Multi-modal intelligent evaluation method for corrosion sensitivity of metal material in deep sea environment

Abstract

The invention provides a multi-modal intelligent evaluation method for corrosion sensitivity of a metal material in a deep sea environment, which comprehensively considers various factors of material corrosion in the deep sea environment and solves the problem of multi-dimensional characterization of corrosion damage of the metal material in a deep sea multi-physics coupling environment. The limitation of a traditional method in cross-modal correlation modeling and few-sample learning is broken through. The method is suitable for corrosion risk dynamic evaluation and life prediction of the high-strength steel material under the coupling action of high pressure, low temperature, low oxygen and microorganisms. The evaluation accuracy is improved, and the data labeling cost is reduced; deep sea adaptability is enhanced, and the actual sea prediction error is smaller than 10% through pressure-sulfide coupling modeling. And data comprehensive analysis and analysis: carrying out coupling analysis on various data, comprehensively considering material failure factors, and carrying out comprehensive assessment on material failure risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of material corrosion monitoring and artificial intelligence, and particularly relates to a deep-sea environment metal material corrosion sensitivity intelligent evaluation method based on multi-modal data fusion, which is suitable for dynamic evaluation and life prediction of corrosion risk of high-strength steel materials under the coupling action of high pressure, low temperature, low oxygen and microorganisms. BACKGROUND

[0002] Due to its important strategic position and rich mineral resources, the deep sea contains rich oil, natural gas, manganese nodules and phosphorite minerals. A large amount of research and development investment is needed for the installation, manufacturing and operation of marine engineering equipment. The seawater has strong corrosiveness, and the corrosion problem is a common problem that must be solved for the application of deep-sea devices and equipment. The high corrosiveness of seawater poses a great risk to equipment underwater, especially in the deep sea where there is great pressure and it is difficult for personnel to reach. Once corrosion failure occurs, it often causes catastrophic accidents and huge losses.

[0003] Therefore, it is of great significance to carry out deep-sea environment test technology and corrosion research, to explore the environmental adaptability behavior of equipment in the deep sea, and to establish deep-sea corrosion protection technology, which is the premise and foundation for the development and utilization of the deep sea, and is of great significance to ensure the safe and reliable operation of deep-sea equipment.

[0004] Deep-sea extreme environment, including high pressure, low temperature, low oxygen, sulfide and microbial activity, has multi-factor strong coupling characteristics on the corrosion mechanism of metal materials, and the corrosion type and corrosion mechanism are quite different from shallow sea environment. Deep-sea extreme environment material corrosion is mainly divided into chemical corrosion and electrochemical corrosion according to the corrosion principle, and can be divided into uniform corrosion and local corrosion according to the corrosion form. Among them, the chemical corrosion includes hydrogen corrosion, in the deep-sea environment, hydrogen reacts with metal materials to cause material performance degradation and structure damage. Hydrogen atoms penetrate into the metal interior, combine with metal atoms to form new phases or compounds, causing material embrittlement and cracking. In the deep-sea environment containing hydrogen sulfide, hydrogen will react with iron atoms in steel to form iron hydride, reducing the toughness and strength of steel. Electrochemical corrosion includes galvanic corrosion, macroscopic inhomogeneity caused by external factors, such as contact of two different metals or alloy phases, electrolyte solution composition or concentration inconsistency on the same metal surface, and different parts of the same metal under different forced current, which will cause electrochemical corrosion, and the influencing factors are numerous. At present, there are many defects in the research methods of material corrosion in deep-sea extreme environment, such as 1) single mode evaluation limitation: the traditional method relies on single electrochemical or mechanical data, which cannot fully characterize the chemical-mechanical-hydrogen embrittlement synergy of corrosion damage (such as CN118976543 A only uses polarization curve analysis); 2) lack of environment coupling modeling: the existing model does not embed the pressure, sulfide concentration environmental parameters in a differentiable form into the algorithm, resulting in insufficient deep-sea adaptability (such as CN 119129361 A does not consider the pressure gradient effect); 3) high dependence on data labeling: corrosion evaluation based on supervised learning requires a large amount of labeled data, which is difficult to meet the scenario of scarce deep-sea test data (such as CN 120154321 A relies on a fully labeled data set).

[0005] The application adopts a deep-sea environment metal material corrosion sensitivity intelligent evaluation method based on multi-modal data fusion, which is suitable for dynamic evaluation and life prediction of corrosion risk of high-quality steel materials under the coupling action of high pressure, low temperature, low oxygen and microorganisms. SUMMARY

[0006] The application aims to provide a deep-sea environment metal material corrosion sensitivity intelligent evaluation method based on multi-modal data fusion, which comprehensively considers various factors of material corrosion in deep-sea environment, solves the multi-dimensional characterization problem of metal material corrosion damage in deep-sea multi-physical field coupling environment, and breaks through the limitations of traditional methods in cross-modal correlation modeling and few-shot learning. It is suitable for dynamic evaluation and life prediction of corrosion risk of high-strength steel materials under the coupling action of high pressure, low temperature, low oxygen and microorganisms.

[0007] The attention fusion model is a model method for effectively integrating features from different information sources (such as text, image, speech different modalities, or different levels and different positions in the same modality) by using an attention mechanism. The core of the attention fusion model is the attention mechanism. The attention mechanism can be understood as a strategy for weighting different parts of the input data, similar to the way humans focus their attention on key parts when processing information. In the model, the importance of different positions of the input features or different modalities of the features is determined by calculating the attention weights of the features. These weights are determined by a learnable function. In the present application, for multi-modal data (environmental parameters representing different corrosion factors and corrosion data), the model learns to evaluate the importance of various environmental parameters on the corrosion of the material, and then assigns corresponding weights according to the importance. In the data fusion-analysis-computation process, more important features can play a greater role.

[0008] To achieve the above purpose, the present application provides the following technical solutions:

[0009] First step, multi-modal data synchronous acquisition

[0010] Synchronously acquire deep-sea environmental parameters, including pressure P, temperature T, dissolved oxygen DO and sulfide concentration [S 2 ], and material response data, including electrochemical impedance spectrum Z EIS , acoustic emission energy E AE and hydrogen permeation flux J H .

[0011] Second step, physical driving feature extraction

[0012] According to the collected multi-modal environmental parameters, various corrosion behavior parameters are calculated to extract physical driving features. The main corrosion products include: passivation film stability index (PSSI), stress corrosion cracking risk coefficient (SCCR), hydrogen induced cracking risk index (HCR) and local pitting corrosion occurrence coefficient (PTOR).

[0013] a. Passivation film stability index (PSSI):

[0014]

[0015] where R p is the polarization resistance, E f is the elastic modulus of the passivation film, h is the film thickness, and D H is the hydrogen diffusion coefficient.

[0016] b. Stress corrosion cracking risk coefficient (SCCR):

[0017]

[0018] a is a material constant, f c is the acoustic emission characteristic frequency, K I is the crack tip stress intensity factor.

[0019] c. Quantification of hydrogen induced cracking (HIC) susceptibility of materials in sulfide containing environments

[0020] Crack morphology statistical parameters:

[0021] CLR (Crack Length Ratio): The ratio of the total length of cracks in the specimen cross section to the total length of the evaluation line segment CLR = (∑ crack length specimen total length / specimen total length) x 100%

[0022] CSR (Crack Sensitivity Ratio): The ratio of the total area of cracks to the total area of the specimen

[0023] CSR = (∑ crack area specimen cross-sectional area / specimen cross-sectional area) x 100%

[0024] CTR (Crack Thickness Ratio): The ratio of the maximum crack thickness to the specimen thickness

[0025] HCR comprehensive formula:

[0026] HCR = a CLR + b CSR + g CTR

[0027] Where the weight coefficients (a, b, g) need to be adjusted according to the material type and environmental severity (such as NACE TM0284 standard recommends a = 0.4, b = 0.3, g = 0.3)

[0028] d. Pitting Occurrence Ratio (PTOR) model

[0029] 1) Pitting initiation randomness (Poisson process)

[0030] Pitting initiation follows a non-homogeneous Poisson process, and its occurrence probability is related to time and environmental parameters:

[0031]

[0032] λ(t): Intensity function (positively correlated with Cl- concentration, temperature)

[0033] k: The number of pits observed in t time.

[0034] 2) Intensity function model:

[0035]

[0036] Where A is a material constant, E_a is the activation energy, and n is the Cl- sensitivity index (304 stainless steel n ≈ 0.6-1.2).

[0037] 3) Pitting growth extreme value statistics (Gumbel distribution)

[0038] The deepest pitting pit depth d_max obeys the Gumbel first extreme value distribution:

[0039]

[0040] Position parameter μ: related to the corrosion resistance of the material (such as 316L stainless steel μ> 304L)

[0041] Scale parameter σ: characterizes the dispersion of pitting depth (σ≈0.2-0.5mm).

[0042] 4) PTOR comprehensive coefficient

[0043] Combining the initiation and growth stages, define PTOR as:

[0044]

[0045] e. Occurrence of corrosion site picture collection and fracture defect photo image data

[0046] The characteristic photos of the corrosion site collected in the laboratory and on site include macroscopic data of the fracture, scanning electron microscope data, and visual data collection of the fracture site.

[0047] Step 3, mixed attention fusion model

[0048] The attention fusion model is a model method that uses attention mechanisms to effectively integrate features extracted from different information sources or different levels and different positions in the same modality. Attention mechanism can be understood as a strategy of weighting different parts of input data, similar to the way humans focus on key parts when processing information. In the model, the importance of different positions or different modalities of input features is determined by calculating the attention weights of the features, and then the corresponding weights are assigned according to the importance, so that more important features can play a greater role in the fusion process.

[0049] The weights are determined by a learnable attention module. For multiple different levels of data, the model learns the importance of various feature data parameters to the current task, and then assigns corresponding weights according to the importance, so that more important features can play a greater role in the fusion process.

[0050] Physical attention module: generate prior weights based on Faraday's law and Fick's diffusion law, strengthen the electrochemical-hydrogen-localized corrosion diffusion coupling relationship;

[0051] Data attention module: use multi-head self-attention mechanism to mine cross-modal temporal correlation;

[0052] The specific method is as follows:

[0053] The goal of data processing is to prepare structured and processable input for the hybrid attention model. The process is as follows:

[0054] I. Modality-specific feature extraction:

[0055] Corrosion data data parameters: feature extraction is performed on various current and potential data parameters, and a feature parameter vector database is established for corrosion feature parameters represented by different parameters, and feature extraction methods in the corresponding field are used.

[0056] Picture data:

[0057] Preprocessing: resize, normalize.

[0058] Feature extraction: use pre-training. Can extract global features and establish a feature vector database.

[0059] Local / space features: reshape the feature map into style features.

[0060] II. Feature representation alignment and standardization:

[0061] Dimension alignment: ensure that the feature sequences extracted by different modalities are in the same dimension. Sometimes padding, truncation or linear projection is needed.

[0062] Standardization: normalize or standardize (subtract mean and divide variance) the features of each modality, so that the features of different modalities are in a similar numerical range, accelerating training and improving stability.

[0063] III. Construction of model input:

[0064] Combine the processed features of each modality into a structured input.

[0065] IV. Core of hybrid attention fusion:

[0066] After the data processing prepares the features, the hybrid attention mechanism inside the model starts to work:

[0067] Intra-modal attention: first perform self-attention within each modality to capture the intra-modal dependencies.

[0068] Input: single-modality feature sequence.

[0069] Output: feature sequence weighted by intra-modal attention, which can better focus on important elements within the modality.

[0070] Inter-modal attention / cross-modal attention: this is the key to hybridization. Take one modality as Query and the other as Key and Value to calculate attention weights.

[0071] Feature fusion: fuse the features processed by the attention mechanism. It includes splicing, element-by-element operation, gating mechanism, and hierarchical fusion method.

[0072] Task-specific output layer: input the fused features into the task-related layer to generate the final output.

[0073] The beneficial effects of the present application are:

[0074] Evaluation accuracy is improved: compared with single-mode method, the corrosion stage recognition accuracy is improved by 35%. Data labeling cost is reduced: self-supervised learning reduces the demand for manual labeling by 80%. Deep sea adaptability is enhanced: pressure-sulfide coupling modeling makes the real sea prediction error less than 10%. Data comprehensive analysis and interpretation: multiple data coupling analysis, comprehensive consideration of material failure factors, and comprehensive evaluation of material failure risk. BRIEF DESCRIPTION OF DRAWINGS

[0075] Figure 1 The data processing flowchart is shown in

[0076] Figure 2 The high-strength steel field hydrogen embrittlement fracture and cracking crack diagram is shown in

[0077] Figure 3 The specific data processing flowchart is shown in DETAILED DESCRIPTION

[0078] The present application will be further described below in conjunction with examples, but is not limited thereto.

[0079] Example 1

[0080] This embodiment provides a deep sea environment metal material corrosion sensitivity multi-modal intelligent evaluation method, taking high-strength steel, corrosion test and data evaluation analysis in simulated deep sea environment as an example. The data processing flowchart is shown in Figure 1 .

[0081] Test conditions: test pressure: 30 MPa, test temperature: 3℃, seawater oxygen content: 0.3ppm, sulfide concentration: 10ppm.

[0082] Data acquisition: ZEIS data (0.1-10e5 Hz) EAE (threshold 50dB) and JH (accuracy 0.1nA / cm2) are collected, and the corrosion picture collection result is shown in Figure 2 . It is a corrosion fracture photo and a metallographic photo.

[0083] Feature calculation: PSSI=0.72 (medium stable), SCCR=2.5 (high risk), HERC=0.18 (critical hydrogen embrittlement)

[0084] Model output: Life prediction 1.5 years, real sea coupon validation error 8.3%.

[0085] Data processing flow:

[0086] As shown in the figure, the specific data processing flowchart is shown. Figure 3

[0087] Embodiment 2

[0088] The embodiment provides a multi-modal intelligent evaluation method for corrosion sensitivity of a metal material in a deep-sea environment:

[0089] First step, multi-modal data synchronous acquisition

[0090] Synchronously acquire deep-sea environmental parameters, including pressure P, temperature T, dissolved oxygen DO and sulfide concentration [S 2 -] and material response data, including electrochemical impedance spectrum Z EIS , acoustic emission energy E AE and hydrogen permeation flux J H .

[0091] Second step, physical driving feature extraction

[0092] According to the acquired multi-modal environmental parameters, various corrosion behavior parameters are calculated, and physical driving features are extracted. Main corrosion products include: passivation film stability index (PSSI), stress corrosion cracking risk coefficient (SCCR), hydrogen induced cracking risk index (HCR) and local pitting corrosion occurrence coefficient (PTOR).

[0093] a. Passivation film stability index (PSSI):

[0094]

[0095] Where R p is the polarization resistance, E f is the elastic modulus of the passivation film, h is the film thickness, and D H is the hydrogen diffusion coefficient.

[0096] b. Stress corrosion cracking risk coefficient (SCCR):

[0097]

[0098] α is a material constant, f c is the acoustic emission characteristic frequency, and K I is the crack tip stress intensity factor.

[0099] c. For quantifying the hydrogen induced cracking (HIC) sensitivity of the material in the sulfide-containing environment

[0100] Crack morphology statistical parameters:​

[0101] CLR (Crack Length Ratio): The ratio of the total length of cracks in the specimen cross section to the total length of the evaluation line segment, CLR = (∑ crack length specimen total length / specimen total length) x 100%

[0102] CSR (Crack Sensitivity Ratio): The ratio of the total area of cracks to the total area of the specimen

[0103] CSR = (∑ crack area specimen cross-sectional area / specimen cross-sectional area) x 100%

[0104] CTR (Crack Thickness Ratio): The ratio of the maximum crack thickness to the specimen thickness

[0105] HCR Comprehensive Formula:

[0106] HCR = a CLR + b CSR + g CRT

[0107] Where the weight coefficients (a, b, g) need to be adjusted according to the material type and environmental severity (such as NACE TM0284 standard suggests a = 0.4, b = 0.3, g = 0.3)

[0108] d. Point corrosion occurrence coefficient (PTOR) model

[0109] 1) Randomness of point corrosion initiation (Poisson process)

[0110] Point corrosion initiation follows a non-homogeneous Poisson process, and its occurrence probability is related to time and environmental parameters:

[0111]

[0112] λ(t): Intensity function (positively correlated with Cl- concentration and temperature)

[0113] k: The number of point corrosion observed in t time.

[0114] 2) Intensity function model:

[0115]

[0116] Where A is a material constant, E_a is the activation energy, and n is the Cl- sensitivity index (304 stainless steel n ≈ 0.6-1.2).

[0117] 3) Extreme value statistics of point corrosion growth (Gumbel distribution)

[0118] The maximum depth of the deepest point corrosion pit d_max follows the first type of Gumbel extreme value distribution:

[0119]

[0120] Position parameter μ: related to the corrosion resistance of the material (e.g., μ > 304L for 316L stainless steel).

[0121] Scale parameter σ: characterizes the dispersion of pitting depth (σ≈0.2~0.5mm).

[0122] 4) PTOR comprehensive coefficient

[0123] Combining the germination and growth stages, PTOR is defined as:

[0124]

[0125] e. Image acquisition of corroded areas and photographic data of fracture defects

[0126] Photographs of corrosion features obtained from laboratory simulations and on-site collection, including macroscopic data of the fracture surface and scanning electron microscopy data, were used to collect visual data of the fractured areas.

[0127] Step 3: Hybrid Attention Fusion Model

[0128] Attention fusion models are a method that utilizes attention mechanisms to effectively integrate features extracted from different information sources or from different levels and locations within the same modality. The attention mechanism can be understood as a strategy of weighting different parts of the input data, similar to how humans focus their attention on key parts when processing information. In the model, the importance of features in the fusion process is determined by calculating the attention weights of features at different locations or from different modalities.

[0129] The weights are determined through a learnable attention module. For data at different levels, the model learns the importance of various feature data parameters to the current task (such as corrosion rate and failure probability), and then assigns corresponding weights based on this importance, so that more important features can play a greater role in the fusion process.

[0130] Physical Attention Module: Generates prior weights based on Faraday's law and Fick's diffusion law to strengthen the electrochemical-hydrogen-localized corrosion diffusion coupling relationship;

[0131] Data Attention Module: Employs a multi-head self-attention mechanism to mine cross-modal temporal correlations;

[0132] The specific method is as follows: Core data processing flow:

[0133] The goal of data processing is to prepare structured, processable inputs for hybrid attention models. The process is as follows:

[0134] I. Modality-Specific Feature Extraction:

[0135] Corrosion data data parameters: feature extraction on various current, potential data parameters, establish feature parameter vector database for different parameter representation of corrosion feature parameters, and use corresponding field feature extraction method.

[0136] Picture data:

[0137] Preprocessing: resize, normalization.

[0138] Feature extraction: use pre-training. Can extract global features, establish feature vector database.

[0139] Local / space features: according to feature map remodeling for style features.

[0140] II. Feature representation alignment and standardization:

[0141] Dimension alignment: ensure that the feature sequences extracted by different modalities are in the same dimension. Sometimes padding, truncation or linear projection is needed.

[0142] Standardization: normalize or standardize (subtract mean and divide variance) the features of each modality, so that the features of different modalities are in similar numerical range, accelerate training and improve stability.

[0143] III. Model input construction:

[0144] Combine the processed features (tensors) of each modality into a structured input.

[0145] IV. The core of mixed attention fusion:

[0146] After the data processing is ready for features, the mixed attention mechanism inside the model starts to work:

[0147] Intra-modal attention: first do self-attention within each modality, capture the intra-modal dependencies (such as long-distance dependencies in text, relationships between different regions in image).

[0148] Input: single-modal feature sequence.

[0149] Output: feature sequence weighted by intra-modal attention, which can focus on important elements within the modality.

[0150] Inter-modal attention / cross-modal attention: this is the key to mixing. Take one modality as Query, another modality as Key and Value, calculate attention weight.

[0151] Feature fusion: fuse the features processed by attention mechanism. Including concatenation, element-wise operation, gating mechanism, hierarchical fusion method.

[0152] Task-specific output layer: the fused features are input into the task-related layer (such as classifier, regressor, decoder) to generate the final output.

[0153] The core advantage of the hybrid attention fusion model lies in its ability to dynamically and contextually integrate heterogeneous information. The data processing link provides a high-quality, structured input feature base for this powerful fusion capability. Understanding and optimizing the data processing flow and the design of the attention fusion mechanism are key to building high-performance multi-modal models.

[0154] The details of the present application are known.

[0155] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A multimodal intelligent evaluation method for the corrosion sensitivity of metallic materials in deep-sea environments, characterized in that: Considering various factors of material corrosion in deep-sea environment, solving the multi-dimensional characterization problem of metal material corrosion damage in deep-sea multi-physical field coupling environment, and breaking through the limitations of traditional methods in cross-modal correlation modeling and few-shot learning; Collecting environmental parameters and other corrosion monitoring and detection data, including electrochemical signals, image data; For the collected electrochemical signals, the corrosion data evaluation method is used to establish the corresponding evaluation equation to calculate the corresponding passivation film stability index (PSSI), stress corrosion cracking risk coefficient (SCCR), hydrogen induced cracking (HIC) sensitivity, and pitting corrosion occurrence coefficient (PTOR); The collected data is imported into the mixed attention model to evaluate the corrosion rate and predict the corrosion life.

2. The deep-sea environment metal material corrosion sensitivity multi-modal intelligent evaluation method according to claim 1, characterized in that: Step 1, multi-modal data synchronous acquisition synchronously acquiring deep-sea environmental parameters, including pressure P, temperature T, dissolved oxygen DO, sulfide concentration [S 2 ]; and material response data, including electrochemical impedance spectrum Z EIS , acoustic emission energy E AE , hydrogen permeation flux J H ; Step 2, physical driving feature extraction According to the collected multi-modal environmental parameters, various corrosion behavior parameters are calculated, and physical driving features are extracted; Corrosion products include: passivation film stability index (PSSI), stress corrosion cracking risk coefficient (SCCR), hydrogen induced cracking risk index (HCR), and local pitting corrosion occurrence coefficient (PTOR); a. Passivation film stability index (PSSI): where R p is the polarization resistance, E f is the elastic modulus of the passivation film, h is the film thickness, D H is the hydrogen diffusion coefficient; b. Stress corrosion cracking risk coefficient (SCCR): a is a material constant, f c is the acoustic emission characteristic frequency, K I is the crack tip stress intensity factor; c. Quantify the hydrogen induced cracking (HIC) sensitivity of the material in the sulfide-containing environment Crack morphology statistical parameters: CLR (crack length percentage): the ratio of the total length of the crack on the sample cross section to the total length of the evaluation line segment CLR = (∑ crack length sample total length / sample total length) × 100% CSR (crack sensitivity percentage): the ratio of the total area of the crack to the total area of the sample CSR = (∑ crack area sample cross-sectional area / sample cross-sectional area) × 100% CTR (crack thickness percentage): the ratio of the maximum crack thickness to the sample thickness HCR comprehensive formula: HCR = α CLR + β CSR + γ CTR Where the weight coefficients (α, β, γ) need to be adjusted according to the material type and environmental severity (such as NACE TM0284 standard recommends α = 0.4, β = 0.3, γ = 0.3) d. Pitting corrosion occurrence coefficient (PTOR) model 1) Pitting initiation randomness (Poisson process) Pitting initiation follows a non-homogeneous Poisson process, and its occurrence probability is related to time and environmental parameters: λ(t): intensity function (positively correlated with Cl- concentration, temperature) k: the number of observed pits in t time; 2) Intensity function model: Where A is a material constant, E_a is the activation energy, and n is the Cl- sensitivity index (304 stainless steel n ≈ 0.6-1.2); 3) Pitting growth extreme value statistics (Gumbel distribution) The maximum pitting pit depth d_max follows the Gumbel first-type extreme value distribution: Position parameter μ: related to material corrosion resistance (such as 316L stainless steel μ > 304L) Scale parameter σ: representing the dispersion of pitting depth (σ ≈ 0.2-0.5 mm); 4) PTOR comprehensive coefficient PTOR is defined as: e、Collecting pictures of corrosion sites and fracture defect image data Collecting visual data of the fracture site from the laboratory simulation and field collected corrosion site characteristic photos, including fracture macroscopic data, scanning electron microscope data, and visual data collection of the fracture site; Third step, mixed attention fusion model Attention fusion model is a model method that uses attention mechanism to effectively integrate features extracted from different information sources or different levels and different positions in the same modality; Attention mechanism can be understood as a strategy of weighting different parts of input data, similar to the way humans focus on key parts when processing information; In the model, the importance of different positions or different modalities of input features is determined by calculating the attention weight of the features.

3. The method according to claim 2, wherein the method is characterized by: The weight is determined by a learnable attention module. For multiple different levels of data, the model learns the importance of various feature data parameters to the current task, and then assigns corresponding weights according to the importance, so that more important features can play a greater role in the fusion process. Physical attention module: generate prior weights based on Faraday's law and Fick's diffusion law to strengthen the electrochemical-hydrogen-localized corrosion diffusion coupling relationship; Data attention module: use multi-head self-attention mechanism to mine cross-modal temporal correlation; The specific method is as follows, the core data processing flow: The goal of data processing is to prepare structured and processable input for the mixed attention model, and the process is as follows: I. Modality-specific feature extraction: Corrosion data parameters: feature extraction of various current and potential data parameters, establishment of feature parameter vector database for corrosion feature parameters represented by different parameters, and use of feature extraction methods in the corresponding field; Picture data: Preprocessing: resize, normalize; Feature extraction: use pre-training; can extract global features, establish feature vector database; Local / space features: reshape the feature map into style features; II. Feature representation alignment and standardization: Dimension alignment: ensure that the feature sequences extracted from different modalities are in the same dimension; sometimes padding, truncation or linear projection is needed; Standardization: normalize or standardize the features of each modality to make the features of different modalities in similar numerical ranges, speed up training and improve stability; III. Model input construction: Combine the processed features of each modality into a structured input; IV. Core of mixed attention fusion: After the features are prepared by data processing, the mixed attention mechanism in the model starts to work: Intra-modal attention: first perform self-attention within each modality to capture the intra-modal dependency; Input: single modality feature sequence; Output: feature sequence weighted by intra-modal attention, which can better focus on important elements within the modality; Inter-modal attention / cross-modal attention: this is the key to mixing; take one modality as Query and the other modality as Key and Value to calculate the attention weight; Feature fusion: fuse the features processed by the attention mechanism; Including concatenation, element-wise operation, gating mechanism, hierarchical fusion method; Task-specific output layer: input the fused features into the task-related layer to generate the final output.

Images