embodiment
The embodiment of the present invention is described below in detail, the example of embodiment is shown in the accompanying drawings, wherein identical from beginning to end
Or similar label indicates same or similar element or element with the same or similar functions.It is retouched below with reference to attached drawing
The embodiment stated is exemplary, it is intended to is used to explain the present invention, and is not considered as limiting the invention.
The present invention provides a kind of to be looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network.Below
In conjunction with attached method flow shown in FIG. 1, with three subsets comprising various working of C-MAPSS fanjet emulation data set
For, hereinafter referred to as challenge round data, FD002 and FD004 do modeling method of the invention and further retouch
It states；
Three data sets all respectively include a training set and a test set.Training set includes multiple similar engines
Life cycle management monitoring data, including 21 sensing data sequences and 3 operating condition sensing data sequences.Every engine by
Different initial healths is started to work, and starts to degenerate after a period of time, until thrashing.Sensor number in test set
According to sometime termination of the sequence before thrashing, target seeks to predict these engines in the last one of record data
The remaining life at time point.In addition, each engine residual life of the test set of two groups of data of FD002 and FD004 be it is known,
But challenge round data is unknown.Therefore, the prediction result of test set and other existing methods knots in FD002 and FD004
Fruit comparison can prove the validity for the model that S3 step of the present invention obtains.And training set has the data of life cycle management, it is each
The following work information of a time step is known, therefore this partial data can be used to verify holistic approach of the present invention not
Carry out the validity of the predicting residual useful life under operating condition known case.
Step 1: initial data is pre-processed.That is, obtain reflection target device working condition slave initial time to
Monitoring data of failure, and pre-processing to monitoring data, wherein monitoring data include going through of obtaining of multiple sensor measurements
History sensing data.
Feature selecting is carried out first, is removed in sensor monitoring data at any time without the data sequence of significant change trend
Column.In this example, in 21 sensor monitoring data sequences, 7 data sequences remain unchanged always under same operating, this
A little data do not have any effect to predicting residual useful life, it should be rejected.
The amplitude of each sensor monitoring data is different, and amplitude of the same sensor under different operating conditions is also different.In order to
So that sensor monitoring data is all had identical amplitude in same range, and reduce the influence of different operating conditions, according to the following formula to biography
Sensor monitoring data are normalized under different operating conditions, in scaling to [0,1] range.
Wherein, x(i, m)Indicate m-th of sensor in the original value at i-th of time point,Indicate x(i, m)After normalization
Value,WithIndicate maximum value and minimum value of m-th of sensor under c kind operating condition.Each x(i, m)All use phase
It answersWithIt is normalized.
The remaining life label at each time point is set.In fact, work of the equipment in early stage be it is normal, for a period of time with
Just start to degenerate afterwards.For example, in this example, the 11st sensor monitor value after the normalization of FD002 First equipment is for example attached
Shown in Fig. 2.It can be seen from the figure that measurement value sensor, in early stage no apparent variation tendency, degeneration just shows after starting
The trend increased out.Noticing other sensors measured value also has same phenomenon, it can be assumed that in early stage remaining life
Label be a constant, the label at each time point is set using a piecewise linear function.In this example, early stage in the remaining longevity
Life is set as 125.
For the operating condition that can be clustered, all kinds of operating conditions are indicated with one-hot vector after cluster.In this example, data are utilized
Three operating condition sensor monitoring data are concentrated to draw scatter plot in three-dimensional figure, it can be seen that it is attached that all the points all concentrate on six points
Closely, then operating condition readily can be divided into six classes, and indicates all kinds of operating conditions with the one-hot vector that length is 6.
Further, in this example, it is missed using a kind of score function used in 08 meeting of PHM ' match and root mean square
Poor (RMSE) carrys out the error of computation model prediction, and specific calculation expression is as follows:
Step 2: modeling sample is obtained according to monitoring data using sliding time window method, and is chosen in advance according to experiment
If time window size.For each sample, all historical sensor datas within the scope of time window collectively form mould
The input matrix of type.For each input matrix, mark of the corresponding remaining life of the last one time step as the input matrix
Label are target output value.By the way that time window is slided into the last one time from first time step of historical sensor data
Step, can obtain the sample comprising input matrix and target output value, for establishing model.The size of time window is a weight
The parameter wanted.Different window sizes is chosen, cross-validation experiments result is folded by 10- and determines suitable value.In this example,
The experimental result of different windows size is as shown in table 1.According to experimental result, selecting window size is 80, i.e., for every equipment
Historical sensor data, the data from the 1st time point to the 80th time point are as first input sample, from the 2nd
The data at time point to the 81st time point make second input sample, so divide, to the last time point
Data are partitioned into a sample.
The experimental result of 1 different time window size of table
Step 3: depth LSTM model, i.e. multilayer LSTM model are established according to modeling sample.Depth LSTM model is constructed,
The modeling sample obtained in S2 is divided into training group, validation group and test group, and depth LSTM model is carried out using training group
Training updates the parameter in depth LSTM model using Adam algorithm, use mean square error as it is trained when loss function, when
Stop training when the prediction error of validation group stops reducing, calculate the prediction error of the depth LSTM model in test group, and makes
Preliminary predicting residual useful life is carried out to target device with the depth LSTM model after training.Each unit in model, i.e. LSTM mind
Structure through member is as shown in Figure 3.Based on LSTM neuron, depth LSTM model as shown in Figure 4 is constructed, and is obtained with step 2
Modeling sample carry out model training and selection.
Specifically, model training and the process of selection are as shown in Figure 5.Firstly, data set is included by pretreatment
Sensor monitoring data and the time series for the one-hot vector for representing different operating conditions after normalization.Then, when passing through sliding
Between windowhood method obtained the input samples of 2 dimensions, the size of window is determined by the experiment of different value.
Depth LSTM model is necessary for establishing multi input end model.When known to the following operating condition, each sample
The following operating condition input data length is the remaining life discreet value of depth LSTM model and its mean absolute error in test group
The sum of.If it is known that the following operating condition sequence length it is insufficient, then generate operating condition at random behind to supply.It should be noted that
The following operating condition list entries length of each sample is not identical, so zero padding makes each sample behind shorter sequence
The dimension of the following operating condition is identical, so as to the training of model.
Then, the sample obtained by training set is further divided into training group, validation group and test according to the ratio of 8:1:1
Group.The sample of training group is used for training pattern, and the sample of validation group is instructed for determining that exercise wheel number prevents over-fitting when in maximum
Then deconditioning, test group are used for the performance of evaluation model and preference pattern when the error of validation group no longer reduces before practicing wheel number
Hyper parameter.
Then, it is based on ready sample, first establishes depth LSTM network, its structural parameters are according to it in test group sample
Performance in sheet determines.Then, it is based on first established model, establishes the multi input end model for the operating condition that looks to the future.Mould
Type is trained according to the mean square error of every small batch sample, and in this example, every batch of includes that 1024 samples are proper.
Finally, the sample of test group, which is input in model, predicts its remaining life, and calculate prediction error.In this example,
The prediction result difference of each target device is as shown in Figure 6 and Figure 7 in FD002 and FD004 test set, the equipment of each test set according to
Label sorts from small to large.It can be found that predicted value relatively true value, especially when equipment soon fails.The present invention
Method it is as shown in table 2 in the prediction result of this step and the comparison of existing other methods:
Remaining life tentatively estimates accuracy in 2 present invention of table and existing other methods prediction accuracy compares
Step 4: according to tentative prediction result and the following floor data, establishing the multi input end model for the operating condition that looks to the future,
Target device remaining life is predicted accordingly.The result of the preliminary predicting residual useful life according to obtained in S3 calculates the following operating condition input
The length of data, establishes the multi input end model of operating condition of looking to the future, and multi input end model includes first group of LSTM network and the
Two groups of LSTM networks, by first group of LSTM network and second group of LSTM network respectively to historical sensor data and target device
The following operating condition input data carry out feature extraction, use the depth LSTM model after being trained in S3 as first group of LSTM
The pre-training model of network, last LSTM layers of first group of LSTM network of accurate adjustment, then it is more by being obtained from first group of LSTM network
The characterize data of dimension health indicator data and the following operating condition input data obtained from second group of LSTM network combines, common right
Target device carries out predicting residual useful life.
Construct multi input end model as shown in Figure 8, and the target device remaining life discreet value obtained according to step 3
And its following operating condition list entries length is calculated in statistical average absolute error, obtains the following operating condition according to future work plan
List entries forms input sample with historical sensor data, carries out model training and selection.The training of model and selection it is total
The same step 3 of body process, as shown in Figure 5.
The fine adjusting method for generating the following operating condition list entries and training multi input end model is as shown in Figure 9.Specifically, first
Step, do not look to the future operating condition, by the remaining life of each target device of depth LSTM model pre-estimating, and calculates each target
The corresponding following operating condition list entries length of equipment；Second step, according to the following operating condition list entries length and future work plan
Obtain the following floor data of each target device；Third step is loaded into depth LSTM model obtained in trained S3
The part LSTM, the good parameter of entire infrastructure and pre-training other than the full articulamentum of the last layer in the model of multi input end as calculating
The pre-training model of multidimensional health indicator；4th step constructs entire multi input terminal nerve network, freezes last in pre-training model
Layer other than one layer LSTM layers, and the other parts of training network.
The prediction result of 3 two kinds of models of table compares
The prediction result pair of the depth LSTM model for the operating condition that do not look to the future and the multi input end model for the operating condition that looks to the future
Such as shown in table 3 above.As can be seen that the prediction result of multi input end model is more quasi- when known to the following operating condition, in addition,
Prediction result is also shown that using the fine adjusting method shown in Fig. 9 for generating the following operating condition list entries and training multi input end model
The prediction accuracy of model can be improved, therefore, method of the invention is more effective.
Finally, it should be noted that embodiment described above is only used to illustrate the technical scheme of the present invention, rather than it is limited
System；Although the present invention is described in detail referring to the foregoing embodiments, those skilled in the art should understand that: its
It can still modify to technical solution documented by previous embodiment, or part of or all technical features are carried out
Equivalent replacement；And these modifications or substitutions, technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution
Range.
Claims (8)
1. a kind of looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network, which is characterized in that it is wrapped
Include following steps:
S1: the monitoring data slave initial time to failure of reflection target device working condition are obtained, and monitoring data are carried out
Pretreatment, wherein the monitoring data include the historical sensor data that multiple sensor measurements obtain；
S2: modeling sample is obtained according to monitoring data using sliding time window method, and the preset time is chosen according to experiment
Window size；
S3: the modeling sample obtained in the S2 is divided into training group, validation group and test group by building depth LSTM model, and
The depth LSTM model is trained using training group, the parameter in depth LSTM model is updated using Adam algorithm, makes
It uses mean square error as loss function when training, stops training when the prediction error of validation group stops reducing, calculate test
The prediction error of the depth LSTM model in group, and target device is carried out tentatively using the depth LSTM model after training
Predicting residual useful life；And
S4: the result of the preliminary predicting residual useful life according to obtained in the S3 calculates the length of the following operating condition input data, builds
The multi input end model of the vertical operating condition that looks to the future, multi input end model include first group of LSTM network and second group of LSTM net
Network, by first group of LSTM network and second group of LSTM network respectively to the future of historical sensor data and the target device
Operating condition input data carries out feature extraction, uses the depth LSTM model after being trained in the S3 as first group of LSTM net
The pre-training model of network, last LSTM layers of first group of LSTM network described in accurate adjustment, then will be obtained from first group of LSTM network
The characterize data phase of the multidimensional health indicator data and the following operating condition input data obtained from second group of LSTM network that obtain
In conjunction with jointly to target device progress predicting residual useful life.
2. according to claim 1 looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network,
It is characterized in that, the monitoring data include reflecting the multiple sensing of target device health status, working environment or load
Device Historical Monitoring data.
3. according to claim 1 looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network,
It is characterized in that, including: to the pretreatment of the monitoring data
S11, the data of each sensor are screened, removes the sensor monitoring number not changed under same operating
According to；
S12, data normalization carry out normalizing according to maximum value of each sensor under different operating conditions and minimum value according to the following formula
Change,
Wherein, x(i, m)Indicate m-th of sensor in the original value at i-th of time point,Indicate x(i, m)Value after normalization,WithIndicate maximum value and minimum value of m-th of sensor under c kind operating condition；Each x(i, m)All with correspondingWithIt is normalized；
S13, each time point remaining life label of setting choose maximum surplus according to the variation tendency of the monitoring data of each sensor
Remaining service life label value, and the remaining life label value at each time point is set using piecewise linear function；And
If the operating condition of S14, target device experience can cluster, the operating condition after cluster is encoded with one-hot, obtains reflection work
The monitoring data of the code expression of condition.
4. being looked to the future the equipment predicting residual useful life of operating condition according to claim 1 or described in 3 based on Recognition with Recurrent Neural Network
Method, which is characterized in that for each sample, all historical sensor monitoring data within the scope of time window are collectively formed
The input matrix of model；For each input matrix, the corresponding remaining life of the last one time step is as the input matrix
Target output value；When by the way that time window being slided into the last one from first time step of historical sensor monitoring data
Spacer step can obtain the sample comprising input matrix and target output value for establishing model, wherein the time window size
It is determined according to the result of preliminary experiment.
5. being looked to the future the equipment predicting residual useful life of operating condition according to claim 1 or described in 3 based on Recognition with Recurrent Neural Network
Method, which is characterized in that the depth LSTM model includes one layer of shielding layer and LSTM layers of several layers, and the shielding layer is for jumping
Cross the time step of filling, the described LSTM layer feature extraction for input data, according to the modeling sample to the model into
Row training, to obtain the depth LSTM model.
6. being looked to the future the equipment predicting residual useful life of operating condition according to claim 1 or described in 3 based on Recognition with Recurrent Neural Network
Method, which is characterized in that before establishing multi input end model, first with the depth LSTM model to the remaining longevity of target device
Life is estimated, and calculates prediction mean absolute error and discreet value the sum of of the depth LSTM model in test group, and will
Length of the calculated result as the following operating condition input data of multi input end model.
7. being looked to the future the equipment predicting residual useful life of operating condition according to claim 1 or described in 3 based on Recognition with Recurrent Neural Network
Method, which is characterized in that the multidimensional obtained by first group of LSTM network of multi input end model and second group of LSTM network
After health indicator and the following operating condition input data characterization data combine, it is input in full articulamentum, by returning for full articulamentum
Return calculating, obtains final predicting residual useful life as a result, the wherein full articulamentum and first group of LSTM network and second
Group LSTM network is common training.
8. according to claim 7 looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network,
It is characterized in that, the training of multi input end model based on the depth LSTM model, constructs multi input end mould first
Then type structure gives the parameter assignment other than the output layer of depth LSTM model obtained in the S3 to multi input end mould
Parameter in first group of LSTM network of type keeps the parameter other than last LSTM layers of first group of LSTM network not
Become, updates remaining all parameter of multi input end model simultaneously using Adam optimization algorithm.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910676192.0A CN110378052B (en) | 2019-07-25 | 2019-07-25 | Equipment residual life prediction method considering future working conditions based on cyclic neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910676192.0A CN110378052B (en) | 2019-07-25 | 2019-07-25 | Equipment residual life prediction method considering future working conditions based on cyclic neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110378052A true CN110378052A (en) | 2019-10-25 |
| CN110378052B CN110378052B (en) | 2020-11-06 |
Family
ID=68255819
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910676192.0A Active CN110378052B (en) | 2019-07-25 | 2019-07-25 | Equipment residual life prediction method considering future working conditions based on cyclic neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110378052B (en) |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111241673A (en) * | 2020-01-07 | 2020-06-05 | 北京航空航天大学 | Health state prediction method for industrial equipment in noisy environment |
| CN111340282A (en) * | 2020-02-21 | 2020-06-26 | 山东大学 | DA-TCN-based method and system for estimating residual service life of equipment |
| CN111460728A (en) * | 2020-03-09 | 2020-07-28 | 华南理工大学 | Method and device for predicting residual life of industrial equipment, storage medium and equipment |
| CN111475921A (en) * | 2020-03-13 | 2020-07-31 | 重庆邮电大学 | Tool residual life prediction method based on edge calculation and L STM network |
| CN112115636A (en) * | 2020-08-18 | 2020-12-22 | 国网山东省电力公司经济技术研究院 | Method and system for predicting insulation aging life of power cable in advance |
| CN112150443A (en) * | 2020-09-27 | 2020-12-29 | 中南大学 | Train-mounted air conditioner residual life prediction method based on air quality data map |
| CN112232557A (en) * | 2020-09-30 | 2021-01-15 | 中国铁道科学研究院集团有限公司通信信号研究所 | Switch machine health degree short-term prediction method based on long-term and short-term memory network |
| CN112465259A (en) * | 2020-12-12 | 2021-03-09 | 武汉七星电气有限公司 | Switch fault prediction method based on deep neural network |
| CN112488295A (en) * | 2020-11-23 | 2021-03-12 | 江苏科技大学 | Method for optimizing storage life prediction of LSTM network relay by cross validation algorithm |
| CN113158348A (en) * | 2021-05-21 | 2021-07-23 | 上海交通大学 | Aircraft engine residual life prediction method based on deep learning coupling modeling |
| CN113505532A (en) * | 2021-07-06 | 2021-10-15 | 新智数字科技有限公司 | Method and device for predicting residual life of equipment, computer equipment and medium |
| CN113536982A (en) * | 2021-06-28 | 2021-10-22 | 长江大学 | Intelligent prediction method, system, equipment and medium for residual life of main reducer |
| CN113570160A (en) * | 2021-08-27 | 2021-10-29 | 三一石油智能装备有限公司 | Service life prediction method and device for sand pump of sand mixing truck |
| CN113567858A (en) * | 2021-07-30 | 2021-10-29 | 北京航空航天大学 | Control moment gyro residual life prediction system |
| CN113570138A (en) * | 2021-07-28 | 2021-10-29 | 朗坤智慧科技股份有限公司 | Method and device for predicting residual service life of equipment of time convolution network |
| CN113779882A (en) * | 2021-09-10 | 2021-12-10 | 中国石油大学(北京) | Method, device, equipment and storage medium for predicting residual service life of equipment |
| CN114035098A (en) * | 2021-12-14 | 2022-02-11 | 北京航空航天大学 | Lithium battery health state prediction method integrating future working condition information and historical state information |
| CN114112030A (en) * | 2021-12-06 | 2022-03-01 | 江西省纳米技术研究院 | Photoelectric detector performance testing device based on artificial intelligence |
| CN115310240A (en) * | 2022-10-11 | 2022-11-08 | 烟台杰瑞石油装备技术有限公司 | Oil-gas fracturing pump equipment service life prediction method and device and nonvolatile storage medium |
| CN115327382A (en) * | 2022-06-28 | 2022-11-11 | 广州汽车集团股份有限公司 | Method and device for generating battery service life prediction model and vehicle |
| CN115824261A (en) * | 2023-01-15 | 2023-03-21 | 北京理工大学 | Control moment gyroscope fault detection method and device and related storage medium |
| WO2024050782A1 (en) * | 2022-09-08 | 2024-03-14 | Siemens Aktiengesellschaft | Method and apparatus for remaining useful life estimation and computer-readable storage medium |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102592171A (en) * | 2011-12-30 | 2012-07-18 | 南京邮电大学 | Method and device for predicting cognitive network performance based on BP (Back Propagation) neural network |
| CN105758661A (en) * | 2016-05-05 | 2016-07-13 | 华电国际电力股份有限公司技术服务中心 | System and method for evaluating service life of boiler heating surface |
| CN106951695A (en) * | 2017-03-09 | 2017-07-14 | 杭州安脉盛智能技术有限公司 | Plant equipment remaining life computational methods and system under multi-state |
| US20190057307A1 (en) * | 2016-10-11 | 2019-02-21 | Hitachi, Ltd. | Deep long short term memory network for estimation of remaining useful life of the components |
| CN109782192A (en) * | 2019-03-08 | 2019-05-21 | 安徽理工大学 | Lithium ion battery residual life prediction technique under different discharge-rates |
| CN109883699A (en) * | 2018-12-20 | 2019-06-14 | 上海理工大学 | A kind of rolling bearing method for predicting residual useful life based on long memory network in short-term |
| CN109991542A (en) * | 2019-03-27 | 2019-07-09 | 东北大学 | Lithium ion battery residual life prediction technique based on WDE optimization LSTM network |
-
2019
- 2019-07-25 CN CN201910676192.0A patent/CN110378052B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102592171A (en) * | 2011-12-30 | 2012-07-18 | 南京邮电大学 | Method and device for predicting cognitive network performance based on BP (Back Propagation) neural network |
| CN105758661A (en) * | 2016-05-05 | 2016-07-13 | 华电国际电力股份有限公司技术服务中心 | System and method for evaluating service life of boiler heating surface |
| US20190057307A1 (en) * | 2016-10-11 | 2019-02-21 | Hitachi, Ltd. | Deep long short term memory network for estimation of remaining useful life of the components |
| CN106951695A (en) * | 2017-03-09 | 2017-07-14 | 杭州安脉盛智能技术有限公司 | Plant equipment remaining life computational methods and system under multi-state |
| CN109883699A (en) * | 2018-12-20 | 2019-06-14 | 上海理工大学 | A kind of rolling bearing method for predicting residual useful life based on long memory network in short-term |
| CN109782192A (en) * | 2019-03-08 | 2019-05-21 | 安徽理工大学 | Lithium ion battery residual life prediction technique under different discharge-rates |
| CN109991542A (en) * | 2019-03-27 | 2019-07-09 | 东北大学 | Lithium ion battery residual life prediction technique based on WDE optimization LSTM network |
Non-Patent Citations (1)
| Title |
|---|
| 彭宝华等: "基于退化与寿命数据融合的产品剩余寿命预测", 《系统工程与电子技术》 * |
Cited By (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111241673A (en) * | 2020-01-07 | 2020-06-05 | 北京航空航天大学 | Health state prediction method for industrial equipment in noisy environment |
| CN111241673B (en) * | 2020-01-07 | 2021-10-22 | 北京航空航天大学 | Health state prediction method for industrial equipment in noisy environment |
| CN111340282B (en) * | 2020-02-21 | 2021-04-23 | 山东大学 | DA-TCN-based method and system for estimating residual service life of equipment |
| CN111340282A (en) * | 2020-02-21 | 2020-06-26 | 山东大学 | DA-TCN-based method and system for estimating residual service life of equipment |
| CN111460728A (en) * | 2020-03-09 | 2020-07-28 | 华南理工大学 | Method and device for predicting residual life of industrial equipment, storage medium and equipment |
| CN111460728B (en) * | 2020-03-09 | 2022-08-12 | 华南理工大学 | Method and device for predicting residual life of industrial equipment, storage medium and equipment |
| CN111475921A (en) * | 2020-03-13 | 2020-07-31 | 重庆邮电大学 | Tool residual life prediction method based on edge calculation and L STM network |
| CN112115636A (en) * | 2020-08-18 | 2020-12-22 | 国网山东省电力公司经济技术研究院 | Method and system for predicting insulation aging life of power cable in advance |
| CN112150443B (en) * | 2020-09-27 | 2022-07-12 | 中南大学 | Train-mounted air conditioner residual life prediction method based on air quality data map |
| CN112150443A (en) * | 2020-09-27 | 2020-12-29 | 中南大学 | Train-mounted air conditioner residual life prediction method based on air quality data map |
| CN112232557A (en) * | 2020-09-30 | 2021-01-15 | 中国铁道科学研究院集团有限公司通信信号研究所 | Switch machine health degree short-term prediction method based on long-term and short-term memory network |
| CN112232557B (en) * | 2020-09-30 | 2024-02-09 | 中国铁道科学研究院集团有限公司通信信号研究所 | Short-term prediction method for health degree of switch machine based on long-short-term memory network |
| CN112488295A (en) * | 2020-11-23 | 2021-03-12 | 江苏科技大学 | Method for optimizing storage life prediction of LSTM network relay by cross validation algorithm |
| CN112465259A (en) * | 2020-12-12 | 2021-03-09 | 武汉七星电气有限公司 | Switch fault prediction method based on deep neural network |
| CN112465259B (en) * | 2020-12-12 | 2023-01-31 | 武汉七星电气有限公司 | Switch fault prediction method based on deep neural network |
| CN113158348A (en) * | 2021-05-21 | 2021-07-23 | 上海交通大学 | Aircraft engine residual life prediction method based on deep learning coupling modeling |
| CN113158348B (en) * | 2021-05-21 | 2023-10-03 | 上海交通大学 | Aircraft engine residual life prediction method based on deep learning coupling modeling |
| CN113536982B (en) * | 2021-06-28 | 2023-02-24 | 长江大学 | Intelligent prediction method, system, equipment and medium for residual life of main reducer |
| CN113536982A (en) * | 2021-06-28 | 2021-10-22 | 长江大学 | Intelligent prediction method, system, equipment and medium for residual life of main reducer |
| CN113505532B (en) * | 2021-07-06 | 2023-12-29 | 新奥新智科技有限公司 | Equipment remaining life prediction method, device, computer equipment and medium |
| CN113505532A (en) * | 2021-07-06 | 2021-10-15 | 新智数字科技有限公司 | Method and device for predicting residual life of equipment, computer equipment and medium |
| CN113570138A (en) * | 2021-07-28 | 2021-10-29 | 朗坤智慧科技股份有限公司 | Method and device for predicting residual service life of equipment of time convolution network |
| CN113570138B (en) * | 2021-07-28 | 2022-04-26 | 朗坤智慧科技股份有限公司 | Method and device for predicting residual service life of equipment of time convolution network |
| CN113567858A (en) * | 2021-07-30 | 2021-10-29 | 北京航空航天大学 | Control moment gyro residual life prediction system |
| CN113570160A (en) * | 2021-08-27 | 2021-10-29 | 三一石油智能装备有限公司 | Service life prediction method and device for sand pump of sand mixing truck |
| CN113779882A (en) * | 2021-09-10 | 2021-12-10 | 中国石油大学(北京) | Method, device, equipment and storage medium for predicting residual service life of equipment |
| CN114112030A (en) * | 2021-12-06 | 2022-03-01 | 江西省纳米技术研究院 | Photoelectric detector performance testing device based on artificial intelligence |
| CN114035098A (en) * | 2021-12-14 | 2022-02-11 | 北京航空航天大学 | Lithium battery health state prediction method integrating future working condition information and historical state information |
| CN115327382A (en) * | 2022-06-28 | 2022-11-11 | 广州汽车集团股份有限公司 | Method and device for generating battery service life prediction model and vehicle |
| CN115327382B (en) * | 2022-06-28 | 2024-02-23 | 广州汽车集团股份有限公司 | Method and device for generating battery service life prediction model and vehicle |
| WO2024050782A1 (en) * | 2022-09-08 | 2024-03-14 | Siemens Aktiengesellschaft | Method and apparatus for remaining useful life estimation and computer-readable storage medium |
| CN115310240A (en) * | 2022-10-11 | 2022-11-08 | 烟台杰瑞石油装备技术有限公司 | Oil-gas fracturing pump equipment service life prediction method and device and nonvolatile storage medium |
| CN115824261A (en) * | 2023-01-15 | 2023-03-21 | 北京理工大学 | Control moment gyroscope fault detection method and device and related storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110378052B (en) | 2020-11-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110378052A (en) | It is looked to the future the equipment method for predicting residual useful life of operating condition based on Recognition with Recurrent Neural Network | |
| CN110147450B (en) | Knowledge complementing method and device for knowledge graph | |
| CN111222549B (en) | Unmanned aerial vehicle fault prediction method based on deep neural network | |
| US9934470B2 (en) | Production equipment including machine learning system and assembly and test unit | |
| CN108564326A (en) | Prediction technique and device, computer-readable medium, the logistics system of order | |
| CN110298663A (en) | Based on the wide fraudulent trading detection method learnt deeply of sequence | |
| CN111768000A (en) | Industrial process data modeling method for online adaptive fine-tuning deep learning | |
| CN107274011A (en) | The equipment state recognition methods of comprehensive Markov model and probability net | |
| CN108734266A (en) | Compression method and device, terminal, the storage medium of deep neural network model | |
| CN108573116A (en) | A kind of aero-engine transition state thrust algorithm for estimating based on long memory network in short-term | |
| CN112185104A (en) | Traffic big data restoration method based on countermeasure autoencoder | |
| CN108805195A (en) | A kind of motor group method for diagnosing faults based on two-value deep-neural-network | |
| CN111226234A (en) | Method, apparatus and computer program for creating a deep neural network | |
| CN112912902A (en) | Artificial neural network and method for training artificial neural network by using epigenetic neurogenesis | |
| CN112116002A (en) | Determination method, verification method and device of detection model | |
| CN112434743A (en) | Fault identification method based on GIL metal particle partial discharge time domain waveform image | |
| CN115982141A (en) | Characteristic optimization method for time series data prediction | |
| CN109491709B (en) | Software code controllability comprehensive evaluation method based on AHP and neural network | |
| CN108073978A (en) | A kind of constructive method of the ultra-deep learning model of artificial intelligence | |
| CN108073979A (en) | A kind of ultra-deep study of importing artificial intelligence knows method for distinguishing for image | |
| CN108073985A (en) | A kind of importing ultra-deep study method for voice recognition of artificial intelligence | |
| CN111079348A (en) | Method and device for detecting slowly-varying signal | |
| CN116403054A (en) | Image optimization classification method based on brain-like network model | |
| CN115902620A (en) | Motor fault enhancement diagnosis method and device based on loop generation countermeasure network | |
| CN114720129A (en) | Rolling bearing residual life prediction method and system based on bidirectional GRU |
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 |