CN110737948A - method for predicting residual life of aero-engine based on deep FNN-LSTM hybrid network - Google Patents

Abstract

The invention discloses aero-engine residual life prediction methods based on a deep FNN-LSTM hybrid network, which comprise the steps of adding -order difference and second-order difference of detection signal data on the basis of original engine detection signal data to form a three-dimensional detection data structure, using the three-dimensional detection data structure as a characteristic item, generating a training target residual life RUL based on a difference accumulation method, establishing an engine residual life prediction model based on the deep FNN-LSTM hybrid network according to the characteristic item and the training target RUL, and obtaining the residual life of an engine through the FNN-LSTM prediction model by using the engine detection signal data to be predicted.

Description

method for predicting residual life of aero-engine based on deep FNN-LSTM hybrid network

Technical Field

The invention belongs to the technical field of the residual life of an aero-engine, and relates to methods for predicting the residual life of the aero-engine based on a deep FNN-LSTM hybrid network.

Technical Field

The PHM is intended to be maintained before a system/facility failure, and further to evaluate risks or predict residual life RUL in real time by evaluating various system conditions according to historical trajectory data.

generally speaking, prediction methods are mainly divided into three types, model-based, data-based and mixed models, model-based prediction refers to a method using a model derived from the th principle or probability theory, common methods include particle filtering, kalman filtering, weibull distribution, and Eyring model, etc. since the model is most likely to represent the actual degradation characteristics of the system, the method has the highest prediction accuracy theoretically, but the premise is that an accurate physical model is constructed based on prior knowledge of system degradation, whereas for a complex system in practice, prior is generally difficult to obtain, and inter-variables, variables and overall characteristics influence each other, it is difficult to construct an accurate model.

In analyzing the remaining life prediction problem, we found that data is generally detection signals recorded from a running system at time intervals, namely time series data, hi the problem of processing time series data, a recurrent neural network RNN with a memory unit is a more suitable choice than a CNN which is good at processing the image field.

Disclosure of Invention

The invention aims to provide methods for predicting the residual life of the aero-engine based on the deep FNN-LSTM hybrid network aiming at the problems in the prior art, and provide a technical scheme for health management and optional maintenance of the aero-engine.

The technical scheme is as follows: in order to achieve the purpose, the technical scheme adopted by the invention is as follows:

1, aviation engine residual life prediction method based on deep FNN-LSTM hybrid network, which comprises the following steps:

step 1) on the basis of detection signal data of a plurality of groups of aeroengines from a healthy state to a degraded state, -order difference and second-order difference of the detection signal data are added to form a three-dimensional detection data structure and the three-dimensional detection data structure is used as a characteristic item;

step 2) generating a training target residual life RUL based on a difference cumulative addition method;

step 3) according to the training set, the feature items and the training target RUL, training an engine residual life prediction model based on the deep FNN-LSTM mixed network, and predicting the test set by adopting an optimal model to obtain the engine residual life RUL;

2. the method for predicting remaining life of aircraft engine based on deep FNN-LSTM hybrid network as claimed in claim 1, wherein the specific steps of adding order difference and second order difference of the detection signal data to form a three-dimensional detection data structure based on the detection signal data of multiple groups of aircraft engines from healthy state to degraded state in step 1) are as follows:

step 1.1) detection signal data order difference and second order difference information are adopted, and then a mathematical model is used for more accurately describing the detection signal data and the nonlinear relation between order difference and second order difference of the detection signal data and the engine change process:

wherein (1)

In the above formulas (1) to (3), k is a time point, r (k) represents a change process of the engine, s_kThe value of the detection signal at k is,

to detect the th order derivative of the value with respect to time,

in practical applications, generally studies the variation process by collecting discrete values, so the difference in the above equations (2) and (3) is used to replace the differential in the continuous process.

Step 1.2) calculating order difference and second order difference of the engine detection signal data by a forward difference method;

step 1.3) in the detection signal data, expanding the corresponding signal column dimension by using -order difference and second-order difference of each signal, and then forming a three-dimensional detection data structure of < detection signal data, -order difference of detection signal data, second-order difference of detection signal data > as a feature item for each detection signal.

3. The method for predicting remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network as claimed in claim 1, wherein the step 2) of generating the training target remaining life RUL based on the difference accumulation method comprises the following steps:

step 2.1) selecting a detection signal with a degradation trend, carrying out smooth filtering to reduce noise interference, and carrying out normalization on the selected signal data;

step 2.2) obtaining a degradation inflection point by adopting a differential accumulation method, namely for each rows of signal values, sequentially accumulating the difference between the numerical values of the rear moment and the numerical value of the front moment, setting a degradation threshold, and when the accumulated sum exceeds the threshold at a certain point and is greater than the threshold at the next continuous 4 points, selecting the point as the degradation inflection point, thereby obtaining the inflection point of each semaphore, and selecting the average value of the minimum 3 inflection points in the semaphore as the inflection point of the RUL of the group of engines;

and 2.3) adopting the reverse order of the time steps of the degradation data as the initial RUL, taking the RUL corresponding to the inflection point to update the RUL in the time period from the initial to the inflection point, and keeping the RULs in the rest time periods unchanged, thereby generating the training target RUL.

4. The method for predicting remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network as claimed in claim 1, wherein step 3) training the model for predicting remaining life of an engine based on a deep FNN-LSTM hybrid network based on the feature term and the training target RUL comprises the following steps:

step 3.1) constructing an aircraft engine residual life prediction model based on the deep FNN-LSTM mixed network, the training set and the training target RUL generated in the step 2), wherein the expression is as follows:

RUL_t＝FNN-LSTM(x₁，x₂，…，x_k，…，x_t)，(k＝1，2，…，t) (3)

wherein, t represents the time of day,

sets of feature term values representing the k-th time instant, i being the ith feature term,

characteristic item

For the ith detection signal

And its order difference

Second order difference

And constructing a three-dimensional detection data structure.

Step 3.2) for sets of training data, assume input data x at the current time_{input_t}Output as x through FNN network_{fnn_t}And the LSTM network hidden state at the time of upper is h_t-1The cell state is c_t-1Then, the output of the FNN-LSTM network at the current time is calculated as follows:

x_{fnn_t}＝σ(w_fnnx_{input_t}+b_fnn) (4)

f_t＝σ(w_fx_{fnn_t}+R_fh_t-1+b_f) (5)

～c_t＝tanh(W_cx_{fnn_t}+R_ch_t-1+b_c) (6)

i_t＝σ(w_ix_{fnn_t}+R_ih_t-1+b_i) (7)

c_t＝f_tc_t-1*i_t～c_t(8)

o_t＝σ(W_ox_{fnn_t}+R_oh_t-1+b_o) (9)

h_t＝o_t*tanh(c_t) (10)

where σ, tanh are activation functions, w represents weight, b represents bias, i_tIs an input , f_tTo forget , o_tIs output , h_tIs the output of the LSTM network at the current moment;

step 3.3) calculating the FNN-LSTM network in the forward direction according to the formulas (4) - (10), performing network training by adopting an Adam optimization algorithm, and obtaining an FNN-LSTM optimal prediction model through multiple parameter adjustment training;

and 3.4) predicting the test set by using the obtained FNN-LSTM prediction model to obtain a predicted value RUL.

The method has the beneficial effects that aviation engine residual life prediction methods based on the deep FNN-LSTM hybrid network are provided, and a reliable implementation scheme is provided for the health management and maintenance cost reduction of the aviation engine.

Drawings

FIG. 1 is a flow chart of a method for predicting the remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network

FIG. 2 shows the trend of the detected signal under the single operating condition

FIG. 3 shows the trend of the detected signal under multiple operating conditions

FIG. 4 is a comparison graph of predicted values and real values of FD001# Unit17

FIG. 5 is a comparison graph of predicted values and real values of FD002# Unit21

FIG. 6 is a comparison graph of predicted values and real values of FD0003# Unit24

FIG. 7 FD004# Unit31 comparison of predicted values to true values

Detailed Description

The following is a description of a specific embodiment of the present invention at step with reference to the drawings.

The invention discloses an method for predicting the residual life of an aircraft engine based on a deep FNN-LSTM hybrid network, which comprises the following steps in the following specific flow chart shown in the attached figure 1:

step 1) on the basis of detection signal data of a plurality of groups of aeroengines from a healthy state to a degraded state, -order difference and second-order difference of the detection signal data are added to form a three-dimensional detection data structure and the three-dimensional detection data structure is used as a characteristic item;

step 1.1) detection signal data order difference and second order difference information are adopted, and then a mathematical model is used for more accurately describing the detection signal data and the nonlinear relation between order difference and second order difference of the detection signal data and the engine change process:

wherein (1)

In the above formulas (1) to (3), k is a time point, r (k) represents a change process of the engine, s_kThe value of the detection signal at k is,

to detect the th order derivative of the value with respect to time,in practical applications, generally studies the variation process by collecting discrete values, so the difference in the above equations (2) and (3) is used to replace the differential in the continuous process.

Step 1.2) calculating order difference and second order difference of the engine detection signal data by a forward difference method;

step 1.3) in the detection signal data, expanding the corresponding signal column dimension by using -order difference and second-order difference of each signal, and then forming a three-dimensional detection data structure of < detection signal data, -order difference of detection signal data, second-order difference of detection signal data > as a feature item for each detection signal.

Step 2) generating a training target residual life RUL based on a difference cumulative addition method;

step 2.1) selecting a detection signal with a degradation trend, carrying out smooth filtering to reduce noise interference, and carrying out normalization on the selected signal data;

step 2.2) obtaining a degradation inflection point by adopting a differential accumulation method, namely for each rows of signal values, sequentially accumulating the difference between the numerical values of the rear moment and the numerical value of the front moment, setting a degradation threshold, and when the accumulated sum exceeds the threshold at a certain point and is greater than the threshold at the next continuous 4 points, selecting the point as the degradation inflection point, thereby obtaining the inflection point of each semaphore, and selecting the average value of the minimum 3 inflection points in the semaphore as the inflection point of the RUL of the group of engines;

and 2.3) adopting the reverse order of the time steps of the degradation data as the initial RUL, taking the RUL corresponding to the inflection point to update the RUL in the time period from the initial to the inflection point, and keeping the RULs in the rest time periods unchanged, thereby generating the training target RUL.

Step 3) training an engine residual life prediction model based on a deep FNN-LSTM mixed network according to a training set and a training target RUL, and predicting a test set by adopting an optimal model to obtain the engine residual life RUL;

step 3.1) constructing an aircraft engine residual life prediction model based on the deep FNN-LSTM mixed network, the training set and the training target RUL generated in the step 2), wherein the expression is as follows:

RUL_t＝FNN-LSTM(x₁，x₂，…，x_k，…，x_t)，(k＝1，2，…，t) (3)

wherein, t represents the time of day,

sets of feature term values representing the k-th time instant, i being the ith feature term,

characteristic item

For the ith detection signal

And its order difference

Second order difference

And constructing a three-dimensional detection data structure.

Step 3.2) for sets of training data, assume input data x at the current time_{input_t}Output as x through FNN network_{fnn_t}And the LSTM network hidden state at the time of upper is h_t-1The cell state is c_t-1Then, the output of the FNN-LSTM network at the current time is calculated as follows:

x_{fnn_t}＝σ(w_fnnx_{input_t}+b_fnn) (4)

f_t＝σ(w_fx_{fnn_t}+R_fh_t-1+b_f) (5)

～c_t＝tanh(w_cx_{fnn_t}+R_ch_t-1+b_c) (6)

i_t＝σ(w_if_{fnn_t}+R_ih_t-1+b_i) (7)

c_t＝f_tc_t-1*i_t～c_t(8)

o_t＝σ(w_ox_{fnn_t}+R_oh_t-1+b_o) (9)

h_t＝o_t*tanh(c_t) (10)

where σ, tanh are activation functions, w represents weight, b represents bias, i_tIs an input , f_tTo forget , o_tIs output , h_tIs the output of the LSTM network at the current moment;

step 3.3) calculating the FNN-LSTM network in the forward direction according to the formulas (4) - (10), performing network training by adopting an Adam optimization algorithm, and obtaining an FNN-LSTM optimal prediction model through multiple parameter adjustment training;

and 3.4) predicting the test set by using the obtained FNN-LSTM prediction model to obtain a predicted value RUL.

In order to verify the effectiveness of the method for predicting the residual life of the aircraft engine based on the deep FNN-LSTM hybrid network, a NASA (network-assisted analysis and maintenance) C-MAPSS turbine engine degradation data set is adopted for experimental verification, hundreds of groups of engines which normally operate but are worn to different degrees are selected for the C-MAPSS turbine engine degradation data set for experiment, under set operation conditions, the experiment makes or two faults on 5 rotating parts of the engine in normal operation, 58 sensors are adopted for detecting the whole process from the normal operation state to the fault state and finally to the complete fault state of the engine, and 21 detection signals are selected as effective signal data.

The experimental data comprises 4 data subsets under different operating conditions and fault types, wherein the FD001 subset is only operating conditions and fault types, the FD004 subset is the most complex and has 6 operating conditions and 2 fault modes, and table 1 shows specific information of each subset, each subset is divided into a training set and a testing set, all operating period data of the engine from to the final complete degradation are recorded in the training set, data in the testing set are recorded to moments before the complete degradation, and the residual operating period number is the predicted target RUL.

TABLE 1C-MAPSS data set essential information

Through the two graphs, the change trend of the signal quantity is obvious under the single operation condition, the signal quantity has obvious rising or falling trend and is easier to reflect the degradation rule of the engine, and under the multiple operation condition, the change of the signal quantity has no specific rule and no obvious trend, and the mining of the degradation rule of the engine is relatively complex and difficult.

It can also be seen from fig. 2 and 3 that the original detection signal data has noise, so the real data is restored by smoothing filtering and normalization processing, fig. 4, 5, 6 and 7 are graphs comparing the predicted results of the selected test engine in FD001-FD004 subset with the real results, respectively, and table 2 is the experimental results of the training set and the test set of 4 subsets.

Table 24 experimental results of training and test sets of subsets

Results fig. 4-7 and table 2 show that the prediction accuracy of the method for predicting the residual life of the aircraft engine based on the deep FNN-LSTM hybrid network is high. The RMSE values of the FD001 subset and the FD003 subset in the prediction results of the test set are low enough, and although the RMSE of the FD002 and the FD004 is relatively high, the prediction results reach high accuracy due to the complex operation environment and more fault types. Also, as can be seen from fig. 4-7, the predicted values for the 4 subsets are overall very close to the true values. In a word, the experimental result proves that the method for predicting the residual life of the aero-engine based on the deep FNN-LSTM hybrid network has effectiveness, and a reliable implementation scheme is provided for the health management and maintenance cost reduction of the aero-engine.

Claims (4)

1, aviation engine residual life prediction method based on deep FNN-LSTM hybrid network, which comprises the following steps:

step 1) on the basis of detection signal data of a plurality of groups of aeroengines from a healthy state to a degraded state, -order difference and second-order difference of the detection signal data are added to form a three-dimensional detection data structure and the three-dimensional detection data structure is used as a characteristic item;

step 2) generating a training target residual life RUL based on a difference cumulative addition method;

and 3) training an engine residual life prediction model based on the deep FNN-LSTM hybrid network according to the training set, the feature items and the training target RUL, and predicting the test set by adopting the optimal model to obtain the engine residual life RUL.

2. The method for predicting remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network as claimed in claim 1, wherein the specific steps of adding order difference and second order difference of the detection signal data to the engine detection signal data in step 1) to form a three-dimensional detection data structure as a characteristic item are as follows:

step 1.1) detection signal data order difference and second order difference information are adopted, and then a mathematical model is used for more accurately describing the detection signal data and the nonlinear relation between order difference and second order difference of the detection signal data and the engine change process:

wherein (1)

In the above formulas (1) to (3), k is a time point, r (k) represents a change process of the engine, s_kThe value of the detection signal at k is,

to detect the th order derivative of the value with respect to time,in practical applications, generally studies the variation process by collecting discrete values, so the difference in the above equations (2) and (3) is used to replace the differential in the continuous process.

Step 1.2) calculating order difference and second order difference of the engine detection signal data by a forward difference method;

step 1.3) in the detection signal data, expanding the corresponding signal column dimension by using -order difference and second-order difference of each signal, and then forming a three-dimensional detection data structure of < detection signal data, -order difference of detection signal data, second-order difference of detection signal data > as a feature item for each detection signal.

3. The method for predicting remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network as claimed in claim 1, wherein the step 2) of generating the training target remaining life RUL based on the difference accumulation method comprises the following steps:

step 2.1) selecting a detection signal with a degradation trend, carrying out smooth filtering to reduce noise interference, and carrying out normalization on the selected signal data;

step 2.2) obtaining a degradation inflection point by adopting a differential accumulation method, namely for each rows of signal values, sequentially accumulating the difference between the numerical values of the rear moment and the numerical value of the front moment, setting a degradation threshold, and when the accumulated sum exceeds the threshold at a certain point and is greater than the threshold at the next continuous 4 points, selecting the point as the degradation inflection point, thereby obtaining the inflection point of each semaphore, and selecting the average value of the minimum 3 inflection points in the semaphore as the inflection point of the RUL of the group of engines;

and 2.3) adopting the reverse order of the time steps of the degradation data as the initial RUL, taking the RUL corresponding to the inflection point to update the RUL in the time period from the initial to the inflection point, and keeping the RULs in the rest time periods unchanged, thereby generating the training target RUL.

4. The method for predicting remaining life of an aircraft engine based on a deep FNN-LSTM hybrid network as claimed in claim 1, wherein step 3) training the model for predicting remaining life of an engine based on a deep FNN-LSTM hybrid network based on the feature term and the training target RUL comprises the following steps:

step 3.1) constructing an aircraft engine residual life prediction model based on the deep FNN-LSTM mixed network, the training set and the training target RUL generated in the step 2), wherein the expression is as follows:

RUL_t＝FNN-LSTM(x₁，x₂，…，x_k，…，x_t)，(k＝1，2，…，t) (3)

wherein, t represents the time of day, sets of feature item values representing the k-th time instant, i being the ith featureThe items are,

characteristic itemFor the ith detection signal

And its order difference

Second order differenceAnd constructing a three-dimensional detection data structure.

Step 3.2) for sets of training data, assume input data x at the current time_{input_t}Output as x through FNN network_{fnn_t}And the LSTM network hidden state at the time of upper is h_t-1The cell state is c_t-1Then, the output of the FNN-LSTM network at the current time is calculated as follows:

x_{fnn_t}＝σ(w_fnnx_{input_t}+b_fnn) (4)

f_t＝σ(w_fx_{fnn_t}+R_fh_t-1+b_f) (5)

～c_t＝tanh(w_cx_{fnn_t}+R_ch_t-1+b_c) (6)

i_t＝σ(w_ix_{fnn_t}+R_ih_t-1+b_i) (7)

c_t＝f_tc_t-1*i_t～c_t(8)

o_t＝σ(w_ox_{fnn_t}+R_oh_t-1+b_o) (9)

h_t＝o_t*tanh(c_t) (10)

wherein σ and tanh are activation functionsNumber, w weight, b bias, i_tIs an input , f_tTo forget , o_tIs output , h_tIs the output of the LSTM network at the current moment;

step 3.3) calculating the FNN-LSTM network in the forward direction according to the formulas (4) - (10), performing network training by adopting an Adam optimization algorithm, and obtaining an FNN-LSTM optimal prediction model through multiple parameter adjustment training;

and 3.4) predicting the test set by using the obtained FNN-LSTM prediction model to obtain a predicted value RUL.

Images