Method for predicting residual service life of aero-engine under variable working conditions based on dynamic weighted adversarial migration network

Abstract

The invention discloses an aero-engine variable working condition residual service life prediction method based on a dynamic weighted adversarial migration network, and aims to solve the problem that the residual service life of an aero-engine is difficult to predict under the existing variable working condition. The method comprises the following steps: 1, establishing a gating circulation unit network based on a self-attention mechanism; 2, constructing a loss function of the migration network oriented to the adaptive edge distribution difference; 3, establishing a loss function of the migration network oriented to the adaptive condition distribution difference; 4, designing a dynamic weighting factor; 5, designing an overall loss function of the dynamic weighted adversarial migration network; 6, obtaining a trained dynamic weighted confrontation-based migration network; and 7, inputting the online data belonging to the target domain into the trained dynamic weighted adversarial migration network to realize online prediction of the remaining service life of the aero-engine. The method is used in the field of aero-engine variable working condition residual service life prediction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of intersection of aero-engine remaining useful life prediction and artificial intelligence, and specifically relates to a method for predicting the remaining useful life of aero-engine under variable working conditions. BACKGROUND

[0002] Under the background of intelligent upgrading of aviation propulsion systems, the precise prediction of the remaining useful life (RUL) of high thrust-to-weight ratio aero-engines, as the power core of modern aircraft, has become a key technical challenge to ensure flight safety and operational economy. For typical degradation modes such as crack propagation of aero-engine high-pressure turbine blades and creep damage of hot-end components in the combustion chamber, the remaining useful life prediction technology fuses time-series monitoring data of gas path parameters, blade vibration spectrum features, and historical maintenance records to build a quantitative mapping relationship between performance degradation and remaining life. This technology provides key support for on-condition maintenance decision-making, spare parts dynamic scheduling, and in-flight parking risk warning of aero-engines, and directly determines the life cycle operation cost and flight safety redundancy of airlines.

[0003] In recent years, researchers have proposed various methods for predicting the remaining useful life of aero-engines, which can be broadly divided into model-based methods and data-driven methods. Model-based methods for predicting the remaining useful life of aero-engines are based on the damage principle and mathematical model of the engine to achieve the prediction task. Since this type of method requires accurate physical parameters and models, it may be limited by model assumptions and parameter estimation in practical applications. On the other hand, data-driven methods for predicting the remaining useful life of aero-engines do not rely on the physical model of the equipment, but are based on historical operation data and monitoring information, and by extracting and learning features from aero-engine degradation data, a prediction model is established to achieve remaining useful life prediction.

[0004] In actual service environment, it is very difficult to obtain sufficient aero-engine life cycle degradation data, because of high cost of fault data acquisition equipment, access restriction of equipment and interruption of work caused by fault. Therefore, it is necessary to study aero-engine residual useful life prediction method under variable working condition, to predict residual useful life under unknown working condition by using labeled data under known working condition, so as to overcome the problem of data scarcity and save computing resources. It is worth mentioning that the traditional data-driven residual useful life prediction method often performs poorly under variable working condition, because there are differences in data distribution and feature representation between different working conditions. In recent years, the transfer learning method guided by domain adaptation provides a good opportunity to solve the problem of residual useful life prediction under variable working condition. The domain adaptation method reduces the feature distribution difference between the source domain data and the target domain data by learning the shared feature representation between the two kinds of data, so as to improve the generalization ability of the residual useful life prediction model. It is worth mentioning that in different degradation stages of aero-engine, the distribution of data will be different, and how to use a fine-grained measurement method to reduce this distribution difference needs further study. In addition, when implementing domain adaptation, it needs to consider two aspects of adaptive marginal distribution difference and adaptive conditional distribution difference, how to measure the relationship between the two and guide the variable working condition residual useful life prediction of aero-engine needs further work. SUMMARY

[0005] The purpose of the present application is to solve the problem of difficulty in predicting the residual useful life of an aero-engine under variable working conditions, and to propose an aero-engine variable working condition residual useful life prediction method based on a dynamic weighted adversarial transfer network.

[0006] The aero-engine variable working condition residual useful life prediction method based on the dynamic weighted adversarial transfer network has the following specific process:

[0007] Step one, build a gated recurrent unit network based on self-attention mechanism;

[0008] Step two, based on the built gated recurrent unit network based on self-attention mechanism, construct the loss function L of the transfer network for adaptive marginal distribution difference g ;

[0009] Step three, based on the built gated recurrent unit network based on self-attention mechanism, build the loss function L of the transfer network for adaptive conditional distribution difference l ;

[0010] Step four, based on the loss function L of the transfer network for adaptive marginal distribution difference L g and the loss function L of the transfer network for adaptive conditional distribution difference L l, a dynamic weighting factor w is designed;

[0011] Step five, the loss function L of the migration network based on adaptive edge distribution difference g , the migration network L based on adaptive conditional distribution difference l And the dynamic weighting factor w, the overall loss function of the dynamic weighting adversarial migration network is designed;

[0012] Step six, based on the overall loss function of the dynamic weighting adversarial migration network, the offline training based on the dynamic weighting adversarial migration network is completed, and the trained dynamic weighting adversarial migration network based on the dynamic weighting adversarial migration network is obtained;

[0013] Step seven, inputting the online data belonging to the target domain into the dynamic weighting adversarial migration network trained in step six, realizing the online prediction of the remaining useful life of the aero-engine.

[0014] The beneficial effects of the present application are:

[0015] (1) The present application learns from the adversarial transfer learning idea, and comprehensively designs the migration network based on adaptive edge distribution difference from the local and global two angles. At the same time, the present application measures the adaptive edge distribution difference of the aero-engine in different degradation stages from a fine-grained perspective.

[0016] (2) On the basis of adaptive edge distribution difference, the present application considers the adaptive conditional distribution difference under variable working condition, and designs multiple local domain discriminators to make up for the conditional distribution difference between the source domain data and the target domain data, and further improves the remaining useful life prediction effect of the aero-engine.

[0017] (3) In order to measure the influence degree of the migration network based on adaptive edge distribution difference and adaptive conditional distribution difference on the domain difference, the present application designs a dynamic weighting factor according to the two kinds of differences, so as to better realize the dynamic weighting adversarial migration learning. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The workflow diagram of the present application;

[0019] Figure 2 The self-attention mechanism diagram of the present application;

[0020] Figure 3 The schematic diagram of the gating recurrent unit of the present application;

[0021] Figure 4a The remaining useful life prediction result diagram of the task OC1-OC2;

[0022] Figure 4bA remaining useful life prediction result graph for the task OC2-OC1. DETAILED DESCRIPTION

[0023] Specific implementation one: the specific process of the aero-engine variable working condition remaining useful life prediction method based on the dynamic weighted adversarial transfer network in this embodiment is as follows:

[0024] The present application is designed from the aspects of adaptive edge distribution difference and adaptive conditional distribution difference. Further, the present application proposes a remaining useful life prediction method based on a dynamic weighted adversarial transfer network, a weighting factor with dynamics is designed according to the influence degree of the two differences, and dynamic weight adjustment between the two differences is realized, so that the dynamic weighted adversarial transfer learning is better realized. The work flow chart of the present application is shown in Figure 1 .

[0025] Step one, build a gated recurrent unit network based on a self-attention mechanism;

[0026] Step two, based on the built gated recurrent unit network based on a self-attention mechanism, a loss function L g of the transfer network for adaptive edge distribution difference is constructed.

[0027] Step three, based on the built gated recurrent unit network based on a self-attention mechanism, a loss function L l of the transfer network for adaptive conditional distribution difference is constructed.

[0028] Step four, based on the loss function L g of the transfer network for adaptive edge distribution difference and the loss function L l of the transfer network for adaptive conditional distribution difference, a dynamic weighting factor w is designed.

[0029] Step five, based on the loss function L g of the transfer network for adaptive edge distribution difference, the loss function L l of the transfer network for adaptive conditional distribution difference, and the dynamic weighting factor w, a total loss function of the dynamic weighted adversarial transfer network is designed.

[0030] Step six, based on the total loss function of the dynamic weighted adversarial transfer network, off-line training based on the dynamic weighted adversarial transfer network is completed, and a trained dynamic weighted adversarial transfer network is obtained.

[0031] Step seven, input online data belonging to the target domain into the trained dynamic weighted adversarial transfer network in step six, and realize online prediction of the aero-engine remaining useful life.

[0032] Specific implementation method two: the difference between this embodiment and specific implementation method one is that: the step one is to build a gated recurrent unit network based on a self-attention mechanism; the specific process is:

[0033] Considering the time sequence of the data collected by the sensors in the aero-engine, the data input into the deep neural network is time window data. Therefore, how to focus on the important time points in the time window data is the key to improving the effect of the remaining useful life prediction of the aero-engine. Therefore, without prior knowledge of the degradation of the aero-engine, the self-attention mechanism is introduced into the remaining useful life prediction problem, and greater weight values are given to important time points.

[0034] For convenience of expression, the input time window data of the self-attention mechanism layer is represented as: X={x i |i=1,2,…,d};

[0035] Wherein, x i ∈R m , m is the feature dimension of the data sample, d is the length of the time window; x i is the feature variable at the i th time; X is the feature of all time in the time window data; R is a real number;

[0036] The self-attention mechanism diagram is shown in Figure 2 ;

[0037] The essence of the self-attention mechanism layer in the application is to construct the mapping relationship between the feature variable x i at the i th time in the time window data and the importance s i of the feature variable x i , which is expressed in the form described in formula (1):

[0038]

[0039] Wherein, g is a sigmoid activation function, x i represents the feature variable at the i th time (the data collected by the sensors in the aero-engine at the i th time), w i is the weight matrix at the i th time, b i is the bias vector at the i th time, and T is the transpose operation.

[0040] Then the normalization of the importance degree of formula (1) is realized by the softmax function, as shown in formula (2):

[0041]

[0042] Wherein, β i is the importance s i of the feature variable x iThe degree of importance after normalization;

[0043] The output of the self-attention mechanism layer is obtained through matrix multiplication. As shown in equation (3):

[0044]

[0045] in, The first time-time feature variable is the output of the self-attention mechanism layer. The feature variables at the second time step are the outputs of the self-attention mechanism layer. Let be the feature variable at time i output by the self-attention mechanism layer. The feature variable at time d is the output of the self-attention mechanism layer;

[0046] β1 represents the importance of feature variable x1 after normalization to s1, and β2 represents the importance of feature variable x2 after normalization to s2. i For the characteristic variable x i Importance i The importance after normalization, β d For the characteristic variable x d Importance d The degree of importance after normalization;

[0047] x1 is the feature variable at time 1, x2 is the feature variable at time 2, x d Let d be the feature variable at time d;

[0048] After weights are assigned to different features within the time window in the self-attention mechanism layer, the input is fed into subsequent gated recurrent units (ROUs). A gated recurrent unit is an improved variant of a recurrent neural network, consisting of a reset gate and an update gate. The reset gate controls the degree to which new input information is combined with past information. The update gate controls the degree to which the network unit retains past information. A schematic diagram of a gated recurrent unit is shown below. Figure 3 As shown;

[0049] The output of the self-attention mechanism layer The input gated loop unit, the update gate, the reset gate, and the memory unit output characteristics of the gated loop unit are calculated using the formulas shown in equations (4) to (7):

[0050]

[0051] Among them, h i-1 This represents the output of the gated loop unit at time i-1; h i This represents the output of the gated loop unit at time i;

[0052] represents the input vector of the representation unit at the i-th moment;

[0053] z i , r i , c i respectively represent the output features of the update gate, the reset gate and the memory cell;

[0054] U z , U r , U c respectively represent the weight matrices of the update gate, the reset gate and the memory cell;

[0055] b z , b r and b c respectively represent the bias vectors of the update gate, the reset gate and the memory cell;

[0056] σ represents a sigmoid activation function, tanh represents a hyperbolic tangent function, represents a dot product operation.

[0057] The other steps and parameters are the same as those in the first embodiment.

[0058] The third embodiment is different from the first or second embodiment in that the loss function L g of the transfer network oriented to adaptive edge distribution difference is constructed based on the built gated recurrent unit network based on the self-attention mechanism in the second step.

[0059] The transfer network oriented to adaptive edge distribution difference in the application contains two sub-aspects in terms of implementation details, which are respectively a global transfer network oriented to adaptive edge distribution difference and a locally weighted transfer network oriented to adaptive edge distribution difference.

[0060] Step two one, constructing a loss function of a global transfer network oriented to adaptive edge distribution difference; the specific process is as follows:

[0061] For the global transfer network oriented to adaptive edge distribution difference, the application borrows the idea of adversarial transfer learning, and respectively corresponds the real samples and random noise of the generative adversarial network to the source domain degraded data and the target domain degraded data in the transfer learning; the reason is that the source domain data has labels, which corresponds to the real samples in the generative adversarial network; on the other hand, the target domain data is unlabeled, which can correspond to the random noise in the generative adversarial network.

[0062] Specifically,

[0063] The source domain data sample D s is represented as:

[0064] wherein, and respectively represent the feature variable represented by the source domain sample at the i-th moment and the remaining service life label at the i-th moment; N s represents the total number of source domain samples;

[0065] target domain data sample D T represents:

[0066] wherein, represents the feature variable represented by the target domain sample at the j-th moment; N T represents the total number of target domain samples;

[0067] Loss function of global transfer network oriented to adaptive edge distribution difference The expression is shown in formula (8):

[0068]

[0069] wherein, x k represents the input sample of the source domain and the target domain at the k-th moment;

[0070] f SAMGRU is the mapping relationship constructed by the gate recurrent unit based on the self-attention mechanism in step 1 (from x k to the update gate, the reset gate and the memory cell output feature of the gate recurrent unit);

[0071] f d is the mapping relationship of the domain discriminator, which is used to judge whether the input data sample comes from the source domain or the target domain;

[0072] d k represents the domain label;

[0073] L p represents the cross-entropy loss function;

[0074] Formula (8) is expressed as described in formula (9):

[0075]

[0076] wherein, d k and are set to 0 or 1 to show that the data comes from the source domain or the target domain;

[0077] d k = 0 indicates that the k-th sample belongs to the true source domain label, and d k = 1 indicates that the k-th sample belongs to the true target domain label;

[0078] represents that the kth sample belongs to a predicted source domain label, represents that the kth sample belongs to a predicted target domain label.

[0079] Step two, constructing a loss function of a locally weighted transfer network oriented to adaptive edge distribution difference The specific process is as follows:

[0080] In order to more comprehensively reflect the edge distribution difference between the source domain and the data domain, the application also designs a globally transferred network oriented to adaptive edge distribution difference and a locally weighted transfer network oriented to adaptive edge distribution difference. The motivation of the network is to consider that the global domain adaptive method will confuse the fine-grained features between the sub-domains represented by different degradation stages of the aero-engine, and a locally weighted network is introduced. Considering the infinity of the remaining useful life label and the internal similarity of different stages in the aero-engine degradation process, before introducing the method, it is necessary to first explain the sub-domain design of the aero-engine degradation;

[0081] Specifically,

[0082] The degradation process of the aero-engine is divided into three stages of the health period (early degradation), the recession period (middle degradation) and the retirement period (late degradation), which can be specifically represented in the form shown in Figure 4.

[0083] The health period (early degradation), the recession period (middle degradation) and the retirement period (late degradation) are three stages of self-defined division;

[0084] Suppose that the maximum limit value of the remaining useful life is T r Then, for the source domain sample at the i th moment, the designed soft label l i is expressed in the form shown in formula (10):

[0085]

[0086] Wherein, round represents a rounding function;

[0087] RUL ti is the real value of the remaining useful life of the source domain sample at the i th moment;

[0088] The maximum mean difference is used to measure the difference between the source domain data distribution u and the target domain data distribution v As shown in formula (11):

[0089]

[0090] Wherein,

[0091] H k Represent the regenerable kernel Hilbert space defined by the significant kernel k;

[0092] H represents k The square norm is often used to measure distances distributed in a high-dimensional feature space;

[0093] x s and x t Representing source domain data samples D respectively s and target domain data sample D T Data in the middle;

[0094] φ defines the mapping from the original data to the reproducible kernel Hilbert space RKHS, and k implies the kernel function relation k(x) s ,x t )=<φ(x s ),φ(x t )>, <.> means inner product operation; k(x s ,x t ) represents the kernel function, used to measure the performance of a sample x. s and x t Similarity in regenerable kernel Hilbert space (high-dimensional feature space);

[0095] φ(x s ) indicates that sample x s The feature mapping result from the original feature space to the reproducible kernel Hilbert space (high-dimensional feature space); φ(x t ) indicates that sample x t Feature mapping results from the original feature space to the reproducible kernel Hilbert space (high-dimensional feature space);

[0096] E x～u E represents the expected value of a variable x with distribution u. x～v This represents the expected value of a variable x with distribution v.

[0097] It is worth mentioning that this invention considers the inherent connections between different stages in the system degradation process, and based on equation (11), introduces a locally weighted maximum mean difference measure to measure the distance between relevant subdomains in the source domain and the target domain. Represented in the form shown in formula (12):

[0098]

[0099] in, The regenerable kernel Hilbert space H is defined by the kernel function k. k The distance between the distribution u and the distribution v in the middle; This represents the expected value of a variable x with distribution u(c); This represents the expected value of a variable x with distribution v(c); u(c) and v(c) represent source domain data samples belonging to the c-th degradation stage, respectively. and target domain data samples Data distribution; the c-th stage of degradation is either the healthy period (early degradation), the declining period (intermediate degradation), or the retirement period (late degradation); E c This represents the mathematical expectation value for the c-th degradation stage;

[0100] Specifically, each sample belonging to each degradation stage will be assigned a different weight, and formula (12) is further rewritten as shown in formula (13):

[0101]

[0102] in, and These represent the i-th source domain data samples belonging to the c-th degradation stage, respectively. and the j-th target domain data sample The weight value;

[0103] The loss function represents a locally weighted migration network that adapts to differences in marginal distributions.

[0104] Indicates that the sample Feature mapping results from the original feature space to the reproducible kernel Hilbert space (high-dimensional feature space);

[0105] Indicates that the sample Feature mapping results from the original feature space to the reproducible kernel Hilbert space (high-dimensional feature space);

[0106] and The conditions described in formula (14) must be met:

[0107]

[0108] For the source domain data sample at time i weight value The expression is shown in formula (15):

[0109]

[0110] For the target domain data sample at time j weight value The expression is shown in the following formula:

[0111]

[0112] wherein, denotes the element of the soft label vector l i of the cth degradation stage; denotes the element of the soft label vector l j of the cth degradation stage;

[0113] denotes the element of the soft label vector l m of the cth degradation stage;

[0114] (x m ,y m ,l m ) denotes a triplet of the sample at the mth moment, including the input feature, the remaining life label and the soft label vector of the degradation stage;

[0115] Step two three, the loss function of the global transfer network based on the constructed adaptive edge distribution difference and the loss function of the locally weighted transfer network of adaptive edge distribution difference obtain the loss function L g of the transfer network facing adaptive edge distribution difference.

[0116] The other steps and parameters are the same as those in embodiment one or two.

[0117] Embodiment four: the embodiment is different from one of embodiments one to three in that: in the step two three, the loss function of the global transfer network based on the constructed adaptive edge distribution difference and the loss function of the locally weighted transfer network of adaptive edge distribution difference obtain the loss function L g of the transfer network facing adaptive edge distribution difference; the specific process is:

[0118] It should be noted that the soft label vector l i is a one-hot encoding vector for the discretized degradation stage; when the sample x i comes from the source domain data sample D s , the soft label vector l i can be directly obtained by formula (10); however, the method proposed in the present application is essentially unsupervised domain adaptation. Therefore, when the sample x i comes from the target domain data D T , the soft label vector l iThen, the prediction value of the remaining useful life is calculated by the forward process of the gated recurrent unit based on the self-attention mechanism in formulas (1)-(8), and then the calculation is performed according to the method in formula (10). According to the results of formula (9) and formula (13), the loss function L of the transfer network oriented to the adaptive edge distribution difference g The expression is in the form described in formula (16):

[0119]

[0120] The other steps and parameters are the same as one of the first to third embodiments.

[0121] Embodiment five: different from one of the first to fourth embodiments is that: in the step three, based on the built gated recurrent unit network based on the self-attention mechanism, the loss function L of the transfer network oriented to the adaptive condition distribution difference is built l ;

[0122] The specific process is:

[0123] The application not only constructs the transfer network oriented to the adaptive edge distribution difference, but also constructs the transfer network oriented to the adaptive condition distribution difference. In the transfer network oriented to the adaptive condition distribution difference, based on the idea of adversarial transfer, multiple local domain discriminators are designed to make up for the condition distribution difference between the source domain data and the target domain data. Similar to the transfer network oriented to the adaptive edge distribution difference, the transfer network oriented to the adaptive condition distribution difference is also based on the high-dimensional features extracted by the gated recurrent unit based on the self-attention mechanism (h t ) in formula 7.

[0124] The loss function L of the transfer network oriented to the adaptive condition distribution difference l As shown in formula (17):

[0125]

[0126] Among them,

[0127] represents the domain discriminant cross-entropy loss function for the cth degradation stage;

[0128] represents the domain discriminator mapping relationship of the cth degradation stage;

[0129] represents the degradation stage prediction label of the sample x k ;

[0130] represents the domain label from the cth degradation stage;

[0131] f SAMGRU represents the mapping relationship constructed by the gated recurrent unit based on the self-attention mechanism in step 1;

[0132] The formula (17) is expressed in the form as described in formula (18):

[0133]

[0134] wherein, and are set to 0 or 1 to show that the data comes from the source domain or the target domain;

[0135] represents that the sample at the kth moment belongs to the real source domain label under the cth degradation stage, represents that the sample at the kth moment belongs to the real target domain label under the cth degradation stage;

[0136] represents that the sample at the kth moment belongs to the predicted source domain label under the cth degradation stage, represents that the sample at the kth moment belongs to the predicted target domain label under the cth degradation stage;

[0137] represents the loss function of the transfer network facing the adaptive conditional distribution difference under the cth degradation stage.

[0138] The other steps and parameters are the same as one of the first to fourth embodiments.

[0139] Embodiment six: the embodiment is different from one of the first to fifth embodiments in that: in the step four, the loss function L g of the transfer network facing the adaptive edge distribution difference l , a dynamic weighting factor w is designed; the specific process is:

[0140] In order to measure the influence degree of the transfer network facing the adaptive edge distribution difference and the adaptive conditional distribution difference on the domain difference, it is necessary to design a weighting factor with dynamics according to the two kinds of differences, so as to better realize the dynamic weighted transfer learning; for this purpose, the present application considers the overall and local properties of the field, and introduces an A-distance to measure the degree of the two kinds of difference losses;

[0141] Specifically, the domain distance d A,g (D S ,D T ) facing the adaptive edge distribution difference is calculated in the form shown in formula (19):

[0142]

[0143] Domain distance d facing adaptive conditional distribution difference A,l (D S ,D T ) is calculated in the form described by formula (20):

[0144]

[0145] Wherein, represents the sample from the cth degradation stage of the source domain, represents the sample from the cth degradation stage of the target domain;

[0146] According to d A,g (D S ,D T ) and The dynamic weighting factor w is designed as shown in formula (21):

[0147]

[0148] The calculation of the dynamic weighting factor w can be carried out at the end of each iteration, according to the influence degree of the adaptive marginal distribution difference and the adaptive conditional distribution difference. Finally, as the training converges, the dynamic weighting factor w will also converge to a specific value.

[0149] Other steps and parameters are the same as those in specific embodiments one to five to one.

[0150] Specific embodiment seven: this embodiment is different from one of specific embodiments one to six: the loss function L g of the transfer network facing the adaptive marginal distribution difference in step five, the transfer network L l facing the adaptive conditional distribution difference, and the dynamic weighting factor w, the overall loss function of the dynamic weighted adversarial transfer network is designed; the specific process is:

[0151] The overall loss function of the dynamic weighted adversarial transfer network is to minimize the difference L y between the source domain real label and the source domain predicted label and maximize the field discriminator loss L d ; which is specifically expressed in the form described by formula (22):

[0152] Wherein,

[0153] L y represents the difference between the source domain real label and the source domain predicted label;

[0154] Ld represents the domain discriminator loss;

[0155] θ SAMGRU represents the parameter of the gated recurrent unit network based on the self-attention mechanism;

[0156] θ d and θ y respectively represent the parameters of the domain discriminator sub-neural network and the regression sub-neural network;

[0157] f y represents the mapping function of the regression sub-neural network;

[0158] λ represents a weight coefficient in the loss function, used to measure the weight between the regression loss and the migration loss in the model;

[0159] f SAMGRU represents the mapping relationship constructed by the gated recurrent unit based on the self-attention mechanism in step 1;

[0160] RUL k represents the remaining useful life of the sample x k ;

[0161] f d represents the mapping relationship of the domain discriminator, used to judge whether the input data sample comes from the source domain or the target domain;

[0162] d k represents the domain label;

[0163] L g represents the cross-entropy loss function; L l represents the loss function of the entire adaptive conditional distribution difference-oriented migration network;

[0164] In the dynamic weighted adversarial migration network proposed in the application, the domain discriminator sub-neural network is the core module for realizing domain adaptation and feature distribution alignment. Its essential role is to discriminate the input high-dimensional feature vector to distinguish whether the sample comes from the source domain or the target domain, thereby guiding the feature extraction network to gradually learn more domain-independent feature expression. Specifically, the domain discriminator sub-neural network takes the time sequence feature vector h k extracted after processing by the self-attention mechanism and the gated recurrent unit as input, and sequentially passes through the full connection layer, the ReLU activation function layer, the full connection layer, and the ReLU activation function layer for feature transformation, and finally outputs a probability value for representing the possibility of the input feature belonging to the target domain; the specific mapping relationship is f d .

[0165] p k =f d (W dh k +b d )(26)

[0166] wherein, p k is the probability value of the kth sample, W d and b d respectively represent the weight matrix and bias vector of the domain discriminant sub neural network.

[0167] In the dynamic weighted adversarial transfer network of the application, the regression sub neural network undertakes the core task of accurately mapping the high-dimensional time sequence feature vector to the aero-engine remaining useful life prediction value. Specifically, the regression sub neural network takes the feature vector h k processed by the self-attention mechanism and the gated recurrent unit as input, sequentially passes through the full connection layer, the ReLU activation function layer, the full connection layer, and the ReLU activation function layer to realize the nonlinear mapping from complex time sequence feature information to the continuous numerical value of the remaining useful life, f y represents the mapping function of the regression sub neural network; and is expressed as:

[0168] RUL pk =f y (W y h k +b y ) (27)

[0169] wherein, RUL pk is the remaining useful life prediction value of the kth sample, W y and b y respectively represent the weight matrix and bias vector of the regression sub neural network.

[0170] The parameter update of the dynamic weighted adversarial transfer network is shown in formulas (23)-(25):

[0171]

[0172] wherein, η represents the learning rate of the neural network;

[0173] By minimizing the defined overall loss function L(θ SAMGRU ,θ d ,θ y ), the gated recurrent unit network based on the self-attention mechanism, the domain discriminant sub neural network and the regression sub neural network are optimized; the domain invariance of the degradation data can be well learned, so that the neural network trained using data from the labeled source domain can correctly predict the remaining useful life from the unlabeled target domain data.

[0174] The other steps and parameters are the same as those in the first to sixth embodiments.

[0175] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One through Seven in that: the mapping function f of the regression sub-neural network... y (f SAMGRU (x k The specific operation process is as follows:

[0176] f SAMGRU (x k The first fully connected layer, the first ReLU activation function layer, the second fully connected layer, and the second ReLU activation function layer are input sequentially. The second ReLU activation function layer outputs the predicted remaining lifetime value.

[0177] Mapping relationship f of the domain discriminator d (f SAMGRU (x k The specific operation process is as follows:

[0178] f SAMGRU (x k The third fully connected layer, the third ReLU activation function layer, the fourth fully connected layer, and the fourth ReLU activation function layer are input sequentially. The fourth ReLU activation function layer outputs the probability value.

[0179] The other steps and parameters are the same as those in specific implementation methods one through seven.

[0180] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One to Eight in that: in step six, offline training of the dynamic weighted adversarial transfer network is completed based on the overall loss function of the dynamic weighted adversarial transfer network, and a trained dynamic weighted adversarial transfer network is obtained.

[0181] The specific process is as follows:

[0182] The framework of the cross-domain remaining lifetime prediction method based on dynamic weighted adversarial migration network proposed in this invention is as follows: Figure 1 As shown.

[0183] 1) Combine sensor data collected from the source and target domains under conditions in the form of time windows to form time window data from the source and target domains;

[0184] The raw time-series sensor data of the aero-engine is divided into multiple input segments according to time windows. Each window contains sensor observations at several consecutive time points, which are used as input to the model.

[0185] 2) Obtain the outputs of the gating loop units in the source domain and the target domain respectively from the time window data from the source domain and the target domain through step 1;

[0186] 3) Based on the outputs of the gated recurrent unit of the source domain and the gated recurrent unit of the target domain obtained in 2), the global and local loss functions of adaptive edge distribution difference are calculated respectively by step 2, and the two kinds of losses are added to obtain the edge distribution difference loss;

[0187] Based on the outputs of the gated recurrent unit of the source domain and the gated recurrent unit of the target domain obtained in 2), the adaptive conditional distribution difference loss is calculated by step 3;

[0188] 4) Based on the adaptive edge distribution difference loss and the adaptive conditional distribution difference loss, the weighting factor of dynamic is calculated by step 4;

[0189] 5) Based on 3), 4), the overall loss function of dynamic weighted adversarial transfer network is calculated by step 5, and the parameters of dynamic weighted adversarial transfer network are updated;

[0190] 6) Repeat 1), 2), 3), 4), 5) until the overall loss function of dynamic weighted adversarial transfer network converges, complete the offline training of dynamic weighted adversarial transfer network, and obtain the trained dynamic weighted adversarial transfer network.

[0191] Other steps and parameters are the same as one of the first to eighth embodiments.

[0192] The tenth embodiment is different from one of the first to ninth embodiments in that the online data belonging to the target domain is input into the dynamic weighted adversarial transfer network trained in step 6 in step 7 to realize online prediction of the remaining useful life of the aero-engine.

[0193] The data collected by the sensors in the online aero-engine belonging to the target domain is divided according to a fixed length time window to obtain a time sequence feature sequence X;

[0194] The time sequence feature sequence X is input into the self-attention mechanism layer, and the self-attention mechanism layer outputs (Formula (1-3));

[0195] The self-attention mechanism layer output is input into the gated recurrent unit, and the gated recurrent unit outputs h (Formula (4-7));

[0196] The gated recurrent unit output h is input into the regression sub-neural network (Formula (27)), and the regression sub-neural network outputs the remaining useful life prediction value; the process is as follows:

[0197] The door control cycle unit output h is sequentially input into a first full connection layer, a first ReLU activation function layer, a second full connection layer, and a second ReLU activation function layer, and the second ReLU activation function layer outputs the remaining service life prediction value.

[0198] Other steps and parameters are the same as those in the first to ninth embodiments.

[0199] The beneficial effects of the present application are verified by the following embodiments:

[0200] Embodiment 1:

[0201] The present application adopts a commercial modular aero-propulsion system simulation (CMAPSS) dataset, which simulates the performance degradation process of an aero-engine through multi-condition coupling modeling. The experimental data covers the full cycle degradation trajectory of a turbofan engine under four differentiated operating modes (including different combinations of thrust load and environmental parameters), and each condition can be divided into a training subset and a verification subset. The training subset contains complete time series monitoring records of the engine from the initial healthy state to functional failure, and the verification subset intercepts part of the degradation data in the early running stage of the engine, and takes the calibrated remaining life at the termination time as the model performance evaluation benchmark. Each condition sample is composed of a 26-dimensional multivariate time series, including: engine number, cumulative cycle number, 3 types of working condition parameters (altitude, Mach number, sea level temperature), and 21 channel sensor measurement values (such as exhaust port temperature, pressure, engine pressure ratio, speed, etc.).

[0202] Table 1 CMAPSS aero-engine dataset description

[0203]

[0204] As shown in Table 1, the four subsets of the data set correspond to different working condition configurations and failure mechanisms, resulting in significant data distribution offset characteristics among the subsets. Therefore, under unsupervised conditions, the field adaptation of variable working conditions faces significant challenges. To solve this problem, the research task of the present application is to select single working condition labeled data as the source domain, train a dynamic weighted adversarial transfer network to establish a degradation feature alignment mechanism between the source domain and the target domain, and then realize the remaining useful life prediction of the aircraft turbofan engine under the target domain working condition. In order to fully verify the robustness of the method, the present application designs multiple sets of variable working condition experiments (source domain→target domain), covering asymmetric working condition transfer scenarios, and the specific configurations are shown in Table 2. First, data preprocessing is performed, and 14 kinds of sensor data such as numbers 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20 and 21 are selected from the multi-working condition aircraft turbofan engine data set. Then, the original time series sensor data of the aircraft engine is divided into multiple input segments according to the time window, each window contains continuous sensor observations at several time points, and the real remaining useful life of the aircraft engine at the corresponding time point constitutes the input and output of the data set.

[0205] Table 2 Remaining life prediction tasks under multiple variable working condition combinations

[0206]

[0207] Step one, build a gated recurrent unit network based on self-attention mechanism;

[0208] The gated recurrent unit network based on self-attention mechanism in the present application includes two parts: a self-attention mechanism subnetwork and a gated recurrent unit subnetwork. The main function is: the gated recurrent unit based on self-attention mechanism gives greater weight to important time points in the window composed of time series degradation data; the gated recurrent unit subnetwork extracts the time dependence implied in the time series degradation data to obtain high-dimensional features for the transfer task.

[0209] Step two, based on the built gated recurrent unit network based on self-attention mechanism, construct the loss function L of the transfer network for adaptive edge distribution difference g ;

[0210] Step three, based on the built gated recurrent unit network based on self-attention mechanism, build the loss function L of the transfer network for adaptive conditional distribution difference l ;

[0211] Step four, based on the loss function L of the transfer network for adaptive edge distribution difference g and the loss function L of the transfer network for adaptive conditional distribution difference l , design a dynamic weighting factor w;

[0212] Step five, the loss function L of the adaptive edge distribution difference-oriented transfer network g , the adaptive condition distribution difference-oriented transfer network L l and the dynamic weighting factor w, the overall loss function of the dynamic weighting adversarial transfer network is designed;

[0213] Step six, based on the overall loss function of the dynamic weighting adversarial transfer network, the offline training based on the dynamic weighting adversarial transfer network is completed, and the trained dynamic weighting adversarial transfer network is obtained;

[0214] The loss function of network training is the mean square error loss function, the optimization algorithm is the Adam optimization algorithm, the learning rate is 0.001, and the network training process is carried out in the hardware environment of 1 GPU (GTX 3080Ti graphics card).

[0215] Step seven, inputting online data belonging to the target domain into the dynamic weighting adversarial transfer network trained in step six, realizing online prediction of the remaining useful life of the aero-engine.

[0216] Evaluate the remaining useful life prediction effect: the absolute average error and the root mean square error are used to evaluate the prediction effect of the adaptive remaining useful life.

[0217] Figure 4a 、 4b The prediction value and the true value of the remaining useful life of the aircraft turbofan engine test data set under the OC1-OC2 task in table 2 are shown. Table 3 shows the comparison of the remaining useful life prediction task results of the present application and various existing transfer learning methods under various variable working condition combinations. Figure 4a 、 4b And from table 3, it can be seen that the method proposed in the present application can effectively predict the remaining useful life of the aero-engine under variable working condition, which can assist users in effectively evaluating the health state of the system and reasonably formulating maintenance plans, thereby significantly optimizing the maintenance strategy of the whole system.

[0218] Table 3 Comparison of prediction results of the present application and various existing transfer learning methods

[0219]

[0220] The present application also has other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, but these corresponding changes and modifications should all belong to the protection scope of the claims attached to the present application.

Claims

1. An aero-engine variable working condition residual useful life prediction method based on a dynamic weighted adversarial transfer network, characterized in that: The method specifically comprises the following steps: Step one, building a gated recurrent unit network based on a self-attention mechanism; Step two, based on the built self-attention mechanism based gated recurrent unit network, the loss function L of the adaptive edge distribution difference oriented transfer network is constructed g ; Step three, based on the built self-attention mechanism based gated recurrent unit network, build the loss function L of the adaptive conditional distribution difference oriented transfer network l ; Step four, the loss function L of the migration network based on adaptive edge distribution difference g and the loss function L of the migration network based on adaptive conditional distribution difference l , design a dynamic weighting factor w; Step five, the loss function L of the migration network based on adaptive edge distribution difference g The migration network L based on adaptive conditional distribution difference l And a dynamic weighting factor w, design the overall loss function of the dynamic weighted adversarial migration network; Step six, obtaining a trained dynamic weighted adversarial transfer network based on a total loss function of the dynamic weighted adversarial transfer network; Step seven, inputting online data belonging to a target domain into the trained dynamic weighted adversarial transfer network in step six to realize online prediction of the remaining useful life of an aero-engine.

2. The aero-engine off-design remaining useful life prediction method based on dynamic weighting adversarial migration network according to claim 1, characterized in that: The step one of building the gated recurrent unit network based on the self-attention mechanism specifically comprises the following steps: The input time window data of the self-attention mechanism layer is represented as: X = {x i i = 1, 2, …, d}; where x i ∈R m , m is the feature dimension of the data sample, d is the length of the time window; x i is the feature variable at the i-th time; X is the features of all times in the time window data; R is a real number; Construct the feature variable x at time i in the time window data i With characteristic variable x i Importance i The mapping relationship between them can be expressed in the form described by formula (1): where g is a sigmoid activation function, x i represents the feature variable at the i-th moment, w i is the weight matrix at the i-th moment, b i is the bias vector at the i-th moment, and T is the transpose operation. Then, the importance degree of formula (1) is normalized through a softmax function, as shown in formula (2): wherein β i is the normalized importance degree of the importance s i of the feature variable x i ; The output of the self-attention mechanism layer is obtained by matrix multiplication As shown in equation (3): wherein, is a feature variable of the 1st time instant output by the self-attention mechanism layer, is a feature variable of the 2nd time instant output by the self-attention mechanism layer, is a feature variable of the i-th time instant output by the self-attention mechanism layer, is a feature variable of the d-th time instant output by the self-attention mechanism layer; β1 is the normalized importance degree of the importance s1 of the characteristic variable x1, β2 is the normalized importance degree of the importance s2 of the characteristic variable x2, β i is the normalized importance degree of the importance s i of the characteristic variable x i is the normalized importance degree of the importance s d of the characteristic variable x d is the normalized importance degree of the importance s d of the characteristic variable x x1 is the characteristic variable at the 1st time, x2 is the characteristic variable at the 2nd time, x d is the characteristic variable at the dth time; The output of the self-attention mechanism layer is input into the input-gated recurrent unit The input-gated recurrent unit, the update gate, the reset gate and the memory cell output features of the gated recurrent unit, and the calculation formula is shown in formula (4)-(7): wherein h i-1 represents the output of the gating recurrent unit at time i-1; h i represents the output of the gating recurrent unit at time i; represents the input vector of the unit at time i; z i 、r i 、c i respectively represent the update gate, the reset gate and the output feature of the memory cell; U z , U r , U c Wm, Wre, and Wm represent the weight matrices of the update gate, the reset gate, and the memory cell, respectively. b z , b r , and b c represent the update gate, the reset gate, and the bias vector of the memory cell, respectively; σ denotes a sigmoid activation function, tanh denotes a hyperbolic tangent function, denotes a dot product operation.

3. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 2, characterized in that: The step two is based on the built self-attention mechanism based gated recurrent unit network, and a loss function L of the adaptive edge distribution difference oriented migration network is constructed g The specific process is as follows: Step two one, building a loss function of a global transfer network facing adaptive edge distribution difference; the specific process is as follows: source domain data samples D s is represented as: wherein, and Xi and Yi represent the feature variable represented by the source domain sample at the i-th moment and the remaining useful life label at the i-th moment, respectively; N s represents the total number of source domain samples; Target domain data samples D T is represented as: wherein, represents the feature variable represented by the target domain sample at the jth moment; N T represents the total number of target domain samples; Loss function for global transfer network oriented to adaptive edge distribution difference The expression is shown as formula (8): wherein x k denotes the input sample of the source domain and the target domain at the kth time instant; f SAMGRU Mapping relationship constructed for the gating recurrent unit based on self-attention mechanism in step 1; f d a mapping relationship of the domain discriminator, used for judging whether an input data sample is from a source domain or a target domain; d k represents a domain tag; L p denotes the cross-entropy loss function; Formula (8) is expressed as formula (9) as follows: Wherein, d k = 0 indicates that the kth sample belongs to the real source domain label, d k = 1 indicates that the kth sample belongs to the real target domain label; denotes that the k-th sample belongs to the predicted source domain label, denotes that the k-th sample belongs to the predicted target domain label; Step two, constructing a loss function of the locally weighted migration network oriented to adaptive edge distribution difference The specific process is as follows: The degradation process of the aero-engine is divided into three stages of a healthy period, a recession period and a retirement period; Assuming that the maximum limit value of the remaining useful life is T r Then, for the source domain sample at the i-th moment, the designed soft label l i is expressed in the form shown in equation (10): Wherein, round represents a rounding function; RUL ti RULi is the true value of the remaining useful life of the source domain sample at the i-th time instant. The maximum mean discrepancy is used to measure the difference between the source domain data distribution u and the target domain data distribution v As shown in equation (11): Wherein, H k denotes the reproducing kernel Hilbert space defined by the kernel k; represents H k the square norm of x s and x t represent data in source domain data samples D s and target domain data samples D T respectively; φ defines a mapping relationship from original data to a renewable kernel Hilbert space RKHS; φ(x s ) denotes the feature mapping result of mapping the sample x s from the original feature space to the reproducing kernel Hilbert space; φ(x t ) denotes the feature mapping result of mapping the sample x t from the original feature space to the reproducing kernel Hilbert space; E x～u denotes the mathematical expectation value of the variable x with distribution u x～v denotes the mathematical expectation value of the variable x with distribution v Based on equation (11), a locally weighted maximum mean difference measure is introduced to measure the distance between the relevant subdomains in the source domain and the target domain is expressed in the form shown in equation (12): wherein, denotes a reproducing kernel Hilbert space H k defined by a kernel function k; denotes the distance between the distributions u and v; (c) denotes the mathematical expectation of the variable x with the distribution v (c) ; u (c) and v (c) denote the data distributions of the source domain data samples and target domain data samples belonging to the cth degradation phase; the cth degradation phase is the healthy phase, the degradation phase, or the decommissioning phase; E c denotes the mathematical expectation of the cth degradation phase;​ Each sample belonging to each degradation stage is given a different weight, and formula (12) is further rewritten as formula (13) as shown: wherein, and respectively denote the weight value of the i-th source domain data sample belonging to the c-th degradation stage and the j-th target domain data sample . a loss function representing a locally weighted transfer network facing adaptive edge distribution difference; representing mapping the sample a feature mapping result from the original feature space to the reproducing kernel hilbert space; representing mapping the sample a feature mapping result from the original feature space to the reproducing kernel hilbert space; and satisfies the condition described by equation (14): a weight value of the i-th time point source domain data sample The expression of the weight value of the i-th time point source domain data sample is shown as formula (15):​ a weight value of the target domain data sample at the jth moment The expression of the weight value of the target domain data sample at the jth moment is shown in the following formula: wherein, denotes the element of the soft label vector l i of the c-th degradation phase; denotes the element of the soft label vector l j of the c-th degradation phase; represents the soft label vector l m the element of the c-th degradation phase of (x m ,y m ,l m ) represents a triplet of samples at the mth moment, including input features, residual life labels and soft label vectors of degradation stages; Step two three, loss function of the global transfer network based on the constructed adaptive edge distribution difference and the loss function of the local weighted transfer network of adaptive edge distribution difference obtain the loss function L of the transfer network for adaptive edge distribution difference g .

4. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 3, characterized in that: The loss function of the global migration network based on the constructed adaptive edge distribution difference in step two and the loss function of the local weighted migration network of adaptive edge distribution difference The loss function L of the migration network for adaptive edge distribution difference is obtained g The specific process is: Loss function L of the migration network oriented to adaptive edge distribution difference g The expression is in the form described by equation (16):

5. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 4, characterized in that: The step three is based on the built self-attention mechanism based gated recurrent unit network, and a loss function L of the migration network for adaptive conditional distribution difference is built l ; The specific process is as follows: Loss function L of the migration network oriented to adaptive conditional distribution difference l As shown in equation (17): Wherein, represents the domain discriminative cross-entropy loss function for the c-th degradation stage; a domain discriminator mapping relationship representing the cth degradation stage; representative sample x k degradation phase prediction label; then represents the domain tag from the cth degradation stage; f SAMGRU represents the mapping relationship constructed by the gating recurrent unit based on the self-attention mechanism in step 1; Formula (17) is expressed as formula (18) as follows: Wherein, Yk,c represents that the sample at the kth moment belongs to the real source domain label under the cth degradation stage, Yk,c represents that the sample at the kth moment belongs to the real target domain label under the cth degradation stage; Yk,c represents the sample at the kth moment belongs to the predicted source domain label under the cth degradation stage, Yk,c represents the sample at the kth moment belongs to the predicted target domain label under the cth degradation stage; Lc= Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ Lc+ L 6. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 5, characterized in that: The loss function L of the migration network based on the adaptive edge distribution difference in step four g The loss function L of the migration network based on the adaptive condition distribution difference l A dynamic weighting factor w is designed; the specific process is as follows: Domain distance d oriented to adaptive edge distribution difference A,g (D S ,D T ) is calculated in the form shown by equation (19): Domain distance d oriented to adaptive conditional distribution difference A,l (D S ,D T ) is calculated in the form described by equation (20): wherein, represents a sample from the cth degradation stage of the source domain, represents a sample from the cth degradation stage of the target domain; According to d A,g (D S ,D T ) and The dynamic weighting factor w is designed as shown in equation (21):

7. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 6, characterized in that: The loss function L of the migration network based on adaptive edge distribution difference in step five g The migration network L based on adaptive condition distribution difference l And a dynamic weighting factor w, the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function of the dynamic weighting based on the overall loss function The overall loss function based on the dynamic weighted adversarial migration network is to realize the difference L between the source domain real label and the source domain predicted label y minimization and field discriminator loss L d maximization; specifically expressed in the form as described in equation (22): Wherein, L y denotes the difference between the source domain true label and the source domain predicted label; L d represents the domain discriminator loss; θ SAMGRU denotes a parameter of a gated recurrent unit network based on a self-attention mechanism; θ d and θ y denote the parameters of the field discriminator neural network and the regression sub-neural network, respectively; f y represents a mapping function of the regression sub-neural network; λ represents a weight coefficient in the loss function; f SAMGRU represents the mapping relationship constructed by the gating recurrent unit based on the self-attention mechanism in step 1; RUL k indicates the remaining useful life of a sample x k ; f d a mapping relationship representing the domain discriminator; d k represents a domain tag; L g denotes the cross-entropy loss function; L l denotes the loss function of the entire transfer network oriented to the difference of adaptive conditional distribution Parameter updating of the dynamic weighted adversarial transfer network is shown in formulas (23)-(25): Wherein, η represents a learning rate of the neural network; By minimizing a defined overall loss function L(θ SAMGRU ,θ d ,θ y ), a gated recurrent unit network based on a self-attention mechanism, a domain discriminator sub-neural network and a regression sub-neural network are optimized.

8. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 7, characterized in that: a mapping function f of the regression sub-neural network y (f SAMGRU (x k )) is as follows: f SAMGRU (x k ) are sequentially input into a first full connection layer, a first ReLU activation function layer, a second full connection layer, and a second ReLU activation function layer, and the second ReLU activation function layer outputs the remaining service life prediction value; Mapping relationship f of domain discriminator d (f SAMGRU (x k ) of the specific operation process is: f SAMGRU (x k ) are sequentially input into a third fully connected layer, a third ReLU activation function layer, a fourth fully connected layer, a fourth ReLU activation function layer, and the fourth ReLU activation function layer outputs a probability value.

9. The aero-engine off-design remaining useful life prediction method based on dynamic weighted adversarial migration network according to claim 8, characterized in that: The step six of the total loss function of the dynamic weighted adversarial transfer network completes offline training of the dynamic weighted adversarial transfer network, and a trained dynamic weighted adversarial transfer network is obtained; the specific process is as follows: 1) Sensor data collected under source domain and target domain conditions are combined in the form of a time window to form time window data from the source domain and the target domain; 2) The time window data from the source domain and the target domain are respectively input into the gated recurrent unit of the source domain and the gated recurrent unit of the target domain to obtain the output of the gated recurrent unit of the source domain and the output of the gated recurrent unit of the target domain; 3) Based on the output of the gated recurrent unit of the source domain and the output of the gated recurrent unit of the target domain obtained in 2), the global and local loss functions of the adaptive edge distribution difference are calculated through step 2, and the edge distribution difference loss is obtained by adding the two kinds of losses; Based on the output of the gated recurrent unit of the source domain and the output of the gated recurrent unit of the target domain obtained in 2), the adaptive conditional distribution difference loss is calculated through step 3; 4) Based on the adaptive edge distribution difference loss and the adaptive conditional distribution difference loss, the dynamic weighting factor is calculated through step 4; 5) Based on 3) and 4), the total loss function of the dynamic weighted adversarial transfer network is calculated through step 5, and the parameters of the dynamic weighted adversarial transfer network are updated; 6), repeat 1), 2), 3), 4), 5), until the overall loss function based on dynamic weighting adversarial migration network converges, complete the offline training of the dynamic weighting adversarial migration network, and obtain the trained dynamic weighting adversarial migration network.

10. The aero-engine off-design remaining useful life prediction method based on dynamic weighting adversarial migration network according to claim 9, characterized in that: The online data belonging to the target domain in the step seven is input into the trained dynamic weighting adversarial migration network in the step six, and the online prediction of the remaining useful life of the aero-engine is realized; the specific process is: The online aero-engine data belonging to the target domain is divided according to the fixed-length time window to obtain the time sequence feature sequence X; The time sequence feature sequence X is input into a self-attention mechanism layer, and the self-attention mechanism layer outputs outputting the self-attention mechanism layer inputting a gating recurrent unit, and outputting h; The gated recurrent unit output h inputs the regression sub-neural network, and the regression sub-neural network outputs the remaining useful life prediction value; the process is: The gated recurrent unit output h is sequentially input into the first full connection layer, the first ReLU activation function layer, the second full connection layer, the second ReLU activation function layer, and the second ReLU activation function layer outputs the remaining useful life prediction value.

Images