A life prediction method for deep-sea watertight connectors

Abstract

The present invention provides a life prediction method for deep-sea watertight connectors, which relates to the field of life prediction and specifically includes: constructing a tensor dataset; inputting the decomposed feature tensor into a multi-scale sparse neural network, extracting cross-scale degradation features through a dilated convolution module and an attention gating mechanism; fusing features using a dual-channel Transformer-CBAM network; using a causal graph autoencoder to learn invariant fault representations under deep-sea environmental disturbances, and achieving cross-depth operating condition distribution alignment through a domain adversarial network driven by a gradient reversal layer; modeling the effects of random fluctuations through a nonlinear Wiener process modulated by environmental parameters, combining Monte Carlo path simulation with Bayesian inference for confidence interval quantization, and outputting a dynamic remaining life prediction result for the watertight connector. The technical solution of the present invention overcomes the problem of the prior art in being unable to accurately predict the life of deep-sea watertight connectors.

Description

Technical Field

[0001] The present invention relates to the field of life prediction, and in particular to a life prediction method for a deep-sea watertight connector. Background Art

[0002] Deepwater seawater connectors are crucial components for power distribution, signal transmission, and communications in deepwater equipment. Their reliability directly determines the long-term operation of critical equipment, such as underwater production systems, in extreme environments. These components are subjected to the multi-field coupling effects of high pressure, high salt corrosion, temperature gradients, and dynamic mechanical loads for extended periods, resulting in highly nonlinear and non-stationary performance degradation. Existing life prediction methods face four major bottlenecks: First, the complex synergistic effects of high pressure, low temperature, corrosion, and mechanical loads make it difficult for traditional single-factor models to accurately decouple the effects of multi-physics coupling. Second, the high cost of deepwater testing leads to a scarcity of failure data, and traditional statistical models lack the generalization capability under small sample sizes. Third, material damage exhibits threshold effects and nonlinear mutation characteristics, with crack propagation exhibiting a three-stage pattern of "slow-fast-failure." Existing linear degradation models are unable to capture the performance inflection point at the end of life. Fourth, significant differences exist between laboratory environments and real-world sea conditions in terms of dynamic pressure fluctuations and bio-adhesion-corrosion coupling, necessitating improved reliability for cross-sea life prediction.

[0003] Therefore, a life prediction method for a deep sea water tight connector is needed that can accurately predict the life of the deep sea water tight connector. Summary of the Invention

[0004] The main purpose of the present invention is to provide a life prediction method for deep sea water-tight connectors, so as to solve the problem in the prior art that the life of deep sea water-tight connectors cannot be accurately predicted.

[0005] To achieve the above object, the present invention provides a life prediction method for a deep-sea watertight connector, which specifically comprises the following steps:

[0006] S1, through the multi-source sensor array in the deep-sea pressure chamber, collects temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress and dielectric loss of watertight connectors in real time to construct a tensor dataset.

[0007] S2, the standardized tensor data set is input into the parallel factor analysis PARAFAC module to decompose the time evolution factor matrix, spatial distribution factor matrix, frequency domain characteristic factor matrix and environmental coupling factor matrix.

[0008] S3, the decomposed feature tensor is input into the multi-scale sparse neural network MSNN, and cross-scale degradation features are extracted through the void convolution module and attention gating mechanism.

[0009] S4, the time-frequency features and spatial environment features are obtained through the CEEMDAN-KPCA algorithm, and the features are fused using the dual-channel Transformer-CBAM network.

[0010] S5, uses a causal graph autoencoder to learn invariant fault representations under deep-sea environmental disturbances, and achieves cross-depth working condition distribution alignment through a domain adversarial network driven by a gradient reversal layer.

[0011] S6, through the nonlinear Wiener process modeling of random fluctuations modulated by environmental parameters, combined with Monte Carlo path simulation and Bayesian inference to quantify the confidence interval, output the dynamic remaining life prediction results of the watertight connector.

[0012] Furthermore, the tensor data set in step S1 is a fourth-order tensor data set ,in, is the time dimension; is the spatial dimension; is the frequency dimension; For the environmental dimension.

[0013] Furthermore, step S2 specifically includes the following steps:

[0014] S2.1, perform Z-score standardization on the data of each dimension:

[0015] ;

[0016] in, and Respectively The mean and standard deviation of the environmental parameters, is a fourth-order tensor data set, After standardization .

[0017] The scalar form formula of the fourth-order parallel factor molecule tensor decomposition is:

[0018] ;

[0019] in, is the number of factors, It is Time factor, It is A spatial factor, It is frequency factors, It is environmental factors, represents the error set of a fourth-rank tensor.

[0020] S2.2, the normalized tensor dataset is input into the parallel factor analysis PARAFAC module to decompose 、 、 、 Four matrices, defining the time evolution factor matrix for × dimensional matrix, is the time factor; define the spatial distribution factor matrix for × dimensional matrix, is the spatial factor; define the frequency domain characteristic factor matrix for × dimensional matrix, is the frequency factor; define the environmental coupling factor matrix for × dimensional matrix, For environmental factors.

[0021] S2.3, Calculation 、 、 and , to calculate The other matrix calculation methods are the same as same:

[0022] ;

[0023] in, Indicates that 、 、 The matrix is ​​block diagonalized on the column vectors of , and the error between the observed data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows:

[0024] ;

[0025] in, represents the generalized inverse.

[0026] S2.4, Repeated Counting 、 、 and , until the result converges or the set number of iterations is reached and then stops running, and .

[0027] Furthermore, in step S3, the multi-scale features are extracted in the dilated convolution module as follows:

[0028] The result obtained in step 2 、 、 and Concatenate into feature tensors , and Input the dilated convolution module of the multi-scale sparse neural network MSNN to extract multi-scale features and output a multi-scale feature map:

[0029] ;

[0030] ;

[0031] in, is the convolution kernel size, The expansion ratio is 1, 2 and 4, For the The convolution kernel weights at each position, is the factor of the current position, is a multi-scale feature map, is the sequence length, is the number of channels.

[0032] Furthermore, in step S3, extracting features of different granularities through the attention gating mechanism specifically includes the following steps:

[0033] S3.1, the multi-scale feature map output by the dilated convolution module Cascade and assign weights through the attention gating mechanism :

[0034] ;

[0035] in, is the learnable weight matrix, is the bias term.

[0036] S3.2, fuse the weighted multi-scale features:

[0037] ;

[0038] in, is the fused multi-scale feature map.

[0039] S3.3, Input residual connection and nonlinear projection module, output fusion features :

[0040] ;

[0041] ;

[0042] in, is the eigenvector after residual connection, is the learnable projection matrix, is the column projection vector of the linear combination concatenated features, is the cascade operator, is the bias, .

[0043] Furthermore, step S4 specifically includes the following steps:

[0044] S4.1, the fusion features output by MSNN are used as input signals and decomposed into multiple intrinsic mode functions through the CEEMDAN algorithm and the residual , as shown below:

[0045] ;

[0046] in, is the characteristic number, For time.

[0047] S4.2, extract the energy of each IMF and calculate the energy entropy value :

[0048] .

[0049] S4.3, multiple IMF energies Input KPCA dimensionality reduction, through the kernel function Mapping to high-dimensional space for principal component analysis, we can obtain time-frequency characteristics and spatial environment characteristics:

[0050] ;

[0051] ;

[0052] in, is an element in the kernel matrix; represents the energy of the IMF, is the kernel matrix eigenvector, is the projection matrix, Score the principal components.

[0053] S4.4, the temporal frequency features and spatial environment features are extracted into global relationships through the multi-head attention module, and then input into the CBAM attention mechanism module.

[0054] S4.5, the overall process of CBAM is expressed as:

[0055] ;

[0056] ;

[0057] in, represents the Hadamard product, For CAM module, For the SAM module, represents the input features, Represents the feature map obtained by the CAM module, Represents the feature map obtained by the SAM module.

[0058] S4.6, 、 and Perform weighted concatenation and map it into health indicators through a fully connected layer :

[0059] ;

[0060] in, and is a learnable parameter, is the Sigmoid function, For splicing operation.

[0061] Furthermore, step S5 specifically includes the following steps:

[0062] S5.1, health indicators Combined with environmental parameters, environmental parameters include: pressure ,temperature ,salinity , construct the input feature matrix :

[0063] ;

[0064] in, , , , , after the merger .

[0065] S5.2, use the fuzzy C-means clustering algorithm to generate pseudo labels corresponding to target domain samples , the objective function of FCM for:

[0066] ;

[0067] in, represents the membership matrix, represents the cluster center, and represent the number of clusters and the sample size of the target domain, respectively. is the fuzziness indicator, for The target domain samples in .

[0068] S5.3, by introducing Lagrange multipliers Constructing Lagrangian functions ,as follows:

[0069] .

[0070] S5.4, respectively, for the membership , cluster center and Lagrange multipliers Find the partial derivative and set it to zero, and we get The cluster centers of the iteration and membership Update formula:

[0071] ;

[0072] .

[0073] Furthermore, step S5 further includes the following steps:

[0074] S5.5, the input of the causal graph autoencoder is ,in, is the source domain data with real labels, is the target domain data with pseudo labels, and the objective function of the causal representation learning module is Including: reconstruction loss, causal structure learning loss and cross entropy loss:

[0075] ;

[0076] in, is the number of samples, express The number of samples, for The reconstructed output of To control the strength of causal constraints, To balance the classification loss, is the regularized weight decay; Indicates the number of hidden layers; and denote the cross entropy loss and classifier respectively; Representation Label Merged Markov blanket; and Represents the first The weight parameters of the hidden layer; and Represent the outputs of the encoder and decoder of the causal graph autoencoder respectively:

[0077] ;

[0078] ;

[0079] in, For the dimension, Indicates mapping the original input into low-dimensional features, Indicates causal logic and label fusion, represents the decoder network, is the adjacency matrix.

[0080] S5.6, to ensure The acyclicity of the corresponding DAG must satisfy the following acyclicity constraints :

[0081] ;

[0082] in, represents the trace of the matrix, express Matrix index of ; is the Hadamard product; Indicates the number of variables in the causal graph model.

[0083] S5.7, after introducing the Lagrange multiplier and penalty term, the augmented Lagrangian is given by:

[0084] ;

[0085] in, represents the Lagrange multiplier, represents the penalty parameter.

[0086] Furthermore, step S5 further includes the following steps:

[0087] S5.8, for the adjacency matrix and The update rules are as follows:

[0088] ;

[0089] ;

[0090] ;

[0091] in, and are two tuning hyperparameters, is the parameter value when the function is minimized.

[0092] S5.9, use 、 and Representation feature extractor , label classifier and domain discriminator The network model parameters in the domain alignment module are:

[0093]

[0094] in, is the feature extractor, For label classifier and For the domain discriminator: and denote the cross entropy loss functions for fault classification and domain classification, respectively; and Represent the domain labels of source domain data and target domain data respectively; is the balance parameter, represents the gradient reversal layer, Represents the true label of the source domain data.

[0095] S5.10, parameters The update rules are as follows:

[0096] ;

[0097] ;

[0098] .

[0099] in, represents the learning rate, and Represent fault classification loss and domain discrimination loss respectively:

[0100] ;

[0101] .

[0102] Furthermore, step S6 specifically includes the following steps:

[0103] S6.1, considering the influence of environmental parameters, introduce nonlinear drift terms and diffusion terms , constructing a nonlinear Wiener process modulated by environmental parameters:

[0104] ;

[0105] Among them, the drift term To reflect the accelerating effect of environmental parameters on degradation, the generalized Arrhenius model is used:

[0106] .

[0107] Diffusion term Characterize random degradation noise caused by environmental fluctuations:

[0108] ;

[0109] in, It is a degradation index, indicating the degradation of watertight connectors over time. The cumulative degradation of is the standard Brownian motion; is the initial degradation amount; These are parameters to be determined.

[0110] S6.2, using accelerated degradation test data, the maximum likelihood estimation function Estimating model parameters :

[0111] ;

[0112] ;

[0113] in, , , is an exponential function; Respectively temperature, pressure, and salinity of each path, For the The degradation index of a path.

[0114] S6.3, using Monte Carlo path simulation to generate multiple degradation paths:

[0115] ;

[0116] in, , independent standard normal distributed random variables, introducing random fluctuations; is the time step.

[0117] S6.4, Assuming a Prior Distribution of Parameters , combined with real-time monitoring data, calculate the posterior distribution:

[0118] ;

[0119] in, is the observed degradation data, Is proportional to .

[0120] S6.5, Define Failure Threshold , calculate the remaining service life RUL:

[0121] ;

[0122] in, is the time increment, is the zth path at time The predicted degradation amount, To take the minimum.

[0123] The present invention has the following beneficial effects:

[0124] 1. This invention constructs a fourth-order temporal, spatial, and frequency environmental tensor dataset using a multi-source sensor array. Combined with parallel factor analysis (PARAFAC), it decomposes the four-dimensional factor matrix of temporal evolution, spatial distribution, frequency domain characteristics, and environmental coupling. This design enables precise decoupling of the multi-field coupling effects of high pressure, cryogenics, corrosion, and mechanical loads, resolving the inability of traditional single-factor models to handle nonlinear degradation.

[0125] 2. Targeting the nonlinear and sudden degradation characteristics of watertight connector performance, a multi-scale sparse neural network (MSNN) was designed. The MSNN module utilizes dilated convolutional layers and a multi-head attention gating mechanism to capture local details, mid-range correlations, and global degradation patterns. Residual connections and nonlinear projections preserve high-order feature interactions, significantly improving the ability to characterize crack growth in the three stages of "slow-fast-failure" and addressing the bottleneck of linear degradation models in capturing the performance inflection point at the end of life.

[0126] 3. Adaptive Intrinsic Mode Decomposition (CEEMDAN) is used to extract the IMF energy entropy of non-stationary signals, combined with kernel principal component analysis (KPCA) for dimensionality reduction. A dual-channel Transformer-CBAM network is designed. A multi-head self-attention mechanism is used in the time-frequency domain, while channel-spatial attention weighting is used in the space-frequency domain. Finally, the time-frequency and space-frequency joint features are integrated to generate a nonlinear health index (HI). This solution enables robust fusion of multi-source heterogeneous signals, addressing the generalization issues of traditional statistical models under small sample conditions.

[0127] 4. A causal graph autoencoder is introduced to learn invariant fault representations under environmental perturbations. The causal relationships between variables are modeled using an acyclic adjacency matrix (Icausal). This is combined with a domain adversarial network (DANN) driven by a gradient reversal layer to align the feature distributions of the source and target domains. This design significantly improves the reliability of cross-domain lifespan prediction, overcoming the dynamic differences between laboratory environments and real-world sea conditions.

[0128] 5. A generalized Arrhenius drift term and a linear diffusion term are constructed, random degradation trajectories are generated through Monte Carlo path simulation, and the parameter posterior distribution is updated in real time based on Bayesian inference. The resulting dynamic remaining useful life (RUL) and its 95% confidence interval are output, enabling uncertainty quantification under dynamic modulation of multiple environmental parameters, providing a highly reliable basis for deep-sea equipment maintenance decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0130] Figure 1 A flow chart showing a life prediction method for a deep-sea watertight connector according to the present invention is shown.

[0131] Figure 2 The life degradation path diagram of deep sea water tight connector is shown. DETAILED DESCRIPTION

[0132] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0133] like Figure 1 The life prediction method of a deep-sea seatight connector shown in the figure specifically includes the following steps:

[0134] S1, through the multi-source sensor array in the deep-sea pressure chamber, collects temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress and dielectric loss of watertight connectors in real time to construct a tensor dataset.

[0135] S2, the standardized tensor data set is input into the parallel factor analysis PARAFAC module to decompose the time evolution factor matrix, spatial distribution factor matrix, frequency domain characteristic factor matrix and environmental coupling factor matrix.

[0136] S3, the decomposed feature tensor is input into the multi-scale sparse neural network MSNN, and cross-scale degradation features are extracted through the void convolution module and attention gating mechanism.

[0137] S4, the time-frequency features and spatial environment features are obtained through the CEEMDAN-KPCA algorithm, and the features are fused using the dual-channel Transformer-CBAM network.

[0138] S5, uses a causal graph autoencoder to learn invariant fault representations under deep-sea environmental disturbances, and achieves cross-depth working condition distribution alignment through a domain adversarial network driven by a gradient reversal layer.

[0139] S6, through the nonlinear Wiener process modeling of random fluctuations modulated by environmental parameters, combined with Monte Carlo path simulation and Bayesian inference to quantify the confidence interval, output the dynamic remaining life prediction results of the watertight connector.

[0140] Specifically, the tensor data set in step S1 is a fourth-order tensor data set ,in, is the time dimension; is the spatial dimension (number of sensor channels); is the frequency dimension; are environmental dimensions (such as temperature, pressure, salinity, etc.).

[0141] Specifically, step S2 includes the following steps:

[0142] S2.1, perform Z-score standardization on the data of each dimension:

[0143] ;

[0144] in, and Respectively The mean and standard deviation of the environmental parameters, is a fourth-order tensor data set, After standardization .

[0145] The scalar form formula of the fourth-order parallel factor molecule tensor decomposition is:

[0146] ;

[0147] in, is the number of factors, It is Time factor, It is A spatial factor, It is frequency factors, It is environmental factors, represents the set of errors for a fourth-rank tensor.

[0148] S2.2, the normalized tensor dataset is input into the parallel factor analysis PARAFAC module to decompose 、 、 、 Four matrices, defining the time evolution factor matrix for × dimensional matrix, is the time factor; define the spatial distribution factor matrix for × dimensional matrix, is the spatial factor; define the frequency domain characteristic factor matrix for × dimensional matrix, is the frequency factor; define the environmental coupling factor matrix for × dimensional matrix, For environmental factors.

[0149] S2.3, Calculation 、 、 and , to calculate The other matrix calculation methods are the same as same:

[0150] .

[0151] in, Indicates that 、 、 The matrix is ​​block diagonalized on the column vectors of , and the error between the time observation data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows:

[0152] ;

[0153] in, represents the generalized inverse.

[0154] Secondly,

[0155] .

[0156] in, Indicates that the column vectors of A, C, and D matrices are block diagonalized and combined, and the error between the spatial observation data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows:

[0157] .

[0158] The matrix C is obtained by the following formula:

[0159] .

[0160] Indicates that the column vectors of A, B, and D matrices are block diagonalized and combined, and the error between the frequency domain observation data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows:

[0161] .

[0162] The matrix D is obtained by the following formula:

[0163] ;

[0164] in, Indicates that the column vectors of matrices A, B, and C are block diagonalized and combined, and the error between the environmental observation data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows:

[0165] .

[0166] S2.4, Repeated Counting 、 、 and , until the result converges or the set number of iterations is reached and then stops running. , which ensures the uniqueness of the parallel factorization solution.

[0167] Specifically, in step S3, the multi-scale features are extracted in the dilated convolution module as follows:

[0168] The result obtained in step 2 、 、 and Concatenate into feature tensors , and Input the dilated convolution module of the multi-scale sparse neural network MSNN to extract multi-scale features and output a multi-scale feature map:

[0169] ;

[0170] ;

[0171] in, is the convolution kernel size, The expansion ratio is 1, 2 and 4, For the The convolution kernel weights at each position, is the factor of the current position in the time dimension, space dimension, frequency dimension or environment dimension, is a multi-scale feature map, is the sequence length, is the number of channels.

[0172] Expansion rate =1, capture local features; expansion rate =2, captures mid-range features; expansion rate =4, capturing global features.

[0173] Specifically, the extraction of features of different granularities through the attention gating mechanism in step S3 includes the following steps:

[0174] S3.1, the multi-scale feature map output by the dilated convolution module Cascade and assign weights through the attention gating mechanism :

[0175] ;

[0176] in, is the learnable weight matrix (transposed form), is the bias term.

[0177] S3.2, fuse the weighted multi-scale features:

[0178] ;

[0179] in, is the fused multi-scale feature map.

[0180] S3.3, Input residual connection and nonlinear projection module, output fusion features :

[0181] ;

[0182] ;

[0183] in, is the eigenvector after residual connection, is the learnable projection matrix, is the column projection vector of the linear combination concatenated features, is the cascade operator, is the bias, .

[0184] Specifically, step S4 includes the following steps:

[0185] S4.1, the fusion features output by MSNN are used as input signals and decomposed into multiple intrinsic mode functions through the CEEMDAN algorithm and the residual , as shown below:

[0186] ;

[0187] in, is the characteristic number, For time.

[0188] S4.2, extract the energy of each IMF and calculate the energy entropy value :

[0189] .

[0190] S4.3, multiple IMF energies Input kernel principal component analysis method KPCA dimensionality reduction, through the kernel function Mapping to high-dimensional space for principal component analysis, we can obtain time-frequency characteristics and spatial environment characteristics:

[0191] ;

[0192] ;

[0193] in, is an element in the kernel matrix; represents the energy of the IMF, is the kernel matrix eigenvector, is the projection matrix, Score the principal components.

[0194] S4.4, the temporal frequency features and spatial environment features are extracted into global relationships through the multi-head attention module, and then input into the CBAM attention mechanism module.

[0195] Step S4.4 specifically includes the following steps:

[0196] Global relations are extracted through multi-head self-attention, and the attention function of a set of queries is calculated and combined into a matrix Q. At the same time, the keys and values ​​are packaged into matrices K and V to improve computational efficiency.

[0197] The principle of the multi-head self-attention module is expressed as:

[0198] ;

[0199] ;

[0200] ;

[0201] Where Q, K, and V represent query, key, and value matrices, respectively. represents the dimension of each vector in the input sequence, Represents learnable weight parameters to fuse , n represents the number of heads.

[0202] CBAM consists of two submodules: the channel attention module (CAM) and the spatial attention module (SAM). Channel attention is fixed on different channels of the feature map and assigns different weights to each channel. CAM is expressed as:

[0203] ;

[0204] in, For CAM module, is the activation function, and represents the weight matrix, and represents the features after average and maximum pooling operations, represents the input features, Represents the feature map after passing through the CAM module.

[0205] SAM is expressed as follows:

[0206] ;

[0207] in, For the SAM module, The kernel size is The convolution operation, represents a stacked feature map, Represents the feature map after passing through the SAM module.

[0208] S4.5, the overall process of CBAM is expressed as:

[0209] ;

[0210] ;

[0211] in, represents the Hadamard product, For CAM module, For the SAM module, represents the input features, Represents the feature map obtained by the CAM module, Represents the feature map obtained by the SAM module.

[0212] S4.6, 、 and Perform weighted concatenation and map it into health indicators through a fully connected layer :

[0213] ;

[0214] in, and is a learnable parameter, is the Sigmoid function, For the concatenation operation, the output is normalized to a health score in the interval (0, 1).

[0215] Specifically, step S5 includes the following steps:

[0216] S5.1, health indicators Combined with environmental parameters, environmental parameters include: pressure ,temperature ,salinity , construct the input feature matrix :

[0217] ;

[0218] in, , , , , after the merger .

[0219] S5.2, use the fuzzy C-means (FCM) clustering algorithm to generate pseudo labels corresponding to target domain samples , the objective function of FCM for:

[0220] ;

[0221] in, represents the membership matrix, represents the cluster center, and represent the number of clusters and the sample size of the target domain, respectively. is the fuzziness indicator, for The target domain samples in .

[0222] S5.3, by introducing Lagrange multipliers Constructing Lagrangian functions ,as follows:

[0223] .

[0224] S5.4, respectively, for the membership , cluster center and Lagrange multipliers Find the partial derivative and set it to zero, and we get The cluster centers of the iteration and membership Update formula:

[0225] ;

[0226] .

[0227] Based on the above update rules, by setting the maximum number of iterations or the convergence threshold, the final membership matrix can be obtained, thereby realizing pseudo-labeling learning of target domain samples.

[0228] The causal representation learning module is performed on the joint source and target domain data. The causal relationship between variables can be represented by the adjacency matrix corresponding to the DAG. To describe. Design a The input of the causal graph autoencoder is a combination of source domain and target domain data. After obtaining the low-dimensional feature representation, the label variable is introduced Merge as additional nodes to construct new hidden layers. Source domain data and target domain data The causal representation can be based on the DAG containing the causal relationship Export. Among them, Indicates that and Cascade to form a unified label variable. Next, the present invention will discuss how to obtain the labeled source domain data and the pseudo-labeled target domain data. .

[0229] S5.5, similar to GeAE, the input of the causal graph autoencoder is ,in, is the source domain data with real labels, is the target domain data with pseudo labels, and the objective function of the causal representation learning module is Including reconstruction loss, causal structure learning loss and cross entropy loss:

[0230] ;

[0231] in, is the number of samples, express The number of samples, for The reconstructed output of To control the strength of causal constraints, To balance the classification loss, is the regularized weight decay; Indicates the number of hidden layers; and denote the cross entropy loss and classifier respectively; Representation Label Merged Markov blanket; and Represents the first The weight parameters of the hidden layer; and Represent the low-dimensional representations in the encoder and decoder of the causal graph autoencoder respectively:

[0232] ;

[0233] ;

[0234] in, For the dimension, Indicates mapping the original input to low-dimensional features, which is used to transform the input data (source and target domain combined data) mapped to a low-dimensional latent space. Represents causal logic and label fusion, which is the second part of the encoder network. Its function is to represent the hidden Further transformation is performed to generate the adjacency matrix of the causal graph Compatible features. represents the decoder network, whose role is to reconstruct the original input data from the hidden representation adjusted by the causal graph, is the adjacency matrix .

[0235] S5.7, to ensure The acyclicity of the corresponding DAG must satisfy the following acyclicity constraints :

[0236] ;

[0237] in, represents the trace of the matrix, express Matrix index of ; is the Hadamard product; Represents the number of variables in the causal graph model, that is, the adjacency matrix The dimension is × After introducing the Lagrange multiplier and penalty term, the augmented Lagrangian is given by:

[0238] ;

[0239] in, represents the Lagrange multiplier, represents the penalty parameter.

[0240] S5.8, for the adjacency matrix and The update rules are as follows:

[0241] ;

[0242] ;

[0243] ;

[0244] in, and are two tuning hyperparameters, is the parameter value when the function is minimized. Following the above update rules, the adjacency matrix Gradient descent can be used for optimization. Then the causal representation of the source and target domain data can be derived, which is recorded as and .

[0245] It can be considered as a causal representation. The low-dimensional representations in can be divided into two groups: causal representations , and other feature representations that may be spurious features irrelevant to the task.

[0246] Following the classic domain adversarial network structure, we design a feature extractor , label classifier and domain discriminator The domain alignment module consists of . The feature extractor consists of three convolutional layers and one fully connected layer, while the label classifier and domain discriminator are both fully connected networks, responsible for fault type classification and domain discrimination, respectively. Based on the above network structure, the objective function of the domain alignment module is as follows:

[0247] .

[0248] S5.9, use 、 and Representation feature extractor , label classifier and domain discriminator The network model parameters in the domain alignment module are:

[0249] ;

[0250] in, is the feature extractor, For label classifier and For the domain discriminator: and denote the cross entropy loss functions for fault classification and domain classification, respectively; and Represent the domain labels of source domain data and target domain data respectively; is the balance parameter, represents the gradient reversal layer, Represents the true label of the source domain data.

[0251] S5.10, in order to use the stochastic gradient descent method to solve the above problem, the parameters The update rules are as follows:

[0252] ;

[0253] ;

[0254] .

[0255] in, represents the learning rate, and Represent fault classification loss and domain discrimination loss respectively:

[0256] ;

[0257] ;

[0258] By solving , it is possible to achieve alignment between source domain features and target domain features. In addition, using the trained label classifier Predict the fault samples in the target domain and finally obtain the state label of the target domain .

[0259] Specifically, step S6 includes the following steps:

[0260] S6.1, consider the influence of environmental parameters (such as temperature T, pressure P, salinity C), and introduce nonlinear drift terms and diffusion terms , constructing a nonlinear Wiener process modulated by environmental parameters:

[0261] ;

[0262] Among them, the drift term To reflect the accelerating effect of environmental parameters (temperature T, pressure P, salinity C) on degradation, the generalized Arrhenius model is used:

[0263] .

[0264] Diffusion term Characterize random degradation noise caused by environmental fluctuations:

[0265] ;

[0266] in, It is a degradation index, indicating the degradation of watertight connectors over time. The cumulative degradation of It is a standard Brownian motion (Wiener process), describing random fluctuations; is the initial degradation amount (usually assumed to be 0 or calibrated through experiments); These are parameters to be determined.

[0267] S6.2, using accelerated degradation test data, the maximum likelihood estimation function Estimating model parameters :

[0268] ;

[0269] ;

[0270] in, , , is an exponential function; Respectively temperature, pressure, and salinity of each path, For the The degradation index of a path.

[0271] S6.3, using Monte Carlo path simulation to generate multiple degradation paths to quantify uncertainty, dynamically inputting environmental parameter time series :

[0272] ;

[0273] in, , independent standard normal distributed random variables, introducing random fluctuations; is the time step.

[0274] S6.4, Assuming a Prior Distribution of Parameters , combined with real-time monitoring data, calculate the posterior distribution:

[0275] ;

[0276] in, is the observed degradation data, is proportional to; Monte Carlo path simulation is used to process high-dimensional parameter space.

[0277] S6.5, Define Failure Threshold , calculate the remaining service life RUL:

[0278]

[0279] in, is the time increment, is the zth path at time The predicted degradation amount, To obtain the minimum, the 2.5% and 97.5% quantiles of the RUL distribution are used as the 95% confidence interval.

[0280] Figure 2 The different colored curves represent 50 simulated degradation paths generated by different Monte Carlo simulations. The divergence of the curve paths reflects the uncertainty of the deep-sea environment, including deep-sea pressure fluctuations and temperature changes. The curves show an overall upward trend, reflecting the linear component of the cumulative damage of the connector material. The curves also show local fluctuations, which are due to the simulated short-term deep-sea environmental disturbances. The predicted values ​​of the normalized degradation amount at the end of the curve are distributed between the 95% confidence interval [-0.35, 0.55], reflecting the uncertainty of the degradation amount.

[0281] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for predicting the life of a deep-sea watertight connector, characterized in that: The specific steps include: S1, uses a multi-source sensor array in a deep-sea pressure chamber to collect temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress, and dielectric loss of watertight connectors in real time to construct a tensor dataset; S2, input the normalized tensor data set into the parallel factor analysis PARAFAC module to decompose the time evolution factor matrix, spatial distribution factor matrix, frequency domain characteristic factor matrix and environmental coupling factor matrix; S3, the decomposed feature tensor is input into the multi-scale sparse neural network MSNN, and the cross-scale degradation features are extracted through the void convolution module and attention gating mechanism; S4, the time-frequency features and spatial environment features are obtained through the CEEMDAN-KPCA algorithm, and the features are fused using the dual-channel Transformer-CBAM network; S5 uses a causal graph autoencoder to learn invariant fault representations under deep-sea environmental disturbances, and achieves cross-depth working condition distribution alignment through a domain adversarial network driven by a gradient reversal layer; S6, through the nonlinear Wiener process modeling of random fluctuations modulated by environmental parameters, combined with Monte Carlo path simulation and Bayesian inference to quantify the confidence interval, output the dynamic remaining life prediction results of the watertight connector.

2. The life prediction method of a deep seawater-tight connector according to claim 1, characterized in that: The tensor dataset in step S1 is a fourth-order tensor dataset ,in, is the time dimension; is the spatial dimension; is the frequency dimension; For the environmental dimension.

3. The life prediction method of a deep seawater-tight connector according to claim 1, characterized in that: Step S2 specifically includes the following steps: S2.1, perform Z-score standardization on the data of each dimension: ； in, and Respectively The mean and standard deviation of the environmental parameters, is a fourth-order tensor data set, After standardization ; The scalar form formula of the fourth-order parallel factor molecule tensor decomposition is: ； in, is the number of factors, It is Time factor, It is A spatial factor, It is frequency factors, It is environmental factors, represents the error set of the fourth-order tensor; S2.2, the normalized tensor dataset is input into the parallel factor analysis PARAFAC module to decompose 、 、 、 Four matrices, defining the time evolution factor matrix for × dimensional matrix, is the time factor; define the spatial distribution factor matrix for × dimensional matrix, is the spatial factor; define the frequency domain characteristic factor matrix for × dimensional matrix, is the frequency factor; define the environmental coupling factor matrix for × dimensional matrix, is the environmental factor; S2.3, Calculation 、 、 and , to calculate The other matrix calculation methods are the same as same: ； in, Indicates that 、 、 The matrix is ​​block diagonalized on the column vectors of , and the error between the observed data and the model prediction value is expressed as express, The least squares estimate of is expressed as follows: ； in, represents the generalized inverse; S2.4, Repeated Counting 、 、 and , until the result converges or the set number of iterations is reached and then stops running, and .

4. The life prediction method of a deep seawater-tight connector according to claim 3, characterized in that: In step S3, the multi-scale features are extracted in the dilated convolution module as follows: The result obtained in step 2 、 、 and Concatenate into feature tensors , and Input the dilated convolution module of the multi-scale sparse neural network MSNN to extract multi-scale features and output a multi-scale feature map: ； ； in, is the convolution kernel size, The expansion ratio is 1, 2 and 4, For the The convolution kernel weights at each position, is the factor of the current position, is a multi-scale feature map, is the sequence length, is the number of channels.

5. The life prediction method of a deep seawater-tight connector according to claim 4, characterized in that: In step S3, the extraction of features of different granularities through the attention gating mechanism specifically includes the following steps: S3.1, the multi-scale feature map output by the dilated convolution module Cascade and assign weights through the attention gating mechanism : ； in, is the learnable weight matrix, is the bias term; S3.2, fuse the weighted multi-scale features: ； in, is the fused multi-scale feature map; S3.3, Input residual connection and nonlinear projection module, output fusion features : ； ； in, is the eigenvector after residual connection, is the learnable projection matrix, is the column projection vector of the linear combination concatenated features, is the cascade operator, is the bias, .

6. The life prediction method of a deep seawater-tight connector according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, the fusion features output by MSNN are used as input signals and decomposed into multiple intrinsic mode functions through the CEEMDAN algorithm and the residual , as shown below: ; in, is the characteristic number, For time; S4.2, extract the energy of each IMF and calculate the energy entropy value : ; S4.3, multiple IMF energies Input KPCA dimensionality reduction, through the kernel function Mapping to high-dimensional space for principal component analysis, we can obtain time-frequency characteristics and spatial environment characteristics: ； ； in, is an element in the kernel matrix; represents the energy of the IMF, is the kernel matrix eigenvector, is the projection matrix, Score the principal components; S4.4, the temporal frequency features and spatial environment features are extracted into the global relationship through the multi-head attention module, and then input into the CBAM attention mechanism module; S4.5, the overall process of CBAM is expressed as: ； ； in, represents the Hadamard product, For CAM module, For the SAM module, represents the input features, Represents the feature map obtained by the CAM module, Represents the feature map obtained by the SAM module; S4.6, 、 and Perform weighted concatenation and map it into health indicators through a fully connected layer : ； in, and is a learnable parameter, is the Sigmoid function, For splicing operation.

7. The life prediction method of a deep seawater-tight connector according to claim 1, characterized in that: Step S5 specifically includes the following steps: S5.1, health indicators Combined with environmental parameters, environmental parameters include: pressure ,temperature ,salinity , construct the input feature matrix : ； in, , , , , after the merger ; S5.2, use the fuzzy C-means clustering algorithm to generate pseudo labels corresponding to target domain samples , the objective function of FCM for: ； in, represents the membership matrix, represents the cluster center, and represent the number of clusters and the sample size of the target domain, respectively. is the fuzziness indicator, for Target domain samples in ; S5.3, by introducing Lagrange multipliers Constructing Lagrangian functions ,as follows: ； S5.4, respectively, for the membership , cluster center and Lagrange multipliers Find the partial derivative and set it to zero, and we get The cluster centers of the iteration and membership Update formula: ； 。 8. The life prediction method of a deep seawater-tight connector according to claim 7, characterized in that: Step S5 also includes the following steps: S5.5, the input of the causal graph autoencoder is ,in, is the source domain data with real labels, is the target domain data with pseudo labels, and the objective function of the causal representation learning module is Including: reconstruction loss, causal structure learning loss and cross entropy loss: ； in, is the number of samples, express The number of samples, for The reconstructed output of To control the strength of causal constraints, To balance the classification loss, is the regularized weight decay; Indicates the number of hidden layers; and denote the cross entropy loss and classifier respectively; Representation Label Merged Markov blanket; and Represents the first The weight parameters of the hidden layer; and Represent the outputs of the encoder and decoder of the causal graph autoencoder respectively: ； ； in, For the dimension, Indicates mapping the original input into low-dimensional features, Indicates causal logic and label fusion, represents the decoder network, is the adjacency matrix; S5.6, to ensure The acyclicity of the corresponding DAG must satisfy the following acyclicity constraints : ； in, represents the trace of the matrix, express Matrix index of ; is the Hadamard product; Indicates the number of variables in the causal graph model; S5.7, after introducing the Lagrange multiplier and penalty term, the augmented Lagrangian is given by: ； in, represents the Lagrange multiplier, represents the penalty parameter.

9. The life prediction method of a deep seawater-tight connector according to claim 8, characterized in that: Step S5 also includes the following steps: S5.8, for the adjacency matrix and The update rules are as follows: ； ； ； in, and are two tuning hyperparameters, is the parameter value when the function is minimized; S5.9, use 、 and Representation feature extractor , label classifier and domain discriminator The network model parameters in the domain alignment module are: ； in, is the feature extractor, For label classifier and For the domain discriminator: and denote the cross entropy loss functions for fault classification and domain classification, respectively; and Represent the domain labels of source domain data and target domain data respectively; is the balance parameter, represents the gradient reversal layer, Represents the true label of the source domain data; S5.10, parameters The update rules are as follows: ； ； ； in, represents the learning rate, and Represent fault classification loss and domain discrimination loss respectively: ； 。 10. The life prediction method of a deep seawater-tight connector according to claim 1, characterized in that: Step S6 specifically includes the following steps: S6.1, considering the influence of environmental parameters, introduce nonlinear drift terms and diffusion terms , constructing a nonlinear Wiener process modulated by environmental parameters: ； Among them, the drift term To reflect the accelerating effect of environmental parameters on degradation, the generalized Arrhenius model is used: ； Diffusion term Characterize random degradation noise caused by environmental fluctuations: ； in, It is a degradation index, indicating the degradation of watertight connectors over time. The cumulative degradation of is the standard Brownian motion; is the initial degradation amount; These are parameters to be determined; S6.2, using accelerated degradation test data, the maximum likelihood estimation function Estimating model parameters : ； ； in, , , is an exponential function; Respectively temperature, pressure, and salinity of each path, For the The degradation index of each path; S6.3, using Monte Carlo path simulation to generate multiple degradation paths: ； in, , independent standard normal distributed random variables, introducing random fluctuations; is the time step; S6.4, Assuming a Prior Distribution of Parameters , combined with real-time monitoring data, calculate the posterior distribution: ； in, is the observed degradation data, is proportional to; S6.5, Define Failure Threshold , calculate the remaining service life RUL: ； in, is the time increment, is the zth path at time The predicted degradation amount, To take the minimum.

Images