Method for predicting service life of deep sea watertight connector

Abstract

The invention provides a service life prediction method for a deep sea watertight connector, and relates to the field of service life prediction, and the method specifically comprises the steps: constructing a tensor data set; inputting the decomposed feature tensor into a multi-scale sparse neural network, and extracting cross-scale degradation features through a cavity convolution module and an attention gating mechanism; the features are fused by using a dual-channel Transform-CBAM (Content Border Administration and Maintenance) network; learning invariant fault representation under deep sea environment disturbance by using a causal graph auto-encoder, and realizing cross-depth working condition distribution alignment through a domain adversarial network driven by a gradient inversion layer; and modeling random fluctuation influence through a nonlinear Wiener process of environmental parameter modulation, performing confidence interval quantization in combination with Monte Carlo path simulation and Bayesian inference, and outputting a dynamic residual life prediction result of the watertight connector. According to the technical scheme, the problem that the service life of the deep sea watertight connector cannot be accurately predicted in the prior art is solved.

Description

Technical Field

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

[0002] Deep - sea watertight connectors are important components for deep - sea equipment to achieve power distribution, signal transmission, and communication. Their reliability directly determines the long - term operation of major equipment such as underwater production systems in extreme environments. This component is long - term subjected to the multi - field coupling effects of high pressure, high - salt corrosion, temperature gradient, and dynamic mechanical loads, resulting in the performance degradation showing strong non - linear and non - stationary characteristics. Existing life prediction methods face four major bottlenecks: First, the synergistic action mechanism of high pressure, low temperature, corrosion, and mechanical loads is complex, and traditional single - factor models are difficult to accurately decouple the multi - physical - field coupling effects; Second, the high cost of deep - sea tests leads to scarce failure data, and the generalization ability of traditional statistical models is insufficient under small - sample conditions; Third, material damage has threshold effects and non - linear mutation characteristics, and crack propagation shows a three - stage mode of "slow - fast - failure". Existing linear degradation models cannot capture the performance inflection point at the end of life; Fourth, there are significant differences between the dynamic pressure fluctuations in the laboratory environment and real sea conditions, and the coupling effect of bio - attachment - corrosion. The migration reliability of cross - sea area life prediction needs to be improved urgently.

[0003] Therefore, there is a need for a life prediction method for deep - sea watertight connectors that can accurately predict the life of deep - sea watertight connectors. Summary of the Invention

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

[0005] To achieve the above - mentioned purpose, the present invention provides a life prediction method for deep - sea watertight connectors, which specifically includes the following steps: S1, real - time collect temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress, and the dielectric loss of the watertight connector through a multi - source sensor array in a deep - sea pressure chamber to construct a tensor data set.

[0006] S2, input the standardized tensor data set into the parallel factor analysis (PARAFAC) module to decompose the time - evolution factor matrix, spatial - distribution factor matrix, frequency - domain feature factor matrix, and environmental - coupling factor matrix.

[0007] S3, input the decomposed feature tensor into a multi - scale sparse neural network (MSNN), and extract cross - scale degradation features through a dilated convolution module and an attention gating mechanism.

[0008] S4. Obtain the time-frequency features and spatial environment features through the CEEMDAN-KPCA algorithm, and fuse the features using the dual-channel Transformer-CBAM network.

[0009] S5. Use the causal graph autoencoder to learn the invariant fault representation under deep-sea environment disturbances, and achieve cross-depth working condition distribution alignment through the domain adversarial network driven by the gradient reversal layer.

[0010] S6. Model the stochastic fluctuation effect through the non-linear Wiener process modulated by environmental parameters, combine Monte Carlo path simulation and Bayesian inference for confidence interval quantification, and output the dynamic remaining life prediction result of the watertight connector.

[0011] Furthermore, the tensor dataset in step S1 is a fourth-order tensor dataset , where is the time dimension; is the spatial dimension; is the frequency dimension; is the environmental dimension.

[0012] Furthermore, step S2 specifically includes the following steps: S2.1. Perform Z-score normalization on the data of each dimension respectively: ; where and are the mean and standard deviation of the th environmental parameter respectively, is the fourth-order tensor dataset, is the after normalization.

[0013] The scalar form formula of the fourth-order parallel factor PARAFAC tensor decomposition is: ; where is the number of factors, is the th time factor, is the th spatial factor, is the th frequency factor, is the th environmental factor, represents the error set of the fourth-order tensor.

[0014] S2.2. Input the normalized tensor dataset into the parallel factor analysis PARAFAC module to decompose , , , Four matrices define the time evolution factor matrix as × dimensional matrix, is the time factor; define the spatial distribution factor matrix as × dimensional matrix, is the spatial factor; define the frequency domain feature factor matrix as × dimensional matrix, is the frequency factor; define the environmental coupling factor matrix as × dimensional matrix, is the environmental factor.

[0015] S2.3. Calculate , , and to calculate, taking the method as an example, the calculation methods of the remaining matrices are the same as : ; where means block diagonalizing and combining the column vectors of the , , matrices. The error between the observed data and the model prediction value is represented by , and the least squares estimate of is expressed by the following formula: ; where represents the generalized inverse.

[0016] S2.4. Repeat the calculation of , , and until the result converges or reaches the set number of iterations and then stops running, and .

[0017] Furthermore, the specific process of extracting multi-scale features in the dilated convolution module in step S3 is as follows: Concatenate the , , and obtained in step 2 into a feature tensor , and The dilated convolution module of the input multi-scale sparse neural network MSNN extracts multi-scale features and outputs a multi-scale feature map: ; ; Among them, is the convolution kernel size, the dilation rate takes values of 1, 2, and 4, is the convolution kernel weight at the th position, is the factor at the current position, is the multi-scale feature map, is the sequence length, is the number of channels.

[0018] Furthermore, the specific steps for extracting features of different granularities through the attention gating mechanism in step S3 are as follows: S3.1, concatenate the multi-scale feature maps output by the dilated convolution module and assign weights through the attention gating mechanism : ; Among them, is the learnable weight matrix, is the bias term.

[0019] S3.2, fuse the weighted multi-scale features: ; Among them, is the fused multi-scale feature map.

[0020] S3.3, input into the residual connection and non-linear projection module, and output the fused feature : ; ; Among them, is the feature vector after the residual connection, is the learnable projection matrix, is the column projection vector of the linearly combined concatenated features, is the concatenation operator, is the bias, .

[0021] Furthermore, step S4 specifically includes the following steps: S4.1, take the fused feature output by MSNN as the input signal and decompose it into multiple intrinsic mode functions through the CEEMDAN algorithm and the residual term , as follows: ; where is the number of features, is the time.

[0022] S4.2, Extract the energy of each IMF and calculate the energy entropy value : .

[0023] S4.3, Input the energies of multiple IMFs into KPCA for dimensionality reduction, and map them to a high-dimensional space through the kernel function for principal component analysis to obtain time-frequency features and spatial environment features: ; ; where is an element in the kernel matrix; represents the energy of the IMF, is the eigenvector of the kernel matrix, is the projection matrix, is the principal component score.

[0024] S4.4, Extract the global relationship of the time-frequency features and spatial environment features through the multi-head attention module, and then input them into the CBAM attention mechanism module.

[0025] S4.5, The overall process of CBAM is expressed as: ; ; where represents the Hadamard product, is the CAM module, is the SAM module, represents the input feature, represents the feature map obtained after passing through the CAM module, represents the feature map obtained after passing through the SAM module.

[0026] S4.6, Perform weighted splicing on , and , and map them to a health index through a fully connected layer : ; where and is a learnable parameter, is the Sigmoid function, is the concatenation operation.

[0027] Furthermore, step S5 specifically includes the following steps: S5.1, Combine the health indicators with the environmental parameters, where the environmental parameters include: pressure , temperature , salinity to construct the input feature matrix : ; wherein, , , , , after merging .

[0028] S5.2, Use the fuzzy C-means clustering algorithm to generate the pseudo-labels corresponding to the target domain samples , and the objective function of FCM is: ; wherein, represents the membership matrix, represents the cluster center, and respectively represent the number of clusters and the number of target domain samples, is the fuzziness index, is the target domain samples in

[0029] S5.3, By introducing Lagrange multipliers construct the Lagrangian function , as follows: .

[0030] S5.4, Respectively take the partial derivatives of the membership , the cluster center and the Lagrange multiplier , and set the partial derivatives to zero to obtain the -th iteration of the cluster center and the membership update formulas: ; .

[0031] Furthermore, step S5 also includes the following steps: S5.5, the input of the causal graph auto - encoder is , where is the source domain data with true labels, is the target domain data with pseudo - labels. The objective function of the causal representation learning module includes: reconstruction loss, causal structure learning loss, and cross - entropy loss: ; where is the number of samples, represents the number of samples of is the reconstruction output of is to control the causal constraint strength, is the balanced classification loss, is the regularization weight decay; represents the number of hidden layers; and represent the cross - entropy loss and the classifier respectively; represents the label merged Markov blanket; and represent the weight parameters of the th hidden layer in the encoding and decoding processes respectively; and represent the outputs of the encoder and decoder of the causal graph auto - encoder respectively: ; ; where is the dimension, represents mapping the original input to low - dimensional features, represents causal logic and label fusion, represents the decoder network, is the adjacency matrix.

[0032] S5.6, to ensure the acyclicity of the DAG corresponding to , the following acyclicity constraints need to be satisfied : ; where represents the trace of the matrix, represents the matrix exponential of is the Hadamard product; represents the number of variables in the causal graph model.

[0033] S5.7, after introducing the Lagrange multiplier and the penalty term, the augmented Lagrangian is given as: ; where, denotes the Lagrange multiplier, denotes the penalty parameter.

[0034] Furthermore, step S5 also includes the following steps: S5.8, for the adjacency matrix and there are the following update rules: ; ; ; where, and are two tuning hyperparameters, is the parameter value when taking the minimum value of the function.

[0035] S5.9, use , and to represent the network model parameters in the feature extractor , the label classifier and the domain discriminator . The objective function of the domain alignment module is:

[0036] where, is the feature extractor, is the label classifier and is the domain discriminator: and respectively represent the cross-entropy loss functions for fault classification and domain classification; and respectively represent the domain labels of the source domain data and the target domain data; is the balance parameter, represents the gradient reversal layer, represents the true label of the source domain data.

[0037] S5.10, the update rule of the parameter is as follows: ; ; .

[0038] where, denotes the learning rate, and respectively represent the fault classification loss and the domain discrimination loss: ; .

[0039] Furthermore, step S6 specifically includes the following steps: S6.1, considering the influence of environmental parameters, introduce the nonlinear drift term and the diffusion term , and construct a nonlinear Wiener process modulated by environmental parameters: ; wherein, the drift term reflects the acceleration effect of environmental parameters on degradation, and adopts the generalized Arrhenius model: .

[0040] The diffusion term characterizes the random degradation noise caused by environmental fluctuations: ; wherein, is the degradation index, indicating the cumulative degradation amount of the watertight connector at time ; is the standard Brownian motion; is the initial degradation amount; are all parameters to be determined.

[0041] S6.2, using the accelerated degradation test data, estimate the model parameters through the maximum likelihood estimation function : ; ; wherein, , , is the exponential function; are respectively the temperature, pressure and salinity of the th path, is the degradation index of the th path.

[0042] S6.3, adopt Monte Carlo path simulation to generate multiple degradation paths: ; wherein, , independent standard normal distribution random variables, introducing random fluctuations; is the time step.

[0043] S6.4, Assume the prior distribution of the parameters , and combine with the real-time monitoring data to calculate the posterior distribution: ; where is the observed degradation data, is proportional to.

[0044] S6.5, Define the failure threshold , and calculate the remaining useful life (RUL): ; where is the time increment, is the predicted degradation amount of the z-th path at time , is to take the minimum.

[0045] The present invention has the following beneficial effects: 1. The present invention constructs a spatio-temporal-frequency environmental fourth-order tensor dataset through a multi-source sensor array, and combines parallel factor analysis (PARAFAC) to decompose the four-dimensional factor matrices of time evolution, spatial distribution, frequency-domain characteristics, and environmental coupling. This design realizes the accurate decoupling of the multi-field coupling effects of high pressure, low temperature, corrosion, and mechanical loads, and solves the problem that traditional single-factor models cannot handle non-linear degradation.

[0046] 2. A multi-scale sparse neural network (MSNN) is designed for the non-linear mutation characteristics of the performance degradation of waterproof connectors. The MSNN module adopts a dilated convolutional layer and a multi-head attention gating mechanism to capture local details, mid-range correlations, and global degradation patterns respectively. High-order feature interactions are retained through residual connections and non-linear projections, significantly improving the characterization ability of the three-stage crack propagation of "slow-fast-failure", and solving the bottleneck that linear degradation models cannot capture the performance inflection point at the end of life.

[0047] 3. The IMF energy entropy of non-stationary signals is extracted through complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), combined with kernel principal component analysis (KPCA) for dimensionality reduction, and a two-channel Transformer-CBAM network is designed: the multi-head self-attention mechanism is adopted in the time-frequency domain, and the channel-spatial attention weighting is used in the space-frequency domain. Finally, the time-frequency - space-frequency joint features are fused to generate a non-linear health index (HI). This scheme realizes the robust fusion of multi-source heterogeneous signals and solves the problem of insufficient generalization ability of traditional statistical models under small sample conditions.

[0048] 4. Introduce a causal graph autoencoder to learn invariant fault representations under environmental perturbations. Model the causal relationships between variables through an acyclic adjacency matrix Icausal, and combine a domain adversarial network (DANN) driven by a gradient reversal layer to align the feature distributions of the source domain and the target domain. This design significantly improves the migration reliability of cross-sea area life prediction and overcomes the influence of the dynamic differences between the laboratory environment and the real sea conditions.

[0049] 5. Construct a generalized Arrhenius drift term and a linear diffusion term, generate random degradation trajectories through Monte Carlo path simulation, and update the posterior distribution of parameters in real time based on Bayesian inference. Finally, output the dynamic remaining useful life (RUL) and its 95% confidence interval, realize the quantification of uncertainty under dynamic modulation of multi-environment parameters, and provide a high-confidence basis for the maintenance decision-making of deep-sea equipment. Brief Description of the Drawings

[0050] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. In the drawings: Figure 1 The flowchart of a method for predicting the life of a deep-sea watertight connector according to the present invention is shown.

[0051] Figure 2 The life degradation path diagram of the deep-sea watertight connector is shown. Detailed Embodiments

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0053] As Figure 1 shown, a method for predicting the life of a deep-sea watertight connector specifically includes the following steps: S1. Real-time collect temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress, and the dielectric loss of the watertight connector through a multi-source sensor array in a deep-sea pressure cabin, and construct a tensor data set.

[0054] S2. Input the standardized tensor data set into a parallel factor analysis PARAFAC module to decompose a time evolution factor matrix, a spatial distribution factor matrix, a frequency domain feature factor matrix, and an environmental coupling factor matrix.

[0055] S3. Input the decomposed feature tensors into the multi-scale sparse neural network (MSNN), and extract cross-scale degradation features through the dilated convolution module and the attention gating mechanism.

[0056] S4. Obtain the time-frequency features and spatial environment features through the CEEMDAN-KPCA algorithm, and fuse the features using the dual-channel Transformer-CBAM network.

[0057] S5. Use the causal graph autoencoder to learn the invariant fault representation under deep-sea environmental perturbations, and achieve cross-depth working condition distribution alignment through the domain adversarial network driven by the gradient reversal layer.

[0058] S6. Model the stochastic fluctuation effect through the non-linear Wiener process modulated by environmental parameters, combine Monte Carlo path simulation and Bayesian inference for confidence interval quantification, and output the dynamic remaining life prediction result of the watertight connector.

[0059] Specifically, the tensor dataset in step S1 is a fourth-order tensor dataset , where is the time dimension; is the spatial dimension (number of sensor channels); is the frequency dimension; is the environmental dimension (such as temperature, pressure, salinity, etc.).

[0060] Specifically, step S2 specifically includes the following steps: S2.1. Perform Z-score standardization on the data of each dimension respectively: ; where and are the mean and standard deviation of the th environmental parameter respectively, is the fourth-order tensor dataset, is the standardized .

[0061] The scalar form formula for the fourth-order parallel factor molecular tensor decomposition is: ; where is the number of factors, is the th time factor, is the th spatial factor, is the th frequency factor, is the th environmental factor, The error set representing the fourth-order tensor.

[0062] S2.2. Input the standardized tensor data set into the parallel factor analysis (PARAFAC) module to decompose , , , into four matrices. Define the time evolution factor matrix as a × dimensional matrix, being the time factor; define the spatial distribution factor matrix as a × dimensional matrix, being the spatial factor; define the frequency domain feature factor matrix as a × dimensional matrix, being the frequency factor; define the environmental coupling factor matrix as a × dimensional matrix, being the environmental factor.

[0063] S2.3. Calculate , , and to calculate, taking the method as an example, the calculation methods of the remaining matrices being the same as : .

[0064] Among them, represents the block diagonalization combination of the column vectors of the , , matrices. The error between the time observation data and the model prediction value is represented by , and the least squares estimate of is expressed by the following formula: ; Among them, represents the generalized inverse.

[0065] Secondly, .

[0066] Among them, represents the block diagonalization combination of the column vectors of matrices A, C, and D. The error between the spatial observation data and the model prediction value is represented by , and the least squares estimate of is expressed by the following formula: 。

[0067] The matrix C is obtained through the following formula: 。

[0068] denotes the block diagonalization combination of the column vectors of matrices A, B, and D. The error between the frequency-domain observed data and the model prediction value is represented by and is denoted as The least squares estimate of is expressed by the following formula: 。

[0069] The matrix D is obtained through the following formula: ; where denotes the block diagonalization combination of the column vectors of matrices A, B, and C. The error between the environmental observed data and the model prediction value is represented by and is denoted as The least squares estimate of is expressed by the following formula: 。

[0070] S2.4, Repeatedly calculate 、 、 and until the result converges or the set number of iterations is reached and then stop running. By , the uniqueness of the parallel factor decomposition solution is guaranteed.

[0071] Specifically, the extraction of multi-scale features in the dilated convolution module in step S3 is as follows: The 、 、 and obtained in step 2 are concatenated into a feature tensor , and is input into the dilated convolution module of the multi-scale sparse neural network MSNN to extract multi-scale features, and the multi-scale feature map is output: ; ; where is the convolution kernel size, is the dilation rate with values of 1, 2, and 4, is the convolution kernel weight at the th position, is the factor at the current position in the time dimension or spatial dimension or frequency dimension or environmental dimension, is a multi-scale feature map, is the sequence length, is the number of channels.

[0072] Dilation rate = 1, capturing local features; dilation rate = 2, capturing medium-range features; dilation rate = 4, capturing global features.

[0073] Specifically, the steps of extracting features of different granularities through the attention gating mechanism in step S3 specifically include the following steps: S3.1, concatenate the multi-scale feature maps output by the dilated convolution module and assign weights through the attention gating mechanism : ; wherein, is a learnable weight matrix (transpose form), is the bias term.

[0074] S3.2, fuse the weighted multi-scale features: ; wherein, is the fused multi-scale feature map.

[0075] S3.3, input into the residual connection and non-linear projection module, and output the fused feature : ; ; wherein, is the feature vector after residual connection, is the learnable projection matrix, is the column projection vector of the linearly combined cascaded features, is the cascade operator, is the bias, .

[0076] Specifically, step S4 specifically includes the following steps: S4.1, take the fused feature output by MSNN as the input signal, and decompose it into multiple intrinsic mode functions and the residual term , as shown below: ; wherein, is the number of features, is the time.

[0077] S4.2, Extract the energy of each IMF and calculate the energy entropy value : .

[0078] S4.3, Input the energies of multiple IMFs into the kernel principal component analysis method KPCA for dimensionality reduction. Through the kernel function map it to a high-dimensional space for principal component analysis to obtain time-frequency features and spatial environment features: ; ; wherein, is an element in the kernel matrix; represents the energy of the IMF, is the eigenvector of the kernel matrix, is the projection matrix, is the principal component score.

[0079] S4.4, Extract the global relationship of the time-frequency features and spatial environment features through the multi-head attention module, and then input them into the CBAM attention mechanism module.

[0080] Step S4.4 specifically includes the following steps: Extract the global relationship through multi-head self-attention, calculate the attention function of a group of queries and combine them into matrix Q, and at the same time pack the keys and values into matrices K and V to improve the calculation efficiency.

[0081] The principle of the multi-head self-attention module is expressed as: ; ; ; wherein, Q, K, and V respectively represent the query, key, and value matrices, represents the dimension of each vector in the input sequence, represents the learnable weight parameter to fuse , and n represents the number of heads.

[0082] CBAM contains two sub-modules, the channel attention module CAM and the spatial attention module SAM. The channel attention is fixed on different channels of the feature map and assigns different weights to each channel. CAM is expressed as: ; wherein, is the CAM module, is the activation function, and represents the weight matrix, and represent the features after average and max pooling operations, represents the input feature, represents the feature map after passing through the CAM module.

[0083] SAM is expressed in the following form: ; where, is the SAM module, represents the convolution operation with a kernel size of , represents the stacked feature maps, represents the feature map after passing through the SAM module.

[0084] S4.5, the overall process of CBAM is expressed as: ; ; where, represents the Hadamard product, is the CAM module, is the SAM module, represents the input feature, represents the feature map obtained after passing through the CAM module, represents the feature map obtained after passing through the SAM module.

[0085] S4.6, perform weighted concatenation on , and and map it to the health index through a fully connected layer : ; where, and are learnable parameters, is the Sigmoid function, is the concatenation operation, and the output is normalized to a health score in the range of (0, 1).

[0086] Specifically, step S5 specifically includes the following steps: S5.1, combine the health index with the environmental parameters, and the environmental parameters include: pressure , temperature , salinity , to construct the input feature matrix : ; Among them, , , , , after merging .

[0087] S5.2, Use the fuzzy C-means (FCM) clustering algorithm to generate pseudo-labels corresponding to the target domain samples , The objective function of FCM is: ; Among them, represents the membership matrix, represents the cluster center, and represent the number of clusters and the number of target domain samples respectively, is the fuzziness index, is the target domain samples in

[0088] S5.3, By introducing Lagrange multipliers construct the Lagrangian function , as follows: .

[0089] S5.4, Respectively take the partial derivatives of the membership , the cluster center and the Lagrange multiplier , and set the partial derivatives to zero to obtain the cluster center and the membership update formulas for the ; .

[0090] Based on the above update rules, set the maximum number of iterations or the convergence threshold to obtain the final membership matrix, thus realizing the pseudo-label learning of the target domain samples.

[0091] Execute the causal representation learning module on the combined source domain and target domain data. The causal relationship between variables can be described by the adjacency matrix corresponding to the DAG. Design a deep autoencoder network for learning to obtain the causal representation. The input of the causal graph autoencoder is the combination of the source domain and target domain data. After obtaining the low-dimensional feature representation, introduce the label variable merged as an additional node to construct a new hidden layer. The source domain data and the target domain data causal representation can be derived according to the corresponding to the DAG containing causal relationships. Among them, denotes concatenating and to form a unified label variable. Next, the present invention will discuss how to obtain based on the labeled source domain data and the target domain data with pseudo labels.

[0092] S5.5. Similar to GeAE, the input of the causal graph autoencoder is , where is the source domain data with true labels, is the target domain data with pseudo labels, and the objective function of the causal representation learning module includes a reconstruction loss, a causal structure learning loss, and a cross-entropy loss: ; Among them, is the number of samples, denotes the number of samples of, is the reconstruction output of , is to control the causal constraint strength, is to balance the classification loss, is the regularization weight decay; denotes the number of hidden layers; and denote the cross-entropy loss and the classifier respectively; denotes the label merged Markov blanket; and denote the weight parameters of the th hidden layer in the encoding and decoding processes respectively; and denote the low-dimensional representations in the encoder and decoder of the causal graph autoencoder respectively: ; ; Among them, is the dimension, denotes mapping the original input to low-dimensional features, and its role is to map the input data (the merged data of the source domain and the target domain) to the low-dimensional hidden space. denotes causal logic and label fusion, which is the second part of the encoder network, and its role is to further transform the hidden representation to generate the adjacency matrix Compatible features. Denote the decoder network, which is used to reconstruct the original input data from the causal graph adjusted hidden representation, is the adjacency matrix .

[0093] S5.7. To ensure the acyclicity of the corresponding DAG, the following acyclicity constraints need to be satisfied : ; where denotes the trace of the matrix, denotes the matrix exponential of; is the Hadamard product; denotes the number of variables in the causal graph model, that is, the dimension of the adjacency matrix is × . After introducing the Lagrange multiplier and penalty term, the augmented Lagrangian is given: ; where denotes the Lagrange multiplier, denotes the penalty parameter.

[0094] S5.8. For the adjacency matrix and there are the following update rules: ; ; ; where and are two tuning hyperparameters, is the parameter value when taking the minimum value of the function. Following the above update rules, the adjacency matrix can be optimized using the gradient descent method. Then the causal representations of the source domain and target domain data can be derived, denoted as and .

[0095] can be considered as causal representations, the low-dimensional representations in can be divided into two groups: causal representations , and other feature representations that may be task-irrelevant spurious features.

[0096] Following the classical domain adversarial network structure, design a feature extractor , label classifier and domain discriminator The domain alignment module consists of a feature extractor, a label classifier, and a domain discriminator. The feature extractor is composed of 3 convolutional layers and 1 fully connected layer, while both the label classifier and the domain discriminator are fully connected networks, responsible for classifying the fault types and discriminating the domains respectively. Based on the above network structure, the objective function of the domain alignment module is as follows: .

[0097] S5.9, use , and to represent the network model parameters in the feature extractor , the label classifier and the domain discriminator . The objective function of the domain alignment module is: ; where, is the feature extractor, is the label classifier and is the domain discriminator: and represent the cross-entropy loss functions for fault classification and domain classification respectively; and represent the domain labels of the source domain data and the target domain data respectively; is the balance parameter, represents the gradient reversal layer, represents the true label of the source domain data.

[0098] S5.10, in order to use the stochastic gradient descent method to solve the above problem, the update rules of the parameter are as follows: ; ; .

[0099] where, represents the learning rate, and represent the fault classification loss and the domain discrimination loss respectively: ; ; By solving , the alignment between the source domain features and the target domain features can be achieved. In addition, using the trained label classifier to predict the fault samples on the target domain, the state label of the target domain can be finally obtained.

[0100] Specifically, step S6 specifically includes the following steps: S6.1. Consider the influence of environmental parameters (such as temperature T, pressure P, salinity C), introduce a non-linear drift term and a diffusion term , and construct a non-linear Wiener process modulated by environmental parameters: ; Among them, the drift term reflects the accelerating effect of environmental parameters (temperature T, pressure P, salinity C) on degradation, and adopts the generalized Arrhenius model: .

[0101] The diffusion term characterizes the random degradation noise caused by environmental fluctuations: ; Among them, is the degradation index, indicating the cumulative degradation amount of the watertight connector at time ; is the standard Brownian motion (Wiener process), describing random fluctuations; is the initial degradation amount (usually assumed to be 0 or calibrated through experiments); are all parameters to be determined.

[0102] S6.2. Use the accelerated degradation test data to estimate the model parameters through the maximum likelihood estimation function : ; ; Among them, , , is an exponential function; are the temperature, pressure and salinity of the -th path respectively, is the degradation index of the -th path.

[0103] S6.3. Adopt Monte Carlo path simulation to generate multiple degradation paths to quantify uncertainty, and dynamically input the time series of environmental parameters : ; Among them, , independent standard normal distribution random variables, introducing random fluctuations; is the time step.

[0104] S6.4. Assume the prior distribution of parameters , in combination with real-time monitoring data, calculate the posterior distribution: ; Among them, is the observed degradation data, is proportional to; Monte Carlo path simulation is used to process the high-dimensional parameter space.

[0105] S6.5, define the failure threshold , calculate the remaining useful life RUL:

[0106] Among them, is the time increment, is the predicted degradation amount of the z-th path at time , is to take the minimum. The 2.5% and 97.5% quantiles of the RUL distribution are statistically taken as the 95% confidence interval.

[0107] Figure 2 In [], the different color curves represent 50 simulated degradation paths generated by different Monte Carlos. The divergence of the curve paths reflects the uncertainties of the deep-sea environment, deep-sea pressure fluctuations and temperature changes; the overall curve shows an upward trend, reflecting the linear component of the cumulative damage of the connector material; there are also local fluctuations in the curve, which is 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 intervals [-0.35, 0.55], reflecting the uncertainty of the degradation amount.

[0108] 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 those skilled in the art within the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A method for predicting the service life of a deep-sea watertight connector, characterized in that, Specifically, it includes the following steps: S1. Real-time collect temperature, pressure, salinity, current, discharge amplitude, vibration spectrum, resistance, contact stress, and dielectric loss of the watertight connector through a multi-source sensor array in a deep-sea pressure chamber, and construct a tensor data set; S2. Input the standardized tensor data set into the parallel factor analysis (PARAFAC) module to decompose the time evolution factor matrix, spatial distribution factor matrix, frequency-domain feature factor matrix, and environmental coupling factor matrix; S3. Input the decomposed feature tensor into the multi-scale sparse neural network (MSNN), and extract cross-scale degradation features through the dilated convolution module and the attention gating mechanism; S4. Obtain time-frequency features and spatial environment features through the CEEMDAN-KPCA algorithm, and fuse the features using a two-channel Transformer-CBAM network; S5. Use a causal graph autoencoder to learn the invariant fault representation under deep-sea environmental disturbances, and achieve cross-depth working condition distribution alignment through a domain adversarial network driven by a gradient reversal layer; S6. Model the stochastic fluctuation effect through a non-linear Wiener process modulated by environmental parameters, combine Monte Carlo path simulation and Bayesian inference for confidence interval quantification, and output the dynamic remaining life prediction result of the watertight connector.

2. The method for predicting the lifespan of a deep-sea watertight connector according to claim 1, wherein The tensor data set in step S1 is a fourth-order tensor data set , where is the time dimension; is the spatial dimension; is the frequency dimension; is the environmental dimension.

3. A method for predicting the lifespan of a deep-sea watertight connector according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.

1. Perform Z-score standardization on the data in each dimension respectively: ； wherein, and are the mean and standard deviation of the th environmental parameter respectively, is a fourth-order tensor dataset, is the after standardization; The scalar form formula for the fourth-order parallel factor molecular tensor decomposition is: ； Among them, is the number of factors, is the th time factor, is the th spatial factor, is the th frequency factor, is the th environmental factor, represents the error set of the fourth-order tensor; S2.2, input the standardized tensor data set into the Parallel Factor Analysis (PARAFAC) module to decompose , , , into four matrices. Define the time evolution factor matrix as a × dimensional matrix, being the time factor; define the spatial distribution factor matrix as a × dimensional matrix, being the spatial factor; define the frequency domain feature factor matrix as a × dimensional matrix, being the frequency factor; define the environmental coupling factor matrix as a × dimensional matrix, being the environmental factor; S2.3, calculate , , and , to calculate, taking the method as an example, the remaining matrix calculation methods are the same as : ； Among them, means that the column vectors of , , matrix are block - diagonalized and combined, and the error between the observed data and the model predicted value is represented by . The least - squares estimate of is expressed by the following formula: ； wherein, represents the generalized inverse; S2.4, Repeated calculation , , and , until the result converges or the set number of iterations is reached and then stop running, and .

4. A method for predicting the lifespan of a deep-sea watertight connector according to claim 3, characterized in that, In step S3, the specific process of extracting multi-scale features in the dilated convolution module is: The result obtained from step 2 , , and are concatenated into a feature tensor , and is input into the dilated convolution module of the multi-scale sparse neural network MSNN to extract multi-scale features and output a multi-scale feature map: ； ； Among them, is the convolution kernel size, the dilation rate takes values of 1, 2, and 4, is the convolution kernel weight at the current position factor, is the multi-scale feature map, is the sequence length, is the number of channels.

5. A method for predicting the service life of a deep-sea watertight connector according to claim 4, characterized in that, In step S3, the specific steps of extracting features with different granularities through the attention gating mechanism specifically include the following steps: S3.1, concatenate the multi-scale feature maps output by the dilated convolution module and assign weights through the attention gating mechanism : ； Among them, is a learnable weight matrix, is a bias term; S3.

2. Fuse the weighted multi-scale features: ； Among them, is the fused multi-scale feature map; S3.3, input the input residual connection into the non-linear projection module and output the fused feature : ； ； Among them, is the feature vector after residual connection, is a learnable projection matrix, is the column projection vector of the linearly combined cascaded features, is a cascading operator, is a bias, .

6. A method for predicting the lifespan of a deep-sea watertight connector according to claim 1, characterized in that Step S4 specifically includes the following steps: S4.1, Take the fused features output by MSNN as the input signal, and decompose them into multiple intrinsic mode functions through the CEEMDAN algorithm and the residual term , as follows: ; Among them, is the characteristic number, is the time; S4.2, extract the energy of each IMF and calculate the energy entropy value : ; S4.3, input multiple IMF energies into KPCA for dimensionality reduction, and perform principal component analysis by mapping through a kernel function to a high-dimensional space to obtain time-frequency features and spatial environment features: ； ； Among them, is an element in the kernel matrix; represents the energy of the IMF, is the eigenvector of the kernel matrix, is the projection matrix, is the principal component score; S4.

4. Extract the global relationship between the time-frequency features and the spatial environment features through the multi-head attention module, and then input them into the CBAM attention mechanism module; The overall process of CBAM is expressed as: ； ； Among them, represents the Hadamard product, is the CAM module, is the SAM module, represents the input feature, represents the feature map obtained through the CAM module, represents the feature map obtained through the SAM module; S4.6, concatenate , and with weighting, and map them to a health indicator via a fully connected layer: ； Among them, and are learnable parameters, is the Sigmoid function, is the concatenation operation.

7. A method for predicting the lifespan of a deep-sea watertight connector according to claim 1, characterized in that, Step S5 specifically includes the following steps: S5.1, combine the health indicators with environmental parameters, where the environmental parameters include: pressure , temperature , salinity , and construct an input feature matrix : ； Among them, , , , , after merging ; S5.2, generating pseudo-labels corresponding to the target domain samples using the fuzzy C-means clustering algorithm , the objective function of FCM is as follows: ； Among them, 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 index, is the sample of the target domain in S5.3, by introducing Lagrange multipliers to construct the Lagrangian function , as follows: ； S5.4, respectively for the membership degree , the cluster center and the Lagrange multiplier take the partial derivatives and set the partial derivatives to zero to obtain the cluster center of the th iteration and the membership degree update formula: ； 。 8. A method for predicting the lifespan of a deep-sea watertight 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 , where is the source domain data with true labels, is the target domain data with pseudo labels, and the objective function of the causal representation learning module includes: reconstruction loss, causal structure learning loss, and cross-entropy loss: ； Among them, is the number of samples, represents the number of samples of is the reconstructed output of To control the strength of causal constraints, To balance the classification loss, is the regularization weight decay; represents the number of hidden layers; and respectively represent the cross-entropy loss and the classifier; represents the label merged Markov blanket; and respectively represent the weight parameters of the th hidden layer in the encoding and decoding processes; and respectively represent the outputs of the encoder and decoder of the causal graph autoencoder: ； ； wherein, is the dimension, represents mapping the original input to low-dimensional features, represents causal logic and label fusion, represents the decoder network, is the adjacency matrix; S5.6, to ensure the acyclicity of the corresponding DAG, the following acyclicity constraints need to be satisfied : ； Among them, represents the trace of a matrix, represents the matrix exponential of; is the Hadamard product; represents the number of variables in the causal graph model; S5.

7. After introducing the Lagrange multiplier and the penalty term, the augmented Lagrangian is given: ； Among them, represents the Lagrange multiplier, represents the penalty parameter.

9. A method for predicting the lifespan of a deep-sea watertight connector according to claim 8, characterized in that, Step S5 also includes the following steps: S5.8, for the adjacency matrix and the following update rules apply: ； ； ； Among them, and are two adjustable hyperparameters, is the parameter value when taking the minimum value of the function; S5.9, use , and to represent the network model parameters in the feature extractor , label classifier and domain discriminator . The objective function of the domain alignment module is: ； Among them, is a feature extractor, is a label classifier, and is a domain discriminator: and respectively represent the cross-entropy loss functions for fault classification and domain classification; and respectively represent the domain labels of the source domain data and the target domain data; is a balance parameter, represents a gradient reversal layer, represents the true label of the source domain data; S5.10, Parameter The update rule is as follows: ； ； ； Among them, represents the learning rate, and respectively represent the fault classification loss and the domain discrimination loss: ； 。 10. The lifespan prediction method of a deep - sea watertight connector according to claim 1, characterized in that, Step S6 specifically includes the following steps: S6.1, considering the influence of environmental parameters, introduce a non - linear drift term and a diffusion term , and construct a non - linear Wiener process modulated by environmental parameters: ； where the drift term reflects the acceleration effect of environmental parameters on degradation and adopts the generalized Arrhenius model: ； Diffusion term Characterize the random degradation noise caused by environmental fluctuations: ； Among them, is the degradation index, indicating the cumulative degradation amount of the watertight connector at time ; is the standard Brownian motion; is the initial degradation amount; are all parameters to be determined; S6.2, using the accelerated degradation test data, through the maximum likelihood estimation function estimate the model parameters : ； ； Among them, , , is an exponential function; are the temperature, pressure, and salinity of the th path respectively, is the degradation index of the th path; S6.

3. Adopt Monte Carlo path simulation to generate multiple degradation paths: ； Among them, , an independent standard normal distribution random variable, introduces random fluctuations; is the time step; S6.4, Assume the prior distribution of the parameters , and calculate the posterior distribution by combining the real-time monitoring data: ； Among them, is the observed degradation data, is proportional to; S6.5, Define the failure threshold , Calculate the remaining useful life RUL: ； wherein, is the time increment, is the predicted degradation amount of the z-th path at time , and is to take the minimum.

Images