Corrosion evaluation method of marine electrical aluminum alloy based on cnn-svm and dose-response function

Abstract

This invention discloses a corrosion evaluation method for marine electrical equipment aluminum alloys based on CNN-SVM and dose-response function, comprising: obtaining an initial feature set; eliminating dimensional differences of different parameters to obtain preprocessed data; inputting the preprocessed data into a pre-trained CNN-SVM hybrid regression prediction model to obtain the aluminum alloy corrosion weight loss result corresponding to the current operating condition; constructing a nonlinear dose-response function; using the obtained corrosion weight loss result as supervised data, fitting and optimizing the nonlinear dose-response function; substituting the annual service time into the fitted and optimized nonlinear dose-response function to obtain the first-year corrosion weight loss; converting the first-year corrosion weight loss into the annual corrosion rate; and obtaining the corresponding corrosion level by referring to the international standard classification table. This invention fills the technical and standard gap in corrosion evaluation under energized operating conditions, can provide core support for ISO standard updates, and is applicable to the anti-corrosion design and life assessment of aluminum alloy components in marine electrical equipment.

Description

Technical Field

[0001] This invention relates to the field of aluminum alloy corrosion prediction technology, and in particular to a corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function. Background Technology

[0002] As offshore wind power projects expand into deeper and deeper waters, the construction of coastal overhead stranded cables, as the core channel for power transmission, is booming. 1050 aluminum alloy, due to its low density, excellent conductivity, stable mechanical properties, and ease of processing and forming, has become the core material for coastal overhead stranded cables, and market demand continues to surge.

[0003] The marine atmospheric environment itself has strong corrosive characteristics due to high humidity, high chloride ion concentration, and frequent temperature fluctuations. During the service of the overhead stranded cable at the seaside, there is also the coupling effect of electric field, current and marine atmospheric environment, forming a multi-factor coupled corrosion environment of temperature, humidity, chloride ion and electrical effect. This makes the corrosion behavior of aluminum alloy exhibit complex nonlinear characteristics. Its corrosion law is far more complex than that of conventional marine structural components, which directly determines the service safety and service life of marine electrical equipment.

[0004] Currently, scholars both domestically and internationally have conducted extensive research on the marine atmospheric corrosion of aluminum alloys. At the corrosion mechanism level, it has been clarified that the synergistic effect of chloride ions and acidic pollutants is the core factor exacerbating aluminum alloy corrosion, and it has also been confirmed that energized conditions significantly accelerate the corrosion process. Regarding corrosion prediction and evaluation, existing studies have introduced machine learning algorithms such as CNN and SVM to improve prediction accuracy by exploring the nonlinear correlation between environmental factors and corrosion behavior. The ISO 9223 and ISO 9226 standards also provide a framework for atmospheric corrosivity classification, offering a reference for the selection of aluminum alloy materials under normal operating conditions.

[0005] The existing patent (CN121075477A) discloses a CNN-SVM-based method for predicting atmospheric corrosion of aluminum alloys. However, it only targets conventional atmospheric corrosion conditions without electricity. It uses environmental meteorological data and aluminum alloy composition characteristics as inputs and adopts a single-branch general CNN architecture. It does not make customized designs for the differences in the electric-environment coupling characteristics under energized conditions, and cannot capture the multi-order coupling characteristics of electric effects and environmental factors. The prediction accuracy drops sharply and it completely fails under energized coupling conditions. At the same time, this scheme can only achieve black-box prediction of corrosion rate. It does not form a closed-loop optimization by extracting model features and constructing a mechanistic interpretation dose-response function, and cannot provide a quantitative basis for the anti-corrosion design of marine electrical equipment.

[0006] Existing technologies for research and evaluation of coupled corrosion under special operating conditions in marine electrical engineering still have core gaps, specifically limited in three aspects: First, existing research and evaluation systems generally ignore the multi-factor coupling effect between electrical effects and the marine atmospheric environment, failing to systematically explore the synergistic mechanism of electric field, current, temperature, humidity, and chloride ions. Existing ISO standards and corrosion classification systems do not incorporate the influence of electrical effects, resulting in a lack of quantitative basis for the corrosion protection design of aluminum alloy components and a lack of quantification in lifespan assessment. Second, there is a long-standing technical bias in this field: it is generally believed that the corrosion of aluminum alloys in offshore overhead stranded cables is solely dominated by the marine atmospheric environment, with current and electric field being merely secondary conditions of service conditions, not requiring inclusion as independent core characteristics in the evaluation system. Furthermore, the existence of irreplaceable differential corrosion mechanisms between these two factors, as well as the third-order synergistic coupling effect with chloride ions, temperature, and humidity, has been overlooked, resulting in a persistent bottleneck in prediction under energized operating conditions. Secondly, existing corrosion prediction models are mostly general-purpose architectures that lack customized design for the electrical-environment coupling characteristics. Their prediction accuracy drops sharply under energized conditions, their generalization ability is insufficient, and they exhibit system biases between indoor and outdoor scenarios, making them unsuitable for actual service conditions. General-purpose CNN-SVM models can only achieve black-box prediction; feature extraction lacks physical mechanism support and cannot provide a scientific basis for constructing mechanistic explanation models. Thirdly, the classic dose-response function of ISO 9223 is only suitable for unenergized, conventional environments. Under energized conditions, the prediction error exceeds 90%, the coefficient of determination R² is negative, and it completely loses its predictive effectiveness. Furthermore, it is only suitable for the first year of corrosion assessment and cannot predict the corrosion life of marine electrical equipment over its long service life. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, the corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function provided by this invention solves the problems of existing aluminum alloy corrosion evaluation methods lacking a coupled evaluation system of electrical effects and marine atmospheric factors under special working conditions of marine electrical engineering, insufficient prediction accuracy of existing models, inability to quantify the coupling effect of multiple factors, and lack of long-term prediction capability.

[0008] To achieve the aforementioned objectives, the technical solution adopted by this invention is: a corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function, comprising: We acquire marine atmospheric environment parameters, electrical effect characteristics parameters during energized service, and corrosion exposure time parameters of overhead stranded lines under service conditions, forming an initial feature set covering all coupled influencing factors; The normalization method is used to map all feature parameters in the initial feature set to the interval [0,1] to eliminate the difference in the units of different parameters and obtain the preprocessed data; The preprocessed data is input into the pre-trained CNN-SVM hybrid regression prediction model to obtain the aluminum alloy corrosion weight loss result corresponding to the current working condition; Construct a nonlinear dose-response function; The corrosion weight loss results predicted by the pre-trained CNN-SVM hybrid regression prediction model were used as supervised data to fit and optimize the nonlinear dose-response function; Substituting the annual service time into the fitted and optimized nonlinear dose-response function, the corrosion weight loss in the first year is obtained; the corrosion weight loss in the first year is then converted into the annual corrosion rate. Based on the obtained annual corrosion rate, the corresponding corrosion level is determined by referring to the international standard classification table.

[0009] The beneficial effects of this invention are as follows: 1. By constructing a customized dual-branch CNN-SVM model, ultra-high accuracy prediction under energized conditions is achieved, fundamentally different from existing general-purpose architectures. This invention, targeting the differentiated characteristics of electricity-environment coupling features, pioneers a dual-branch differentiated convolutional architecture. It customizes the convolutional kernel size, network layer number, and activation function for slow time-varying environmental features and fast-response electrical effect features respectively, overcoming the limitation of general-purpose single-branch CNNs in simultaneously adapting to two types of heterogeneous features. By adding a multi-level coupling feature interaction module, the synergistic coupling effect of multiple factors is accurately quantified. Simultaneously, a closed-loop mechanism of "ranking importance screening - channel attention weight optimization" is constructed, providing physical mechanism support for feature extraction. This eliminates system bias in indoor and outdoor scenes, and compared to existing CNN-SVMs, solves the industry pain point of general-purpose models completely failing under energized coupling conditions.

[0010] 2. The constructed dose-response function breaks through the limitations of traditional models in full-cycle prediction and scenario boundaries, achieving a qualitative leap in core performance. This invention is the first to incorporate the coupling effects of current and electric field into the dose-response function system, constructing a nonlinear function with clear corrosion electrochemical mechanism support. By introducing a time power correction term and an electrical effect coupling interaction term, it overcomes the limitation of the traditional ISO 9223 dose-response function being only applicable to the first year of corrosion assessment, achieving high-precision quantification of short- to long-cycle corrosion from 10 days to 2000 days.

[0011] 3. By constructing a full-chain innovation system of "data-driven feature extraction - mechanism interpretation function construction - standard adaptation and hierarchical evaluation", it is not only applicable to the target 1050 aluminum alloy, but also to typical electrical metal materials such as Cu, Ag, and Sn, demonstrating outstanding cross-material generalization ability. The corrosion classification system based on ISO general standards can be applied to the corrosion protection design and life assessment of various marine electrical equipment such as coastal overhead stranded cables, offshore wind power converters, and submarine cable terminals. It has strong generalization ability and engineering applicability, and has replicable technical promotion value, providing a replicable technical path for the coupled corrosion evaluation of similar metal materials. Attached Figure Description

[0012] Figure 1 A flowchart of a corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function is provided for the embodiments. Figure 2 This is a graph showing the comparison between the predicted and actual values ​​of the CNN-SVM model's training and test sets provided by this invention. Figure 3 A scatter plot of predicted and measured values ​​for the dose-response function; Figure 4 This is a combined plot of the relative error frequency distribution and cumulative distribution of the dose-response function; Figure 5 The dose-response function constructed for this invention and the MRE and R of the conventional function 2 Performance comparison chart. Detailed Implementation

[0013] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0014] like Figure 1 As shown, in one embodiment of the present invention, a corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function includes: S1. Obtain marine atmospheric environment parameters, electrical effect characteristic parameters during energized service, and corrosion exposure time parameters of overhead stranded lines under service conditions, forming an initial feature set covering all coupled influencing factors.

[0015] The initial feature set is divided into three categories: the first is core marine atmospheric environmental parameters, including temperature T, relative humidity RH, and chloride ion deposition rate S. d and sulfation rate P d The electrical effect characteristic parameters during energized service include current I and electric field strength E; the third is the corrosion exposure time parameter t, which provides a time dimension basis for the quantitative calculation of corrosion weight loss over the entire cycle.

[0016] In this embodiment, the current I ranges from 0 to 4 A / mm², and the electric field strength E ranges from 0 to 300 kV / m, fully covering the actual energized service conditions of the overhead stranded line by the sea. The current I drives the corrosion and dissolution of the aluminum alloy through anodic polarization, and the electric field strength E modifies the corrosion process by changing the migration behavior of corrosion ions at the interface. The two are independent core characteristic parameters and cannot be substituted for each other.

[0017] S2. The normalization method is used to map all feature parameters in the initial feature set to the interval [0,1] to eliminate the difference in the dimensions of different parameters and obtain the preprocessed data.

[0018] S3. Input the preprocessed data into the pre-trained CNN-SVM hybrid regression prediction model to obtain the aluminum alloy corrosion weight loss result corresponding to the current working condition.

[0019] The front end of the CNN-SVM hybrid regression prediction model uses a dual-branch CNN network to extract differentiated high-dimensional nonlinear features from environmental features and electrical effect features, which are then fused by a gated multi-level coupled feature interaction module and weighted by a CBAM hybrid attention module.

[0020] The backend uses an ε-SVR type VM model with a customized weighted RBF kernel function to complete the corrosion weight loss regression prediction. The high-dimensional fusion features output by the CNN are used as input and the corrosion weight loss normalization label is used as output. Weighting coefficients are set according to the power coupling condition. A dual constraint is formed by kernel function weighting and loss function amplification. The hyperparameters are optimized by combining grid search and 5-fold cross-validation to complete the training. Finally, the model performance is evaluated by MRE, MAE, and R².

[0021] Specifically, the CNN-SVM hybrid regression prediction model includes a two-branch CNN network, a 1×1 convolutional layer, an L2 normalization layer, a gated multi-level coupled feature interaction module, a splicing layer, a CBAM hybrid attention module, and a back-end SVM model connected in sequence; the output of the two-branch CNN network serves as another input to the splicing layer.

[0022] The dual-branch CNN network includes an environmental feature branch and an electrical effect feature branch. The environmental feature branch consists of two convolutional layers with a kernel size of 3×1 and a stride of 1, coupled with a ReLU activation function, connected in sequence, and a dropout layer with a dropout rate of 0.1, used to capture the long-term cumulative effect features of corrosion.

[0023] The electrical effect feature branch consists of a convolutional layer with a kernel size of 3×1 and a stride of 1, coupled with a Leaky-ReLU activation function, and a dropout layer with a dropout rate of 0.1, connected sequentially. This is used to capture local corrosion feature changes caused by abrupt changes in electrical parameters. Both branch convolutional layers are followed by dropout layers with a dropout rate of 0.1 to prevent overfitting of the model.

[0024] The gated multi-level coupled feature interaction module adopts a fusion architecture of decomposition machine and gated convolution; the processing methods of the gated multi-level coupled feature interaction module for input features include: A latent vector is assigned to each input feature by a decomposition machine; the inner product of the latent vectors corresponding to the input features is calculated to generate second-order coupled features; a three-branch tensor product operation is performed on the latent vectors corresponding to the input features to generate third-order coupled features; the weights of the second-order and third-order coupled features are learned by sigmoid-gated convolution to set a threshold to filter out noisy coupling terms, and the second-order effective features and the third-order effective coupled features are output.

[0025] Specifically, the high-dimensional single features output by the dual branches first undergo dimension alignment and preprocessing: the two types of heterogeneous features are uniformly mapped to a 32-dimensional feature space through 1×1 convolution, and L2 normalization is performed simultaneously to eliminate the difference in feature amplitude. Then, the features are input into the gated multi-order coupled feature interaction module. This module adopts a fusion architecture of decomposer and gated convolution. Second-order coupled features are generated through the inner product of latent vectors of the decomposer. Third-order coupled features are generated through the three-branch tensor product operation. Feature combinations related to the corrosion mechanism are retained throughout the process. The weights of the coupled features are learned through Sigmoid gated convolution. Noisy coupled terms are filtered out with a weight ≥ 0.3 as a threshold, and second-order and third-order effective coupled features are output.

[0026] After concatenating the original single features of the two branches with the effectively coupled features, the concatenation is input into the CBAM hybrid attention module: First, the channel attention unit performs global average pooling and global max pooling on the features simultaneously to complete channel-dimensional weighted fusion; then, the spatial attention unit performs double pooling along the channel dimension and concatenates to generate spatial feature descriptors. The spatial attention weights are learned through 3×1 convolution to focus on strengthening the local corrosion feature regions corresponding to abrupt changes in electrical effects and sudden increases in chloride ion concentration, thus completing spatial dimension feature enhancement; finally, the features with dual channel-spatial weighting are global average pooled and flattened into a one-dimensional feature vector, which is then output to the backend SVM model.

[0027] The backend SVM model is an ε-SVR type VM model that includes a customized weighted RBF kernel function; The customized weighted RBF kernel function is specifically as follows:

[0028] In the formula, For weighted RBF kernel functions; , , are the high-dimensional feature fusion vectors output by the two branches of the dual-branch CNN network; exp represents the exponential function with the natural constant e as the base; γ is the kernel function bandwidth parameter that controls the local feature fitting ability; Denotes the Euclidean norm; , All are weighting coefficients for the electrical effect operating conditions, with values ​​assigned according to the operating condition level; they are 1, 1.5, or 2 depending on the operating condition level. By embedding the weighting coefficients to change the high-dimensional spatial distance metric of the samples, the sample spacing of the energized operating conditions is reduced, and its influence on the regression hyperplane is strengthened, achieving targeted enhancement of the energized operating conditions from the kernel function level.

[0029] During model training, the weighting coefficients for the electrical effect condition are enhanced for the energized coupling condition, specifically as follows: First, based on the full service operating range of overhead stranded lines at the seaside (0~4A / mm² current density, 0~300kV / m electric field strength), the training set samples were classified into three levels: normal operating condition (current <0.5A / mm² and electric field <50kV / m), medium operating condition (0.5A / mm² ≤ current <2A / mm² or 50kV / m ≤ electric field <150kV / m), and extreme operating condition (current ≥2A / mm² or electric field ≥150kV / m). Corresponding settings were then applied. , Differential weighting coefficients of 1, 1.5, and 2 are used to specifically enhance the model contribution of samples under electrically coupled operating conditions.

[0030] The specific method for training a CNN-SVM hybrid regression prediction model includes two stages: primary optimization and secondary optimization. One of the optimizations specifically includes: Initial feature set data and corresponding aluminum alloy weight loss measurement results under different coupling conditions were collected to obtain a sample dataset, including indoor accelerated corrosion experimental data and on-site electrostatic exposure experimental data at the seaside. The indoor accelerated corrosion experiments included constant current accelerated corrosion experiments and DC electric field accelerated corrosion experiments. Each sample in the initial feature set data contained complete 7 feature parameters and corresponding 1050 aluminum alloy corrosion weight loss measurement labels. In this embodiment, initial feature set data and corresponding aluminum alloy corrosion weight loss measurement results under different coupling conditions were collected, including 34 indoor accelerated corrosion experimental data covering multiple factors such as temperature and humidity, chloride ions, current, and electric field coupling, and 38 on-site electrostatic exposure experimental data at the seaside, constructing a total of 72 sets of original aluminum alloy coupled corrosion datasets.

[0031] The Min-Max normalization method is used to map all feature parameters in the initial feature set data to the [0,1] interval, eliminating the difference in the units of different parameters and obtaining a standardized dataset; The standardized dataset was divided into a training set and a test set in a 7:3 ratio. The division process used random stratified sampling to ensure the consistency of the current and electric field gradient distribution between the training set and the test set. The feature parameters in the training set are input into the CNN-SVM hybrid regression prediction model to obtain the corresponding aluminum alloy corrosion weight loss prediction results. Based on the prediction results of aluminum alloy corrosion weight loss, the hyperparameters of the CNN-SVM hybrid regression prediction model were optimized by combining grid search and five-fold cross-validation method to complete the first optimization; finally, the model performance was evaluated by MRE, MAE, and R².

[0032] The specific training of the backend SVM model is as follows: taking the high-dimensional fusion features output by the CNN module as input and the erosion weight loss normalization label as output, minimizing the weighted regression loss as the objective function, and using a grid search combined with 5-fold hierarchical cross-validation, the optimal combination of the penalty coefficient c and the kernel function bandwidth γ is globally optimized within the exponential interval [-10, 10]. In each round of cross-validation, by setting sample weight parameters, the prediction loss of the energized coupling condition samples is amplified by weight, forcing the model to prioritize fitting the corrosion law of the energized condition. This allows the weighting enhancement effect of the kernel function and the weight amplification of the loss function to form a dual constraint, completely solving the pain point of the sharp drop in prediction accuracy of the general model under the energized condition, and finally achieving high-precision regression prediction of erosion weight loss.

[0033] The 5-fold cross-validation employs a stratified random sampling strategy to ensure that each fold's training and validation sets contain full-condition samples with different current and electric field gradients. The model performance criteria are: goodness of fit R² ≥ 99% between the training and test sets, mean relative error (MRE) ≤ 10%, and mean absolute error (MAE) ≤ 20%. The curves comparing the predicted and actual values ​​of the CNN-SVM model's training and test sets are shown below. Figure 2 As shown.

[0034] Secondary optimization specifically includes: Based on the optimized CNN-SVM hybrid regression prediction model, the contribution of each feature parameter in the initial feature set to the corrosion weight loss of aluminum alloy is quantified. Feature parameters with positive importance values ​​are selected as core corrosion influencing factors. Noise interference terms with negative importance values ​​are removed to address the difficulty of interpreting complex relationships between features in CNN-SVM. The selected core corrosion influencing factors are fed back into the CNN-SVM hybrid regression prediction model for secondary optimization, resulting in a pre-trained CNN-SVM hybrid regression prediction model.

[0035] In this embodiment, the core corrosion influencing factors selected based on marine atmospheric environmental parameters and electrical effect characteristics during energized service include temperature, current value, chloride ion deposition rate, relative humidity, and electric field strength; the noise interference item removed is the sulfation rate.

[0036] The contribution of each feature parameter in the initial feature set to the corrosion weight loss of aluminum alloy was quantified, and the feature parameters with positive importance values ​​were selected as the core influencing factors of corrosion, including: Step A1: Calculate the baseline root mean square error RMSE1 of the optimized CNN-SVM hybrid regression prediction model on the test set; Step A2: Shuffle the value order of individual feature parameters, input the shuffled feature parameters into the optimized CNN-SVM hybrid regression prediction model, and obtain the prediction result after shuffling the feature parameters; Step A3: Calculate the baseline root mean square error RMSE2 of the CNN-SVM hybrid regression prediction model based on the shuffled prediction results; Step A4: Repeat steps A2 and A3 multiple times for the same feature parameter, using the difference between RMSE2 and RMSE1 as the ranking importance quantification index, and calculate the average value of the ranking importance quantification index multiple times; Step A5: Select the feature parameters whose average value of the quantitative indicators of importance is positive as the core influencing factors of corrosion.

[0037] S4. Construct a nonlinear dose-response function.

[0038] Based on the ISO 9223 atmospheric corrosion assessment standard as the core framework, and combined with the screened core corrosion influencing factors and corrosion time parameters, a nonlinear mechanism function for the corrosion weight loss of 1050 aluminum alloy was constructed, incorporating the coupling effects of current and electric field and adapting to long-term predictions of 10d to 2000d. The function structure closely matches the physical nature of material corrosion, accurately matching the power-law and Arrhenius-type variation laws of corrosion behavior with electrochemical coupling effects. The mechanistic expression of the nonlinear dose-response function is as follows:

[0039] In the formula, Y represents the corrosion weight loss, in g / m²; S d RH is the chloride ion deposition rate, in mg / (100cm²·d); t is the relative humidity, in %; T is the corrosion time, in days; T is the ambient temperature, in °C; I is the current density, in A / mm². 2 E represents the electric field strength, in kV / m; a, b, c, d, e, f, g, h, and k are all parameters to be fitted.

[0040] S5. Using the corrosion weight loss results predicted by the pre-trained CNN-SVM hybrid regression prediction model as supervised data, fit and optimize the nonlinear dose-response function.

[0041] Fitting and optimizing the nonlinear dose-response function, including: Based on 72 sets of actual corrosion weight loss data of 1050 aluminum alloy under multiple coupled working conditions, after 3 rounds of iterative removal of outliers, 62 sets of valid samples were retained and divided into training set and test set in a 7:3 ratio. With the goal of minimizing the relative prediction error, a weighted nonlinear least squares method is used to construct the objective function, and an L2 regularization term is introduced into the objective function. The obtained core factors are then used as feature parameters in the dose-response function to complete the final model iteration.

[0042] The optimal solution for the parameters to be fitted is found by a two-stage global optimization strategy that combines coarse search and fine search.

[0043] The nonlinear dose-response function is applicable for periods ranging from 10 days to 2000 days, covering the entire corrosion prediction scenario from short-term accelerated corrosion to ultra-long-term service.

[0044] The prediction results of the quadratically optimized CNN-SVM model were used as the supervision data for fitting, and the function fitting error was used to back-verify the model's feature extraction effect, thus completing the bidirectional iterative optimization of the model and the function. A scatter plot of the predicted and measured values ​​of the dose-response function is shown below. Figure 3 As shown, the scatter plot of the dose-response function between predicted and measured values ​​is as follows: Figure 4 As shown, the coefficient of determination R² for the fitted dose-response function is 0.9632 for the training set and 0.9203 for the test set.

[0045] In this embodiment, the mechanistic expression of the fitted nonlinear dose-response function is: .

[0046] The fitted dose-response function needs to pass the unbiasedness test of the residual sequence and the normality test of the residual distribution. The median of the relative error of the full sample prediction is 11.71%, and the mean error is 17.48%, to ensure the statistical validity and engineering applicability of the model.

[0047] S6. Substitute the annual service time into the fitted and optimized nonlinear dose-response function to obtain the first year's corrosion weight loss; convert the first year's corrosion weight loss into the annual corrosion rate.

[0048] S7. Based on the obtained annual corrosion rate, refer to the international standard classification table to obtain the corresponding corrosion level.

[0049] The atmospheric corrosion classification system for aluminum alloys strictly follows the ISO 9226 standard framework. It substitutes the corrosion exposure time t=365d into the dose-response function, and quantitatively calculates the predicted value of the first-year corrosion weight loss of 1050 aluminum alloy through environmental and operating condition parameters. After converting it into the annual corrosion rate, it matches 6 corrosion levels from C1 to CX, which can realize one-stop determination of corrosion level under special working conditions of marine electrical engineering.

[0050] As verified by actual testing, such as Figure 5As shown, the R² of the function in this invention reaches 0.9632 on the training set and 0.9203 on the test set. The average relative error of the full sample is only 17.48%, and the median error is as low as 11.71%. Under electrically coupled conditions, the prediction error of the function in this invention is only 19.0%, while the error of the traditional ISO 9223 model is as high as 90.1%, completely losing its predictive effectiveness. It has achieved a qualitative leap in both prediction accuracy and applicability to operating conditions, and can provide core quantitative support for the update of ISO standards.

[0051] This invention breaks through long-standing technical biases in the field and fills the technical and standard gap in the evaluation of coupled corrosion under special operating conditions of marine electrical engineering. For the first time, this invention incorporates electrical effects such as electric field and current as independent core characteristic systems into the corrosion evaluation system of aluminum alloys, clarifying the differentiated corrosion mechanisms that the two cannot be substituted for each other, as well as the third-order synergistic coupling effect with chloride ions, temperature, and humidity. This breaks the industry's technical bias that "electrical effects are merely ancillary conditions for service." The constructed corrosion prediction and classification method adapted to the energized service conditions of offshore overhead stranded cables solves the core problem of existing technologies failing to consider the coupling effect of multiple factors between electricity and the environment. It overcomes the limitation that existing ISO standards are only applicable to non-electrical operating conditions and can provide core technical support for the updating of ISO standards for metal corrosion systems under energized coupled environments.

Claims

1. A corrosion evaluation method for marine electrical aluminum alloys based on CNN-SVM and dose-response function, characterized in that, include: We acquire marine atmospheric environment parameters, electrical effect characteristics parameters during energized service, and corrosion exposure time parameters of overhead stranded lines under service conditions, forming an initial feature set covering all coupled influencing factors; The normalization method is used to map all feature parameters in the initial feature set to the interval [0,1] to eliminate the difference in the units of different parameters and obtain the preprocessed data; The preprocessed data is input into the pre-trained CNN-SVM hybrid regression prediction model to obtain the aluminum alloy corrosion weight loss result corresponding to the current working condition; Construct a nonlinear dose-response function; The corrosion weight loss results predicted by the pre-trained CNN-SVM hybrid regression prediction model were used as supervised data to fit and optimize the nonlinear dose-response function; Substituting the annual service time into the fitted and optimized nonlinear dose-response function, the corrosion weight loss in the first year is obtained; the corrosion weight loss in the first year is then converted into the annual corrosion rate. Based on the obtained annual corrosion rate, the corresponding corrosion level is determined by referring to the international standard classification table.

2. The method according to claim 1, characterized in that, Marine atmospheric environmental parameters include temperature, relative humidity, chloride ion deposition rate, and sulfation rate; electrical effect characteristic parameters during energized operation include current and electric field strength.

3. The method according to claim 1, characterized in that, The CNN-SVM hybrid regression prediction model consists of a two-branch CNN network, a 1×1 convolutional layer, an L2 normalization layer, a gated multi-level coupled feature interaction module, a concatenation layer, a CBAM hybrid attention module, and a back-end SVM model connected in sequence; the output of the two-branch CNN network serves as another input to the concatenation layer; The dual-branch CNN network includes an environmental feature branch and an electrical effect feature branch. The environmental feature branch consists of two convolutional layers with a kernel size of 3×1 and a stride of 1, coupled with a ReLU activation function, connected in sequence, and a dropout layer with a dropout rate of 0.

1. The electrical effect feature branch consists of a convolutional layer with a kernel size of 3×1, a stride of 1, and a Leaky-ReLU activation function, connected in sequence, and a dropout layer with a dropout rate of 0.

1.

4. The method according to claim 3, characterized in that, The gated multi-level coupled feature interaction module adopts a fusion architecture of decomposition machine and gated convolution; The gated multi-level coupled feature interaction module processes input features using the following methods: A latent vector is assigned to each input feature using a decomposition machine; Calculate the inner product of the latent vectors corresponding to the input features to generate second-order coupled features; Perform a three-branch tensor product operation on the latent vectors corresponding to the input features to generate third-order coupled features; We learn the weights of second-order and third-order coupled features through sigmoid-gated convolution, set a threshold to filter out noisy coupled terms, and output second-order effective features and third-order effective coupled features.

5. The method according to claim 3, characterized in that, The backend SVM model is an ε-SVR type VM model that includes a customized weighted RBF kernel function; The customized weighted RBF kernel function is specifically as follows: In the formula, For weighted RBF kernel functions; , , are the high-dimensional feature fusion vectors output by the two branches of the dual-branch CNN network; exp represents the exponential function with the natural constant e as the base; γ is the kernel function bandwidth parameter that controls the local feature fitting ability; , All are weighting coefficients for electrical effect operating conditions, and are assigned values ​​according to the operating condition level; This represents the Euclidean norm.

6. The method according to claim 1, characterized in that, The specific method for training a CNN-SVM hybrid regression prediction model includes two stages: primary optimization and secondary optimization. One of the optimizations specifically includes: We collected initial feature set data and corresponding aluminum alloy weight loss test results under different coupling conditions to obtain sample datasets, including indoor accelerated corrosion test data and on-site electrical exposure test data at the seaside. The Min-Max normalization method is used to map all feature parameters in the initial feature set data to the [0,1] interval, eliminating the difference in the units of different parameters and obtaining a standardized dataset; The standard dataset is divided into a training set and a test set using random stratified sampling. The feature parameters in the training set are input into the CNN-SVM hybrid regression prediction model to obtain the corresponding aluminum alloy corrosion weight loss prediction results. Based on the prediction results of aluminum alloy corrosion weight loss, the hyperparameters of the CNN-SVM hybrid regression prediction model are optimized by combining grid search and five-fold cross-validation method, and the first optimization is completed. Secondary optimization specifically includes: Based on the CNN-SVM hybrid regression prediction model that has been optimized once, the contribution of each feature parameter in the initial feature set to the corrosion weight loss of aluminum alloy is quantified, and the feature parameters with positive importance values ​​are selected as the core corrosion influencing factors. Noise interference terms with negative importance values ​​are removed. The selected core corrosion influencing factors are fed back into the CNN-SVM hybrid regression prediction model to complete the second optimization and obtain the pre-trained CNN-SVM hybrid regression prediction model.

7. The method according to claim 6, characterized in that, The contribution of each feature parameter in the initial feature set to the corrosion weight loss of aluminum alloy was quantified, and the feature parameters with positive importance values ​​were selected as the core influencing factors of corrosion, including: Step A1: Calculate the baseline root mean square error RMSE1 of the optimized CNN-SVM hybrid regression prediction model on the test set; Step A2: Shuffle the value order of individual feature parameters, input the shuffled feature parameters into the optimized CNN-SVM hybrid regression prediction model, and obtain the prediction result after shuffling the feature parameters; Step A3: Calculate the baseline root mean square error RMSE2 of the CNN-SVM hybrid regression prediction model based on the prediction results after shuffling the feature parameters; Step A4: Repeat steps A2 and A3 multiple times for the same feature parameter, using the difference between RMSE2 and RMSE1 as the ranking importance quantification index, and calculate the average value of the ranking importance quantification index multiple times; Step A5: Select the feature parameters whose average value of the quantitative indicators of importance is positive as the core influencing factors of corrosion.

8. The method according to claim 7, characterized in that, Based on the marine atmospheric environment parameters and the electrical effect characteristics of energized service, the core corrosion influencing factors screened out include temperature, current value, chloride ion deposition rate, relative humidity, and electric field strength; the noise interference item removed is the sulfation rate.

9. The method according to claim 8, characterized in that, The nonlinear dose-response function is as follows: In the formula, Y represents the corrosion weight loss, in g / m²; S d RH is the chloride ion deposition rate, in mg / (100cm²·d); t is the relative humidity, in %; T is the corrosion time, in days; T is the ambient temperature, in °C; I is the current density, in A / mm². 2 E represents the electric field strength, in kV / m; a, b, c, d, e, f, g, h, and k are all parameters to be fitted.

10. The method according to claim 9, characterized in that, Fitting and optimizing the nonlinear dose-response function, including: To minimize the relative prediction error, a weighted nonlinear least squares method is used to construct the objective function, and an L2 regularization term is introduced into the objective function. The optimal solution for the parameters to be fitted is found by a two-stage global optimization strategy that combines coarse search and fine search.

Images