Remaining useful life prediction method based on variational local weighted subdomain adaptive network

Abstract

The present invention relates to a remaining service life prediction method based on a variational local weighted subdomain adaptive network. The purpose of the present invention is to solve the problem of reduced remaining service life prediction accuracy caused by the confusion of fine-grained features between subdomains in the global domain adaptive method in transfer learning. The process is: 1. Data preprocessing to obtain a training data set and a verification data set; 2. Building a variational local weighted subdomain adaptive network; 3. Training the variational local weighted subdomain adaptive network based on the training set to obtain a trained variational local weighted subdomain adaptive network; 4. Online prediction of target domain data, performing data preprocessing of step 1 on the target domain data, inputting the preprocessed target domain data into the trained variational local weighted subdomain adaptive network, and outputting the remaining service life. The present invention is used in the interdisciplinary field of combining artificial intelligence technology with fault prediction in an industrial context.

Description

Technical Field

[0001] The present invention relates to the interdisciplinary field of combining artificial intelligence technology with fault prediction in an industrial context, and in particular to the use of deep learning methods based on transfer learning to solve the problem of predicting the remaining useful life of complex systems. Background Art

[0002] With the advent of Industry 4.0 and the continuous improvement in performance and increasing complexity of large and complex equipment, prognostics and health management (PHM) technology is gaining increasing application in fields such as aviation, navigation, atomic energy, and structural engineering. As a key component of PHM technology, remaining useful life (RUL) prediction holds significant research value. Accurate RUL prediction effectively assesses system health, assists users in developing appropriate maintenance plans, and significantly optimizes overall system maintenance strategies.

[0003] Existing RUS prediction methods can be roughly divided into model-based methods and data-driven methods. Due to the complexity of the research object and failure modes, it is difficult to establish a suitable model, and the model-based method is greatly limited. On the other hand, data-driven RUS prediction methods, which do not require any prior knowledge of the object system, have received increasing attention. The essence of data-driven RUS prediction methods is to construct a mapping relationship between system degradation data and RUS through a machine learning algorithm model. Deep learning methods that have emerged in recent years, such as convolutional neural networks (CNN), recurrent neural networks (RNN) and their variants, long short-term memory networks (LSTM), have expanded the scope of data-driven RUS methods.

[0004] It's worth noting that the aforementioned supervised data-driven learning methods all assume that offline training data and online test data must come from the same distribution. However, this assumption is difficult to meet in real-world RLS prediction tasks. This is because different devices operate under varying operating conditions, which leads to differences in data characteristics. Furthermore, different devices may exhibit different feature distributions at the degradation data level due to varying failure modes. Consequently, domain distribution differences between training and test data are inevitable. In recent years, transfer learning (TL), guided by domain adaptation (DA), has been widely applied to cross-domain RLS prediction tasks. Cross-domain RLS prediction uses labeled data from the source domain for training and a small number of unlabeled samples from the target domain to extract domain-invariant representations between the source and target domains, thereby achieving RLS prediction in the target domain. It's important to note that traditional transfer learning methods typically achieve global alignment of the source and target domains. However, complex systems exhibit different distributions in their data representations at different degradation stages during the degradation process. Therefore, solely using global domain adaptation methods will confuse the fine-grained features between subdomains representing different degradation stages, thereby reducing the prediction effect of the remaining useful life of complex systems. Further research is needed to extract time-dependent features from complex system degradation data and combine them with deep subdomain adaptation networks. Summary of the Invention

[0005] The purpose of the present invention is to solve the problem of reduced accuracy of remaining useful life prediction caused by the confusion of fine-grained features between subdomains in the global domain adaptive method in transfer learning, and to propose a remaining useful life prediction method based on a variational local weighted subdomain adaptive network.

[0006] The specific process of the remaining service life prediction method based on the variational local weighted subdomain adaptive network is as follows:

[0007] Step 1: Data preprocessing to obtain training data set and verification data set;

[0008] Step 2: Build a variational local weighted subdomain adaptive network; the specific process is:

[0009] Step 21: Build a variational autoencoder-long short-term memory subnetwork;

[0010] Step 22: Based on the variational autoencoder-long short-term memory subnetwork, design soft labels of different categories;

[0011] Step 2 and 3: Build a local weighted deep sub-domain adaptive sub-network based on soft labels;

[0012] Step 3: training a variational local weighted subdomain adaptive network based on the training set to obtain a trained variational local weighted subdomain adaptive network;

[0013] Step 4: Predict the target domain data online. Perform data preprocessing as in step 1 on the target domain data. Input the preprocessed target domain data into the trained variational local weighted subdomain adaptive network to output the remaining useful life.

[0014] Beneficial effects of the present invention

[0015] This paper proposes a novel method based on variational local weighted subdomain adaptive network to predict the remaining useful life of complex systems, in order to help users effectively evaluate the health status of the system and reasonably formulate maintenance plans, thereby significantly optimizing the maintenance strategy of the entire system.

[0016] The present invention aims to reduce the domain distribution differences between offline and online data in the task of remaining useful life prediction, thereby achieving unsupervised prediction of the target domain's remaining useful life. Specifically, the present invention compresses and maps the degradation data of complex systems into an interpretable latent space to describe the domain differences between the source domain and the target domain. At the same time, the present invention also aims to adaptively align subdomains of the source and target domains, capturing fine-grained features between subdomains characterized by different degradation stages, thereby improving the prediction of the remaining useful life of complex systems.

[0017] (1) Traditional supervised data-driven learning methods have difficulty reducing the domain distribution differences between offline and online data. To this end, without the need for labeled data in the target domain, the present invention can achieve unsupervised RSU prediction of target domain degradation data based on labeled source domain data.

[0018] (2) This paper introduces a variational autoencoder-long short-term memory network (LSTM) with good interpretability in the proposed method. It can compress and map the input data into an interpretable latent space that describes the data distribution and construct a mapping relationship between the latent space and the remaining useful life. At the same time, the interpretable latent space can well describe the domain difference between the source and target domains.

[0019] (3) Different from the traditional global alignment method, this paper proposes a local weighted deep subdomain adaptive network to capture the fine-grained information of complex systems at different degradation stages, so as to better learn the domain-invariant features between the source domain and the target domain and improve the remaining useful life prediction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a workflow diagram of the present invention;

[0021] Figure 2The variational autoencoder-long short-term memory subnetwork structure diagram in the present invention; where f FCE and g LSTM Represent the mapping relationship between the encoder's fully connected layer and the LSTM network, where f FCD Indicates the mapping relationship of the decoder.

[0022] Figure 3 This is a diagram of the structural unit of the long short-term memory network in the present invention;

[0023] Figure 4 A comparison chart of the actual value of the remaining service life of the present invention and the corresponding soft label;

[0024] Figure 5 This is a structural diagram of the variational local weighted subdomain adaptive network of the present invention;

[0025] Figure 6a The remaining useful life prediction results for tasks OC1-OC2;

[0026] Figure 6b The remaining useful life prediction results for task OC2-OC1. DETAILED DESCRIPTION

[0027] Specific implementation method 1: The specific process of the remaining useful life prediction method based on the variational local weighted subdomain adaptive network in this implementation is as follows:

[0028] The remaining useful life of complex systems is predicted using a variational autoencoder-long short-term memory network-local weighted deep subdomain adaptive network. The main steps are:

[0029] Step 1: Data preprocessing to obtain training data set and verification data set;

[0030] Step 2: Build a variational local weighted subdomain adaptive network; the specific process is:

[0031] Step 21: Build a variational autoencoder-long short-term memory subnetwork;

[0032] Step 22: Based on the variational autoencoder-long short-term memory subnetwork, design soft labels of different categories;

[0033] Step 2 and 3: Build a local weighted deep sub-domain adaptive sub-network based on soft labels;

[0034] Step 3: training a variational local weighted subdomain adaptive network based on the training set to obtain a trained variational local weighted subdomain adaptive network;

[0035] Step 4: Predict the target domain data online. Perform data preprocessing as in step 1 on the target domain data. Input the preprocessed target domain data into the trained variational local weighted subdomain adaptive network to output the remaining useful life.

[0036] Specific embodiment 2: This embodiment differs from specific embodiment 1 in that the data is pre-processed in step 1; the specific process is as follows:

[0037] Step 11: Collect historical sensor data of the industrial system (multi-working-condition industrial system) during operation as source domain data and target domain data;

[0038] Step 1 and 2: Remove constant features with constant values ​​in the sensor's historical data to complete feature screening, so as to prevent the negative impact of redundant data on the remaining service life prediction and achieve data dimensionality reduction;

[0039] Step 13: Combine the source domain data that have been feature-filtered in steps 1 and 2 in the form of a time window (a period of time). The combined time window data is the source domain dataset (the source domain dataset is a labeled dataset).

[0040] The target domain data that have been feature-filtered in steps one and two are combined in the form of a time window (a period of time). The combined time window data is the target domain dataset (the target domain dataset is an unlabeled dataset).

[0041] The dataset is obtained by taking the labeled source domain dataset and the unlabeled target domain dataset, and 80% of the dataset is divided into the training dataset and 20% of the dataset is used as the validation dataset.

[0042] Other steps and parameters are the same as those in the first embodiment.

[0043] Specific embodiment three: This embodiment differs from specific embodiment one or two in that a variational autoencoder-long short-term memory subnetwork is constructed in step 21; the specific process is as follows:

[0044] The variational autoencoder-long short-term memory subnetwork includes an encoder layer, a re-parameter layer, and a decoder layer;

[0045] The mapping relationship g of the input data through the long short-term memory subnetwork LSTM LSTM , the fully connected layer mapping relationship of the encoder f FCE , the fully connected layer parameters of the encoder Ξ FCE (weights and biases) and the network weight parameter Ξ of the LSTM network LSTM , get the mean μ and standard deviation σ;

[0046] Based on the mean μ and standard deviation σ, we can get the potential low-dimensional space z of the data;

[0047] Based on the potential low-dimensional space z of the data, the mapping relationship of the decoder and the network weight parameters of the fully connected layer of the decoder, the predicted value of the remaining service life RUL is obtained. pi .

[0048] The variational autoencoder-long short-term memory subnetwork in the present invention includes two parts: the variational autoencoder module and the long short-term memory network module. Its structure diagram is as follows: Figure 2 shown.

[0049] Other steps and parameters are the same as those in the first or second embodiment.

[0050] Specific embodiment 4: This embodiment is different from any one of the specific embodiments 1 to 3 in that the long short-term memory sub-network is composed of an input gate i t 、Forget Gate t , output gate o t and memory module c t These 4 parts are composed;

[0051] LSTM determines the information that needs to be input through the input gate, the forget gate selectively forgets the secondary information, and the output gate determines the proportion of input information that can be used for the final output. The memory module determines the memory value of the long and short-term memory network unit at the previous moment. The final hidden layer output is the product of the output value of the output gate and the output value of the memory module. The mathematical expression of the long and short-term memory subnetwork can be expressed as shown in Equations (1) to (5).

[0052] f t =σ(U f [h t-1 ,x t ]+b f ) (1)

[0053] i t =σ(U i [h t-1 ,x t ]+b i ) (2)

[0054] c t =f t *c t-1 +i t *tanh(U c [h t-1 ,x t ]+b c ) (3)

[0055] o t =σ(W o [h t-1 ,x t ]+bo ) (4)

[0056] h t =o t *tanh(c t ) (5)

[0057] Among them, x t represents the input vector of LSTM at time t, h t-1 and h t Represent the hidden layer states of LSTM at time t-1 and t, respectively, c t-1 represents the output of the memory module at time t-1, f t 、i t 、o t 、c t Respectively represent the output of the forget gate, input gate, output gate and memory module at time t, U f and b f Represent the weight and bias vector of the forget gate, U i and b i Represent the weight and bias vector of the input gate, U c and b c Represent the weight and bias vector of the memory module respectively, W o and b o Represent the weight and bias vector of the output gate respectively, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent function, and * represents the product;

[0058] The state value of the LSTM hidden layer is jointly determined by the forget gate, input gate, output gate and memory module, which can effectively avoid problems such as gradient vanishing and gradient explosion caused by repeated multiplication of the weight matrix.

[0059] As a recurrent neural network that can mine the long-term time-dependent information of data during system degradation, the long short-term memory network controls the transmission and control of information through its unique gating structure. Its structural units are as follows: Figure 3 shown.

[0060] The other steps and parameters are the same as those in the first to third embodiments.

[0061] Specific embodiment 5: This embodiment differs from any one of specific embodiments 1 to 4 in that the variational autoencoder-long short-term memory subnetwork includes an encoder layer, a re-parameter layer, and a decoder layer;

[0062] The mapping relationship g of the input data through the long short-term memory subnetwork LSTM LSTM , the fully connected layer mapping relationship of the encoder f FCE , the fully connected layer parameters of the encoder Ξ FCE(weights and biases) and the network weight parameter Ξ of the LSTM subnetwork LSTM , get the mean μ and standard deviation σ;

[0063] Based on the mean μ and standard deviation σ, we can get the potential low-dimensional space z of the data;

[0064] Based on the potential low-dimensional space z of the data, the mapping relationship of the decoder and the network weight parameters of the fully connected layer of the decoder, the predicted value of the remaining service life RUL is obtained. pi ;

[0065] The specific process is:

[0066] As an important generative network model, the Variational Auto-Encoder (VAE) aims to compress the input data and map it to a latent space that can describe the data distribution through variational inference, and obtain the network output by demapping, in which the output result corresponds to the mean and variance of the target distribution.

[0067] Specifically, suppose the input data x i It is generated by a two-step random process. First, according to the prior distribution Generate latent variable z i , then the conditional distribution Generate data x i The optimal parameter θ * It can be solved by the maximum likelihood estimation shown in formula (6):

[0068]

[0069] where θ * is the optimal parameter matrix, θ is the parameter matrix, N is the number of input data, i is the sample number, x i is the i-th input sample data. Due to the difficulties of traditional sampling estimation methods, the core of the variational autoencoder is to find the posterior distribution Q of the potential low-dimensional space z of the generated data x by constructing a neural network. φ (z|x). To this end, the variational autoencoder performs variational inference to maximize the objective function The original objective function is indirectly optimized by a lower bound method, as shown in formula (7):

[0070]

[0071] The first term is the log-likelihood estimate of the sample under the posterior probability, the second term is the KL divergence used to measure the difference between the approximate posterior distribution and the standard Gaussian distribution, z is the potential low-dimensional space of the data, Q φ(z|x) is the posterior distribution of the potential low-dimensional space z that generated the data x, is the objective function, KL is the Kullback–Leibler divergence loss function, KL(Q φ (z|x i )||P θ (z|x i )) is the distribution Q φ (z|x i ) and distribution P θ (z|x i ), the KL divergence between z～Q is the mathematical expectation under the condition that the potential low-dimensional space z obeys the distribution Q. In order to solve the problem that the model cannot implement the backpropagation algorithm due to random sampling, VAE introduces the re-parameterization technique, as shown in the following formula:

[0072] z i =μ(x i ;φ)+σ(x i ;φ)×ε

[0073] where μ(x i ;φ) is the mean of the distribution, σ(x i ; φ) is the standard deviation of the distribution, ε is the constant number of reparameter sampling, z i is the low-dimensional space corresponding to the i-th data sample. The encoder learns the distribution of the input data and maps the input data to the representation distribution mean μ(x i ;φ) and standard deviation σ(x i ; φ), and sample ε from the standard normal distribution to generate the latent variable z i On this basis, the variational autoencoder-long short-term memory network proposed in this paper has the following structure: Figure 2 As shown. The encoder based on the long short-term memory network learns the temporal dependency of the data and compresses the space x represented by the input time window data into the potential interpretable latent variable space z, which is the mean μ(x i ;φ) and standard deviation σ(x i ; φ), can be expressed as shown in formula (9); then, the interpretable latent variable z can be calculated by the above formula, and the mapping relationship from the potential latent variable to the remaining service life can be realized by the decoder shown in the following formula:

[0074] RUL pi =f FCD (z i ;Ξ FCD )

[0075] 1) The mapping relationship g of input data through the long short-term memory subnetwork LSTM LSTM, the fully connected layer mapping relationship of the encoder f FCE , the fully connected layer parameters of the encoder Ξ FCE and the network weight parameter Ξ of the long short-term memory subnetwork LSTM , get the mean μ and standard deviation σ;

[0076]

[0077] where f FCE Represents the mapping relationship of the encoder's fully connected layer, g LSTM Represents the mapping relationship of the long short-term memory subnetwork LSTM network, Ξ FCE represents the fully connected layer parameters of the encoder, LSTM represents the network weight parameter of the long short-term memory subnetwork; x i represents the i-th input sample data; φ represents the input sample data x i Generate a latent low-dimensional space z i Model parameters of

[0078] 2) Get the potential low-dimensional space z of the data based on the mean μ and standard deviation σ i ;

[0079] z i =μ(x i ;φ)+σ(x i ;φ)×ε (9)

[0080] where μ(x i ;φ) is the mean of the distribution, σ(x i ; φ) is the standard deviation of the distribution, ε is the constant number of reparameter sampling, z i is the potential low-dimensional space of the i-th input data sample;

[0081] The encoder learns the distribution of the input data and maps the input data to the representation distribution mean μ(x i ;φ) and standard deviation σ(x i ; φ), and sample ε from the standard normal distribution to generate the latent variable z i On this basis, the variational autoencoder-long short-term memory network proposed in this paper has the following structure: Figure 2 As shown;

[0082] 3) Potential low-dimensional space z based on data i , the mapping relationship of the decoder and the network weight parameters of the fully connected layer of the decoder to obtain the predicted value of the remaining service life RUL pi ;

[0083] RUL pi =f FCD (z i;Ξ FCD ) (10)

[0084] where f FCD Denotes the mapping relationship of the decoder, FCD Represents the network weight parameters of the fully connected layer of the decoder, RUL pi is the estimated value of the remaining useful life.

[0085] In this process, the Kullback–Leibler divergence loss function, which measures the difference in probability distribution, is imposed to limit the loss.

[0086] The other steps and parameters are the same as those in the first to fourth embodiments.

[0087] Specific embodiment 6: This embodiment differs from any one of specific embodiments 1 to 5 in that the loss function of the variational autoencoder-long short-term memory subnetwork can be expressed as shown in formula (11):

[0088]

[0089] Among them L VAE-LSTM is the loss function of the variational autoencoder-long short-term memory subnetwork, the first term is the squared error term between the true and predicted values ​​of the remaining useful life, and the second It is a latent space regularization constraint solved by variational inference, RUL ti is the true value of the remaining useful life of the i-th sample; KL(Q φ (z i |x i )||P θ (z i )) is the distribution Q φ (z i |x i ) and distribution P θ (z i ), Q φ (z i |x i ) is the i-th input sample data x i The potential low-dimensional space z i The posterior distribution of θ (z i ) is used to generate input sample data x i The potential low-dimensional space z i The prior distribution of , (||) represents the sum operator in KL divergence, KL is the divergence, N is the number of input data, and i is the sample number.

[0090] The other steps and parameters are the same as those in the first to fifth embodiments.

[0091] Specific embodiment seven: This embodiment differs from specific embodiments one to six in that, in step 22, soft labels of different categories are designed based on the variational autoencoder-long short-term memory subnetwork; the specific process is:

[0092] For deep subdomain adaptive networks, aligning the source and target domains in subdomain adaptation requires a finite number of subdomains, which allows for a clear description of conditional probabilities. However, the RLS prediction problem is essentially a regression problem with an infinite number of labels, which poses a significant challenge in measuring the differences between conditional distributions. Considering the inherent similarities between different stages of the system degradation process and the infinite number of RLS labels, this paper proposes a discretization strategy that discretizes the entire system degradation process into multiple degradation stages.

[0093] Specifically, the present invention assumes that the maximum limit value of the remaining service life is The;

[0094] For the mth source domain sample The label directly uses the true value of the remaining service life RUL i Calculating soft labels It is expressed as shown in formula (12):

[0095]

[0096] Where round represents the rounding function, RUL ti is the true value of the remaining useful life of the i-th sample, and C is the total number of categories;

[0097] in Represents the soft label vector in the source domain sample The cth element of , c = 1, 2, ..., C;

[0098] For target domain samples, soft labels are calculated based on the remaining service life prediction value.

[0099]

[0100] in Represents the soft label vector l in the target domain data j T The cth element of ;

[0101] Taking the aircraft turbofan engine dataset in the embodiment as an example, The is set to 125, C is set to 10, and the comparison relationship between the actual value of the remaining service life and the corresponding soft label is as follows: Figure 4In summary, the present invention divides the complex system into degradation stages and designs soft labels, thereby reducing the distance between the source domain and the target domain.

[0102] The other steps and parameters are the same as those in the first to sixth embodiments.

[0103] Specific embodiment eight: This embodiment differs from any one of specific embodiments one to seven in that in steps two and three, a local weighted deep sub-domain adaptive sub-network is built based on soft labels; the specific process is as follows:

[0104] The essence of the deep subdomain adaptive network proposed in this paper is an unsupervised domain adaptation method. Assume that the source domain data sample D s It can be formally expressed as: D s ={(x m s ,y m s )|m=1,2,…,n s}, where x m s and y m s They represent the feature variables and corresponding remaining useful life labels represented by the mth source domain sample; the target domain data sample D T It can be formally expressed as: D T ={(x j T ,y j T )|j=1,2,…,n T}, where x j T and y j T They represent the feature variables represented by the j-th target domain sample. Traditional domain adaptation methods usually use the Maximum Mean Discrepancy (MMD) to measure the difference between the source domain data distribution p and the target domain data distribution q, as shown in formula (13):

[0105]

[0106] Among them H k represents the Reproducing Kernel Hilbert Space (RKHS) defined by the salient kernel k, x s and x T Represents the source domain D s and target domain D T The data in , φ defines the mapping relationship from the original data to RKHS, and k means the kernel function relationship k(xs ,x T )=<φ(x s ),φ(x T )>, <.> means inner product operation, p is the source domain data distribution, and q is the target domain data distribution.

[0107] It is worth mentioning that the present invention considers the intrinsic connection between different stages in the system degradation process and introduces a local weighted maximum mean discrepancy (LWMMD) to measure the distance between relevant subdomains in the source domain and the target domain, which can be expressed as shown in formula (14):

[0108]

[0109] in The Local Weighted Maximum Mean Discrepancy (LWMMD) measures the distance between the relevant subdomains in the source and target domains, E c represents the conditional expectation under the subdomain partitioning condition, represents the conditional expectation of samples whose source domain data belongs to category c, represents the conditional expectation of the sample whose target domain data belongs to category c, H k Denotes the Reproducing Kernel Hilbert Space (RKHS) defined by the saliency kernel k, in which the average embedding of the probability distribution p can be expressed as u k (p), then u k (p) is the space H k A unique element in the space H k Any function f∈H in k , there exists the expression in formula (15):

[0110]

[0111] x s and x T Represents the source domain D s and target domain D T The data in φ defines the mapping relationship from the original data to RKHS, p (c) and q (c) They represent the source domain D belonging to category c. s (c) and target domain D T (c) Data distribution;

[0112] Specifically, each sample belonging to each category will be assigned a different weight. The local weighted maximum mean difference in formula (14) measures the distance between the relevant subdomains in the source domain and the target domain, which can be further rewritten as shown in formula (16):

[0113]

[0114] in and Respectively represent the mth source domain sample belonging to the cth category and the jth target domain sample The weight value, C is the total number of categories, ||.|| represents the two norm;

[0115] It is worth mentioning that and Satisfy the conditions described in formula (17):

[0116]

[0117] On this basis, for the mth source domain sample and the jth target domain sample The weight value, weight and The expressions of are shown in formulas (18)-(19):

[0118]

[0119]

[0120] in Represents the soft label vector in the source domain data The cth category of ; Represents the soft label vector l in the target domain data j T The cth category of ;

[0121] It should be noted that the soft label vector l i It is a one-hot encoding vector of the discretized degradation phase.

[0122] When the sample x i From the source domain data D s When the soft label vector l i It can be calculated directly by the method described in formula (12);

[0123] However, the method proposed in this invention is essentially unsupervised domain adaptation. i From the target domain data D T When the soft label vector li It is necessary to calculate the predicted value of the remaining useful life through the forward process of the variational autoencoder-long short-term memory subnetwork in formulas (8)-(10), and then calculate it according to the method in formula (12).

[0124] Through the above description, the weight value of each source domain sample and target domain sample can be obtained, and then the calculation of the local weighted maximum mean difference of the samples can be realized. If the local domain adaptation of the source domain and the target domain is to be achieved, a very important link is the design of the local weighted maximum mean difference in formula (16). Considering the representation distribution mean μ(x i ;φ) and standard deviation σ(x i φ) can well describe the different stages of the system degradation process, with special feature representation and good interpretability. Therefore, using the interpretable latent space spanned by the corresponding mean and variance as the input of formula (16) can well achieve domain adaptation within different degradation stages.

[0125] To this end, the local weighted maximum mean difference loss function in formula (16) can be expanded to the form shown in formula (20):

[0126]

[0127] Where γ is the bandwidth of the Gaussian RBF kernel function, n S is the number of source domain data samples, n T is the number of target domain data samples, and Respectively represent the mth source domain sample belonging to the cth category and the jth source domain sample x j T The weight value of and Respectively represent the mth source domain sample belonging to the cth category and the jth target domain sample The weight value of is the distribution mean of the jth sample in the source domain under the condition of the generative model parameter φ, To characterize the distribution standard deviation of the j-th sample in the source domain under the condition of the generative model parameter φ, To characterize the distribution standard deviation of the j-th sample in the source domain under the condition of the generative model parameter φ, To represent the distribution mean of the mth sample in the target domain under the condition of the generative model parameter φ, It is the standard deviation of the distribution of the mth sample in the target domain under the condition of the generative model parameter φ;

[0128] After obtaining the local weighted maximum mean difference between the source domain data and the target domain data, the content in formula (20) can be minimized to achieve alignment as the difference between the various subdomains at different degradation stages.

[0129] Therefore, combined with formula (11), the objective loss function of the entire variational local weighted subdomain adaptive network can be defined as described in formula (21):

[0130]

[0131] where λ is the weighted trade-off coefficient between the variational autoencoder-long short-term memory subnetwork and the local weighted maximum mean difference loss.

[0132] By designing the loss function as described in formula (20), it is possible to take into account both the remaining useful life prediction accuracy and the local weighted distance between the source domain and the target domain data, thereby achieving a good remaining useful life prediction without the need for the target domain remaining useful life label. Figure 5 The structure of the variational local weighted subdomain adaptive network of the present invention is shown. The loss function of the entire variational local weighted subdomain adaptive network is Figure 5 The sum of the 3 losses.

[0133] The other steps and parameters are the same as those in the first to seventh embodiments.

[0134] Specific embodiment 9: This embodiment differs from any one of specific embodiments 1 to 8 in that the value of the total number of categories C is determined by the training process of the variational local weighted subdomain adaptive network constructed in step 2; the specific process is:

[0135] During the training process of the variational local weighted subdomain adaptive network, the total number of categories C is set as a hyperparameter. The value of C is traversed through grid search, and the total number of categories C with the best prediction effect of the network's remaining service life in the validation dataset is selected as the total number of divided categories.

[0136] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0137] Specific embodiment ten: This embodiment differs from any one of specific embodiments one to nine in that, in step three, a variational local weighted subdomain adaptive network is trained based on a training set to obtain a trained variational local weighted subdomain adaptive network; the specific process is as follows:

[0138] The training set is used as the input of the variational local weighted subdomain adaptive network, and the actual remaining service life of the industrial system at the corresponding moment is used as the output of the variational local weighted subdomain adaptive network. The variational local weighted subdomain adaptive network is trained to obtain a trained variational local weighted subdomain adaptive network.

[0139] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0140] Train the variational local weighted subdomain adaptive network; the specific process is:

[0141] Input the training data set that has undergone data preprocessing in step 1 into the variational local weighted subdomain adaptive network built in step 2, train the network, and construct a mapping relationship between the input sensor data and the output remaining service life;

[0142] The offline historical data of a certain working condition is used as the source domain data and the unlabeled historical data of another working condition is used as the target domain data to train the model;

[0143] The target domain data corresponding to the working conditions is used as online data for the final prediction. During this process, 80% of the source domain data is selected as the training dataset, and 20% of the source domain data is selected as the validation dataset. The network model with the best training effect is selected and saved as the final prediction model. The loss function for network training is the mean square error loss function, the optimization algorithm is the Adam optimization algorithm, and the learning rate is 0.001. The network training process is carried out in the hardware environment of a single GPU (GTX 3080Ti graphics card).

[0144] Online prediction of remaining service life; the specific process is:

[0145] The data of the working condition corresponding to the target domain is used as online data. After data preprocessing in step 1, it is input into the variational local weighted subdomain adaptive network trained in step 5. This can realize the real-time prediction of the online remaining service life under cross-domain conditions.

[0146] The present invention uses the root mean square error (RMSE) indicator to evaluate the effectiveness of the proposed neural network model in predicting the remaining useful life of complex systems or equipment during online operation. The expression of the root mean square error is shown in formula (22):

[0147]

[0148] Among them, N is the number of samples of test data, i is the serial number of the sample, RUL pi and RUL ti are the predicted value and true value of the remaining useful life of the i-th sample respectively.

[0149] The following examples are used to verify the beneficial effects of the present invention:

[0150] Example 1

[0151] This paper uses the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) simulation software provided by NASA's Ames Forecasting Center to simulate degradation data for aircraft turbofan engines under multiple operating conditions. The dataset contains data from four different operating conditions. Each operating condition includes a training dataset for model training and a test dataset for model validation. The training dataset includes data from the entire lifecycle of the turbofan engine, while the test dataset contains initial degradation data over a period of time. The remaining useful life of the turbofan engine at the end of each period is used for model validation. The dataset configuration is shown in Table 1. The degradation data for each operating condition consists of a multivariate time series with 26 variables, including engine ID, operating time, operating parameters describing the operating conditions, such as altitude, Mach number, and sea surface temperature, as well as sensor data from 21 operating processes.

[0152] Table 1 Description of CMAPSS dataset

[0153]

[0154] As can be seen from Table 1, the subsets under the four operating conditions of the data set have different operating conditions and failure modes. In other words, the data distributions of these four data subsets are significantly different. Therefore, how to achieve unsupervised cross-domain migration between data under different working conditions is a great challenge. On this basis, the research task of the present invention is to use only the data under a certain working condition as labeled source domain data to train the variational local weighted subdomain adaptive network, and to make up for the domain distance between the source domain data and the unlabeled target domain data under another working condition, thereby realizing the remaining service life prediction of aircraft turbofan engines under the target domain working condition. In order to improve the experimental details, the present invention studies the remaining service life prediction tasks under a variety of cross-domain combinations, which can be summarized in the form shown in Table 2.

[0155] Table 2 Remaining useful life prediction tasks under various cross-domain combinations

[0156]

[0157] Step 1: Data Preprocessing: Feature screening was performed on a multi-condition aircraft turbofan engine dataset provided by the NASA Ames Forecasting Center. Fourteen sensor data types, numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21, were selected. This feature-screened data was then combined using a time window, and combined with the actual remaining useful life of the complex equipment at the corresponding moment to form the input and output of the dataset.

[0158] Step 2: Build a variational local weighted subdomain adaptive network; the specific process is:

[0159] Step 21: Build a variational autoencoder-long short-term memory subnetwork;

[0160] Step 22: Based on the variational autoencoder-long short-term memory subnetwork, design soft labels of different categories;

[0161] Step 23: Building a local weighted deep subdomain adaptive subnetwork based on soft labels;

[0162] Step 3: Train the variational local weighted subdomain adaptive network based on the training set to obtain a trained variational local weighted subdomain adaptive network;

[0163] 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).

[0164] Step 4: Predict the target domain data online, perform data preprocessing as in step 1 on the target domain data, input the preprocessed target domain data into the trained variational local weighted subdomain adaptive network, and output the remaining service life.

[0165] Figure 6a 、 6b The predicted and true values ​​of the remaining useful life of the aircraft turbofan engine test dataset under the OC1-OC2 task of the proposed method are shown in Table 2. Table 3 shows the comparison of the remaining useful life prediction task results of the present invention and various existing transfer learning methods under various cross-domain combinations. Figure 6a 、 6b As can be seen from Table 3, the method proposed in the present invention can effectively predict the remaining service life of complex systems or equipment under cross-domain conditions, helping users to effectively evaluate the health status of the system and reasonably formulate maintenance plans, thereby significantly optimizing the maintenance strategy of the entire system.

[0166] Table 3 Comparison of prediction results between the present invention and various existing transfer learning methods

[0167]

[0168] The present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A method for predicting remaining useful life based on a variational local weighted subdomain adaptive network, characterized by: The specific process of the method is: Step 1: Data preprocessing to obtain training data set and verification data set; Step 2: Build a variational local weighted subdomain adaptive network; the specific process is: Step 21: Build a variational autoencoder-long short-term memory subnetwork; Step 22: Based on the variational autoencoder-long short-term memory subnetwork, design soft labels of different categories; Step 2 and 3: Build a local weighted deep sub-domain adaptive sub-network based on soft labels; Step 3: training a variational local weighted subdomain adaptive network based on the training set to obtain a trained variational local weighted subdomain adaptive network; Step 4: Online prediction of target domain data: perform data preprocessing in step 1 on the target domain data, input the preprocessed target domain data into the trained variational local weighted subdomain adaptive network, and output the remaining service life; In step 21, a variational autoencoder-long short-term memory subnetwork is constructed; the specific process is as follows: The variational autoencoder-long short-term memory subnetwork includes an encoder layer, a re-parameter layer, and a decoder layer; The mapping relationship g of the input data through the long short-term memory subnetwork LSTM LSTM , the fully connected layer mapping relationship of the encoder f FCE , the fully connected layer parameters of the encoder Ξ FCE and the network weight parameter Ξ of the long short-term memory subnetwork LSTM , get the mean μ and standard deviation σ; Based on the mean μ and standard deviation σ, we can get the potential low-dimensional space z of the data; Based on the potential low-dimensional space z of the data, the mapping relationship of the decoder and the network weight parameters of the fully connected layer of the decoder, the predicted value of the remaining service life RUL is obtained. pi ; The specific process is: 1) The mapping relationship g of input data through the long short-term memory subnetwork LSTM LSTM , the fully connected layer mapping relationship of the encoder f FCE , the fully connected layer parameters of the encoder Ξ FCE and the network weight parameter Ξ of the long short-term memory subnetwork LSTM , get the mean μ and standard deviation σ; where f FCE Represents the mapping relationship of the encoder's fully connected layer, g LSTM Represents the mapping relationship of the long short-term memory subnetwork LSTM network, Ξ FCE represents the fully connected layer parameters of the encoder, LSTM represents the network weight parameter of the long short-term memory subnetwork; x i represents the i-th input sample data; φ represents the input sample data x i Generate a latent low-dimensional space z i Model parameters of 2) Based on the mean μ and standard deviation σ, we can get the potential low-dimensional space z of the data. i ; z i =μ(x i ;φ)+σ(x i ;φ)×ε(9) where μ(x i ;φ) is the mean of the distribution, σ(x i ; φ) is the standard deviation of the distribution, ε is the constant number of reparameter sampling, z i is the potential low-dimensional space of the i-th input data sample; 3) Potential low-dimensional space z based on data i , the mapping relationship of the decoder and the network weight parameters of the fully connected layer of the decoder to obtain the predicted value of the remaining service life RUL pi ; NUMBER pi =f FCD (z i ;Ξ FCD )(10) where f FCD Denotes the mapping relationship of the decoder, FCD Represents the network weight parameters of the fully connected layer of the decoder, RUL pi is the estimated value of the remaining useful life.

2. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 1, characterized in that: The data preprocessing in step 1 is as follows: Step 1: Collect historical sensor data of the industrial system during operation as source domain data and target domain data; Step 1 and 2: Eliminate constant features with constant values ​​in the sensor's historical data to complete feature screening; Step 13: Combine the source domain data that have been feature-filtered in steps 1 and 2 in the form of a time window. The combined time window data is the source domain dataset. The target domain data that have been feature-filtered in steps one and two are combined in the form of a time window. The combined time window data is the target domain dataset. The dataset is obtained by taking the labeled source domain dataset and the unlabeled target domain dataset, and 80% of the dataset is divided into the training dataset and 20% of the dataset is used as the validation dataset.

3. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 2, characterized in that: The long short-term memory sub-network consists of an input gate i t 、Forget Gate t , output gate o t and memory module c t These 4 parts are composed; The mathematical expression of the long short-term memory subnetwork can be expressed as shown in formula (1) to formula (5): f t =σ(U f [h t-1 ,x t ]+b f ) (1) i t =σ(U i [h t-1 ,x t ]+b i ) (2) c t =f t *c t-1 +i t *tanh(U c [h t-1 ,x t ]+b c ) (3) the t =σ(W o [h t-1 ,x t ]+b o ) (4) h t =o t *tanh(c t ) (5) Among them, x t represents the input vector of LSTM at time t, h t-1 and h t Represent the hidden layer states of LSTM at time t-1 and t, respectively, c t-1 represents the output of the memory module at time t-1, f t 、i t 、o t 、c t Respectively represent the output of the forget gate, input gate, output gate and memory module at time t, U f and b f Represent the weight and bias vector of the forget gate, U i and b i Represent the weight and bias vector of the input gate, U c and b c Represent the weight and bias vector of the memory module respectively, W o and b o Represent the weight and bias vector of the output gate respectively, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent function, and * represents the product.

4. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 3, characterized in that: The loss function of the variational autoencoder-long short-term memory subnetwork can be expressed as shown in formula (11): Among them L VAE-LSTM is the loss function of the variational autoencoder-long short-term memory subnetwork, the first term is the squared error term between the true and predicted values ​​of the remaining useful life, and the second It is a latent space regularization constraint solved by variational inference, RUL ti is the true value of the remaining useful life of the i-th sample; KL(Q φ (z i |x i )||P θ (z i )) is the distribution Q φ (z i |x i ) and distribution P θ (z i ), Q φ (z i |x i ) is the i-th input sample data x i The potential low-dimensional space z i The posterior distribution of θ (z i ) is used to generate input sample data x i The potential low-dimensional space z i The prior distribution of , (||) represents the sum operator in KL divergence, KL is the divergence, N is the number of input data, and i is the sample number.

5. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 4, characterized in that: In step 22, soft labels of different categories are designed based on the variational autoencoder-long short-term memory subnetwork; The specific process is: Assume that the maximum limit of the remaining useful life is The; For the mth source domain sample The label directly uses the true value of the remaining service life RUL i Calculating soft labels It is expressed as shown in formula (12): Where round represents the rounding function, RUL ti is the true value of the remaining useful life of the i-th sample, and C is the total number of categories; in Represents the soft label vector in the source domain sample The cth element of , c = 1, 2, ..., C; For target domain samples, soft labels are calculated based on the remaining service life prediction value. in Represents the soft label vector l in the target domain data j T The cth element of .

6. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 5, characterized in that: In steps 2 and 3, a local weighted depth subdomain adaptive subnetwork is built based on soft labels; the specific process is as follows: A local weighted maximum mean difference is introduced to measure the distance between the relevant subdomains in the source domain and the target domain, which can be expressed as shown in formula (14): in represents the local weighted maximum mean difference metric between the relevant subdomains in the source and target domains, E c represents the conditional expectation under the subdomain partitioning condition, represents the conditional expectation of samples whose source domain data belongs to category c, represents the conditional expectation of the sample whose target domain data belongs to category c, H k represents the reproducible kernel Hilbert space defined by the saliency kernel k; x s and x T Represents the source domain D s and target domain D T The data in φ defines the mapping relationship from the original data to RKHS, p (c) and q (c) They represent the source domain D belonging to category c. s (c) and target domain D T (c) Data distribution; The local weighted maximum mean difference in formula (14) to measure the distance between the relevant subdomains in the source domain and the target domain can be further rewritten as shown in formula (16): in and Respectively represent the mth source domain sample belonging to the cth category and the jth target domain sample The weight value, C is the total number of categories, ||.|| represents the two norm; and Satisfy the conditions described in formula (17): On this basis, for the mth source domain sample and the jth target domain sample The weight value, weight and The expressions of are shown in formulas (18)-(19): in Represents the soft label vector in the source domain data The cth category of ; Represents the soft label vector l in the target domain data j T The cth category of ; The local weighted maximum mean difference loss function in formula (16) can be expanded to the form shown in formula (20): Where γ is the bandwidth of the Gaussian RBF kernel function, n S is the number of source domain data samples, n T is the number of target domain data samples, and Respectively represent the mth source domain sample belonging to the cth category and the jth source domain sample x j T The weight value of and Respectively represent the mth source domain sample belonging to the cth category and the jth target domain sample The weight value of is the distribution mean of the jth sample in the source domain under the condition of the generative model parameter φ, To characterize the distribution standard deviation of the j-th sample in the source domain under the condition of the generative model parameter φ, To characterize the distribution standard deviation of the j-th sample in the source domain under the condition of the generative model parameter φ, To represent the distribution mean of the mth sample in the target domain under the condition of the generative model parameter φ, It is the standard deviation of the distribution of the mth sample in the target domain under the condition of the generative model parameter φ; Therefore, combined with formula (11), the objective loss function of the entire variational local weighted subdomain adaptive network can be defined as described in formula (21): where λ is the weighted trade-off coefficient between the variational autoencoder-long short-term memory subnetwork and the local weighted maximum mean difference loss.

7. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 6, characterized in that: The value of the total number of categories C is determined by the training process of the variational local weighted subdomain adaptive network constructed in step 2; the specific process is: During the training process of the variational local weighted subdomain adaptive network, the total number of categories C is set as a hyperparameter. The value of C is traversed through grid search, and the total number of categories C with the best prediction effect of the network's remaining service life in the validation dataset is selected as the total number of divided categories.

8. The method for predicting remaining useful life based on a variational local weighted subdomain adaptive network according to claim 7, characterized in that: In the step 3, the variational local weighted subdomain adaptive network is trained based on the training set to obtain a trained variational local weighted subdomain adaptive network; the specific process is: The training set is used as the input of the variational local weighted subdomain adaptive network, and the actual remaining service life of the industrial system at the corresponding moment is used as the output of the variational local weighted subdomain adaptive network. The variational local weighted subdomain adaptive network is trained to obtain a trained variational local weighted subdomain adaptive network.

Images