Method for constructing and labeling visual data of organic coating corrosion based on electrochemical restraint

Abstract

This invention belongs to the field of intelligent detection and data engineering technology for material corrosion, and relates to a method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints. This method constructs an experimental system for the corrosion process of organic coatings, simultaneously collects image data and electrochemical impedance spectroscopy data of the coating surface, uses electrochemical parameters as objective calibration criteria, automatically grades and annotates the image data, and performs consistency verification by combining the evolution law of visual features, thereby forming a highly reliable corrosion visual dataset. Furthermore, through data screening and quality assessment mechanisms, a standardized dataset for training intelligent recognition models is constructed. This method can solve the problems of strong subjectivity and poor consistency in traditional manual annotation, providing a reliable data foundation for intelligent recognition of organic coating corrosion.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and data engineering technology for material corrosion, and specifically to a method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints. Background Technology

[0002] In recent years, with the rapid development of artificial intelligence technology, deep learning-based intelligent recognition methods have shown great potential in the field of organic coating corrosion status assessment. These methods train deep neural networks to automatically identify and determine the corrosion level from coating surface images, potentially enabling automated and intelligent field corrosion detection. However, the performance of deep learning models is highly dependent on the quantity and quality of training data, especially the accuracy and consistency of data labels.

[0003] Currently, the annotation of visual datasets used to train corrosion recognition models relies primarily on manual visual interpretation. This traditional data annotation method has several inherent drawbacks. First, manual annotation is highly subjective. Different annotators, and even the same annotator at different times, may have different judgments on the level of the same corrosion image, leading to inconsistent labeling standards and high data noise. Second, manual annotation lacks a unified physical benchmark. The classification of corrosion levels is often based solely on the visual impression of surface morphology, such as discoloration area and blister density. However, these visual appearances do not have a strict linear correspondence with the actual, internal corrosion state of the coating. Especially in the early stages of corrosion, the medium may have already penetrated into the coating, and the coating resistance may have decreased significantly, but the surface morphology changes may not be obvious. At this time, annotation based solely on visual images is prone to omissions or mislabeling, misclassifying a "corroded" state as "intact." This annotation bias transmits incorrect information to the model, severely limiting the model's recognition accuracy and generalization ability.

[0004] Furthermore, existing data construction processes often lack effective data screening and quality assessment mechanisms. Due to fluctuations in experimental conditions and inconsistent image acquisition quality, abnormal samples are inevitable in the raw data. If these low-quality or mislabeled samples are used directly for model training without screening, the overall quality of the dataset will be further reduced, affecting the final performance of the model. Therefore, there is an urgent need for a visual data construction and automatic annotation method that can break away from excessive reliance on human experience and use objective physical parameters as the benchmark for the true state, thereby fundamentally improving the credibility and consistency of erosion visual datasets and laying a solid data foundation for the application of intelligent recognition models. Summary of the Invention

[0005] By constructing an experimental system for the corrosion process of organic coatings, image data and electrochemical impedance spectroscopy data of the coating surface are collected simultaneously. Electrochemical parameters are used as objective calibration criteria to automatically classify and label the image data, and consistency verification is performed based on the evolution of visual features, thus forming a highly reliable corrosion visual dataset. Furthermore, a standardized dataset for training intelligent recognition models is constructed through data screening and quality assessment mechanisms. This method can solve the problems of strong subjectivity and poor consistency in traditional manual annotation, providing a reliable data foundation for intelligent recognition of organic coating corrosion.

[0006] The purpose of this invention is to provide a method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints. It aims to use electrochemical parameters as an objective and quantitative "real label source" to achieve automatic and accurate annotation of corrosion image data, and to construct a high-quality standardized dataset through consistency verification and data filtering mechanisms.

[0007] To achieve the above objectives, the core of the technical solution provided by this invention lies in using electrochemical parameters that accurately reflect the internal corrosion state of the coating as the "gold standard" for calibrating visual data. By establishing a quantitative mapping relationship between electrochemical parameters and corrosion levels, automated and standardized annotation of massive image data is achieved. Furthermore, the annotation results are verified and cleaned based on the constraints of electrochemical evolution laws. The data construction system upon which this invention relies includes: an image acquisition device, an electrochemical testing device, a data synchronization unit, a feature extraction module, an automatic annotation module, a consistency verification module, and a dataset construction module. Specifically, this invention is achieved through the following technical solution: The first step is to build a multimodal data synchronous acquisition system (technical means). Input: Untreated coating sample, preset corrosion test conditions (such as temperature, humidity, salt spray cycling, etc.).

[0008] Methodology: A corrosion test covering the entire lifecycle of the coating was designed and conducted under controlled environmental conditions to ensure data repeatability and regularity. Standardized image acquisition equipment (including a ring standard light source, high-resolution industrial camera, and fixed shooting position) and an electrochemical testing system (three-electrode system, electrochemical workstation, and synchronous controller) were employed. Under a strictly unified time series, high-resolution surface images and corresponding electrochemical impedance spectroscopy (EIS) data were simultaneously acquired for the same coating measurement points. A data synchronization unit was responsible for triggering the acquisition, ensuring time alignment between the images and EIS data.

[0009] Output: Synchronously aligned image-electrochemical impedance spectroscopy data pairs.

[0010] Technical benefits: It ensures the temporal and spatial consistency of the data, providing a reliable physical basis for subsequent automatic labeling based on electrochemical parameters, and avoiding labeling errors caused by time misalignment.

[0011] The second step is to extract electrochemical characteristic parameters as a labeling benchmark (technical means). Input: The raw electrochemical impedance spectroscopy data collected.

[0012] Processing method: In-depth analysis of the synchronously acquired electrochemical impedance spectroscopy data. By establishing a reasonable equivalent circuit model (such as Randle's circuit model), automatic fitting is performed using the nonlinear least squares method to extract electrochemical parameters that can quantitatively and sensitively characterize the protective performance and corrosion degree of the coating. These parameters include at least the low-frequency impedance modulus (specifically, the impedance modulus at 0.01 Hz |Z|0.01 Hz), the coating resistance Rc, and the charge transfer resistance Rct when the substrate is corroded.

[0013] Output: A set of electrochemical parameters that quantitatively characterize the corrosion state of the coating (including |Z| 0.01 Hz, Rc, Rct, etc.).

[0014] Technical effect: Objective and sensitive physical quantities of corrosion state were obtained. These parameters are directly related to the barrier performance of the coating and the corrosion activity of the substrate. As the "gold standard" for subsequent automatic labeling, the physical meaning and accuracy of the label are guaranteed from the source.

[0015] The third step involves setting grading thresholds based on electrochemical parameters and implementing automatic labeling (technical means). Input: Electrochemical parameter set and image data at corresponding time points.

[0016] Processing method: Statistical analysis of a large amount of experimental data, combined with an understanding of the physical process of coating failure, was conducted to establish a corrosion state grading standard based on the order-of-magnitude changes in electrochemical parameters, especially low-frequency impedance modulus. For example, based on the change in impedance value from high to low, the coating state can be divided into multiple levels such as "intact," "initial penetration," "intermediate failure," and "severe corrosion." Then, an automated script (such as a Python script) was written to read the electrochemical parameter values ​​of each sample and apply them according to a preset threshold rule (e.g., |Z| 0.01Hz > 10). 9 Ω·cm 2 For good condition, 10 7 < |Z|0.01Hz ≤ 10 9 For initial infiltration, 10 5 < |Z|0.01Hz ≤ 10 7 For intermediate-term failure, |Z| 0.01Hz ≤ 10 5For severely corroded images, the system automatically adds corrosion level labels to the corresponding image files and saves key electrochemical parameter values ​​as supplementary labels, forming labeled data containing rich information.

[0017] Output: An image dataset with corrosion level labels and electrochemical metadata.

[0018] Technical results: The annotation process has been fully automated, eliminating the subjective differences of manual annotation, ensuring the consistency and repeatability of labels, and significantly improving annotation efficiency, laying the foundation for building large-scale datasets.

[0019] The fourth step is to perform consistency verification based on the evolutionary patterns of visual features (technical means). Input: A labeled image dataset and its corresponding electrochemical parameters.

[0020] Processing method: Quantitative visual features of all images are extracted, such as image entropy, texture contrast, and defect area ratio. According to the principles of corrosion kinetics, these visual features should exhibit a monotonic evolution trend with increasing corrosion intensity (i.e., decreasing electrochemical parameters). Utilizing this principle, all automatically labeled samples are scanned: Correlation analysis: Calculate the Spearman rank correlation coefficient between visual features and electrochemical parameters (e.g., |Z| 0.01 Hz) to assess whether the correlation direction is as expected (e.g., the image entropy should increase when the impedance decreases, showing a negative correlation).

[0021] Anomaly detection: Statistical methods, such as the 3σ criterion, are used to calculate the mean μ and standard deviation σ of the fitting residuals between the visual features and electrochemical parameters of all samples. Samples with an absolute residual value greater than 3σ, as well as samples whose visual feature change trends deviate significantly from the electrochemical parameter change trends (e.g., electrochemical parameters show that the coating has failed severely, but the image features still show that the surface is smooth), are identified as abnormal data and removed.

[0022] Output: A high-confidence image dataset that has passed consistency checks.

[0023] Technical results: It effectively identified and eliminated random errors (such as inaccurate focusing and dirt interference) and electrochemical test errors in the data acquisition process, further improving the purity and reliability of the dataset and ensuring the quality of training data.

[0024] Step 5: Construct a standardized visual dataset (technical means) Input: Valid image data that has passed the consistency check.

[0025] Processing method: For qualified samples that pass the consistency check, further data screening and organization are carried out. Low-quality images with blurry images, uneven lighting, or interference from non-corrosive stains are filtered out. The remaining high-quality, highly consistent data are standardized and organized, stored according to a unified naming rule (e.g., "Sample ID Corrosion Level"), file format (e.g., JPEG / PNG), and directory structure, and an annotation file (e.g., CSV or JSON format) containing all key information such as image path, corrosion level, electrochemical parameters, and visual features is generated.

[0026] Output: A standardized, structured visual dataset of organic coating corrosion.

[0027] Technical results: A high-quality dataset that can be directly used for training deep learning models has been formed. The data has clear physical labels and good distribution coverage, providing solid data support for the development of intelligent recognition models and helping to improve the recognition accuracy and generalization ability of the models.

[0028] Compared with the prior art, the advantages of this invention are: First, using electrochemical parameters as an objective physical quantity as the labeling benchmark fundamentally eliminates the subjectivity of manual labeling, ensuring the authenticity and consistency of data labels and providing a high-quality "standard answer" for model training. Second, by introducing a consistency verification mechanism based on the evolution of visual features, abnormal samples during data collection and experimentation are effectively identified and eliminated, further improving the purity and reliability of the dataset. Third, this method automates data labeling, significantly reducing the cost and time of manual labeling, making it possible to construct large-scale, high-quality corrosion datasets. Finally, the dataset constructed by this method, with its labels possessing clear electrochemical and physical meanings, enables intelligent recognition models trained on it to not only identify surface morphology but also understand the underlying corrosion nature, providing an indispensable and solid data foundation for the subsequent development of a high-precision and high-reliability intelligent corrosion assessment system. Attached Figure Description

[0029] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the overall method of the present invention, which shows the complete technical route from experimental design, data acquisition, feature extraction, model construction to final calibration output; Figure 2 This diagram illustrates the automatic labeling mechanism based on electrochemical parameter thresholds in this invention, demonstrating how low-frequency impedance modulus values ​​are mapped to different corrosion level labels. (The threshold division in the diagram is based on the order-of-magnitude variation of electrochemical parameters, such as 10...) 9 10 7 105 Ω·cm 2 ); Figure 3 This is a schematic diagram of data screening and quality control in this invention, showing the process of identifying and removing abnormal samples through correlation analysis of visual features and electrochemical parameters (abnormality identification is performed by Spearman correlation coefficient calculation and 3σ criterion). Detailed Implementation

[0030] The present invention will be further explained below with reference to specific implementation schemes, but this explanation does not limit the invention. The structures, proportions, sizes, etc., shown in the accompanying drawings are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportional relationships, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention. At the same time, terms such as "upper," "lower," "front," "rear," and "middle" used in this specification are only for clarity of description and are not intended to limit the scope of the present invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the present invention.

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to two specific embodiments. Embodiment 1 will elaborate on how to construct a high-quality visual dataset of organic coating corrosion from scratch in a controlled laboratory environment. Embodiment 2 will demonstrate how this method, in a real engineering context, constructs, annotates, and cleans data collected on-site for a coastal steel structure tower, to verify its engineering applicability and robustness.

[0032] Example 1: Construction of a visual dataset of organic coating corrosion in a laboratory environment Experimental conditions and material preparation.

[0033] This embodiment aims to establish a standardized dataset covering the entire lifecycle, from intact to complete failure. 2A12 aluminum alloy, widely used in the aerospace industry, was selected as the substrate and processed into 60 flat samples measuring 150mm × 75mm × 3mm. After standard pretreatment, all samples were uniformly coated with an epoxy polyamide primer and an aliphatic polyurethane topcoat using automated spraying technology, controlling the total dry film thickness to be 100 ± 10 micrometers. The samples were then cured under standard conditions for 10 days before use.

[0034] To simulate the marine atmospheric environment, accelerated corrosion tests were conducted in a cyclic corrosion test chamber. The test spectrum was set as follows: 4 hours of salt spray (5 wt% NaCl, 35℃), 4 hours of drying (60℃, RH<30%), 4 hours of damp heat (50℃, RH>95%), and 4 hours of ultraviolet irradiation (60℃, 0.8W / m²). 2 (@340nm), forming a 16-hour cycle, with a total test time of 720 hours, equivalent to 45 complete cycles.

[0035] Multimodal data synchronous acquisition.

[0036] Before the start of the experiment (0 hours), and at 24, 72, 168, 360, 528, and 720 hours into the experiment, the experiment was paused, and five samples were removed for data acquisition. Each sample was handled according to the following procedure: First, the sample was placed on a customized standard image acquisition station equipped with an adjustable LED ring standard light source (color temperature 5500K, color rendering index Ra=95). A 24-megapixel industrial camera was used to capture a 100mm × 100mm macroscopic image of the central area of ​​the sample at a fixed height and angle. Three images were taken for each sample, and the clearest image was saved in a lossless format. Immediately after image acquisition, electrochemical impedance spectroscopy (EIS) measurements were performed on the same area of ​​the sample surface using an absorbable polytetrafluoroethylene (PTFE) electrolytic cell. A 3.5wt% NaCl solution was injected into the electrolytic cell, a three-electrode system was used, and a 10mV sinusoidal excitation was applied at an open-circuit potential. The measurement frequency range was 100kHz to 0.01Hz. Ultimately, at 7 time points, a total of 35 sets of valid data (5 blocks × 7 times) were obtained. Each set of data included one or more standard images and a corresponding electrochemical impedance spectroscopy (EIS) curve. Considering the sample size, this experiment prepared multiple batches of samples and conducted 4 rounds of repeated experiments, ultimately reaching a total sample size of approximately 2000 sets to ensure the statistical significance and coverage of the data. The data acquisition system used in this embodiment included: a standard light source dark box (custom model), an industrial camera (model MV-CA060-10GC), and an electrochemical workstation (model PARSTAT 4000A). Acquisition was triggered by a synchronous controller to ensure time alignment between the images and EIS data.

[0037] Rules for electrochemical parameter analysis and automatic labeling were established.

[0038] All collected electrochemical impedance spectroscopy (EIS) data were fitted using ZView software with a unified equivalent circuit. For samples exhibiting a single capacitive arc in their impedance spectra, Randle's simplified equivalent circuit R(Q[R]) was used for fitting to extract the coating resistance Rc and coating capacitance Qc. For samples with two time constants or diffusion tails, a more complex equivalent circuit containing Warburg impedance was used for fitting to extract the charge transfer resistance Rct, and the impedance modulus |Z|0.01Hz at 0.01Hz was recorded. Through statistical analysis of all fitted data and combined with macroscopic phenomena observed during the experiment, an automatic labeling rule centered on |Z|0.01Hz was established. This rule divides the corrosion state into four levels: Level 0 (intact) corresponds to |Z|0.01Hz > 10. 9 Ω•cm 2 Level 1 (Initial Infiltration) corresponds to 10 7 < |Z|0.01Hz ≤ 10 9 Ω•cm 2 Level 2 (Mid-term Failure) corresponds to 10 5 < |Z|0.01Hz ≤ 10 7 Ω•cm 2 Level 3 (Severe Corrosion) corresponds to |Z| 0.01Hz ≤ 10 5 Ω•cm 2 .

[0039] Automatic annotation and consistency verification.

[0040] Based on the above rules, a Python script was written to automate the processing of all 2000 sets of data. The script first reads the EIS fitting result for each sample, and automatically adds corrosion level labels (0, 1, 2, 3) to the corresponding image files based on their |Z|0.01Hz values. It also stores the precise value of |Z|0.01Hz, coating resistance Rc, and other metadata in a structured JSON file. After automatic labeling, a consistency check based on visual features is performed. The script further extracts two key visual features from all images: image entropy and defect area ratio. According to electrochemical principles, as |Z|0.01Hz decreases, both image entropy and defect area ratio should show an increasing trend. To quantify this relationship, the script calculates the Spearman rank correlation coefficient ρ between the image entropy X and the logarithm of |Z|0.01Hz Y for all samples. The calculation steps are: sorting X and Y to obtain ranks, and then calculating the Pearson correlation coefficient between the ranks, resulting in ρ ≈ -0.85, indicating a significant negative correlation, consistent with the expected trend. Simultaneously, the 3σ criterion is used for anomaly detection: the residual between the image entropy of each sample and the fitted value based on |Z|0.01Hz is calculated, yielding the mean μ and standard deviation σ of the residual. Samples with an absolute residual value greater than 3σ are marked as anomalies. Samples deviating from the overall trend by more than 3 standard deviations, and those with a low |Z|0.01Hz (e.g., less than 10), are also considered anomalies. 5 Samples with low image entropy and defect area ratios (indicating potential surface contamination or image acquisition errors) were marked as "awaiting verification" by the script. These marked samples were then manually verified by two experienced researchers. The verification results showed that approximately 80 sets (about 4%) out of 2000 data sets were identified as abnormal. The main reasons for these anomalies included: blurring due to focusing issues during image acquisition (approximately 30 sets); interference from non-corrosive contaminants due to inadequate cleaning of the sample surface before photography (approximately 20 sets); and data anomalies caused by leakage in the electrolytic cell during electrochemical testing (approximately 30 sets). These abnormal samples were removed from the dataset.

[0041] Standardized dataset construction.

[0042] After automatic annotation and consistency verification, a total of 1920 sets of high-quality, highly consistent, and valid data were obtained. The script then processed the final data, renaming all images to the format "Sample ID Corrosion Level" and generating a CSV-formatted annotation file containing all key information such as sample ID, image path, corrosion level, |Z| 0.01Hz, Rc, image entropy, and defect area ratio. Finally, all data and annotation files were archived according to a predetermined directory structure, forming a standard organic coating corrosion visual dataset that can be directly used for deep learning model training. The entire data processing pipeline is based on a Python-developed data processing pipeline, including an EIS fitting submodule (based on the ZView engine), an image feature extraction submodule (based on OpenCV), an automatic annotation submodule, and a consistency verification submodule (including Spearman correlation calculation and 3σ anomaly detection).

[0043] The above results are basically consistent with the actual corrosion situation. A certain error range is allowed, but the overall trend of change remains consistent.

[0044] Example 2: Data Construction for Coastal Steel Structure Tower Project Application targets and data sources.

[0045] This embodiment aims to verify the applicability of the method of the present invention in a real and complex engineering environment. A broadcast television signal transmission tower in a coastal area of ​​southern my country, which has been in service for 12 years, was selected as the research object. The tower is 120 meters high, with a main structure of Q345B steel and a surface coating system consisting of epoxy zinc-rich primer (60 microns), epoxy micaceous iron oxide intermediate paint (80 microns), and acrylic polyurethane topcoat (60 microns). The tower is constantly exposed to sea winds, salt spray, heavy rain, and strong sunlight, resulting in significant coating aging issues. To construct a dedicated dataset for this type of structure, the project team selected 20 typical measuring points at different heights and orientations (windward, leeward, and sunny sides) of the tower and conducted tracking data collection for 6 months (summer to winter).

[0046] On-site data acquisition plan.

[0047] On-site data acquisition faces numerous challenges, such as working at heights, varying lighting conditions, and the inability to use liquid electrolytic cells. To address these challenges, a specialized on-site acquisition solution was designed. Image acquisition utilizes a drone equipped with a high-resolution camera and a ring light, hovering approximately 0.5 meters from the measurement point. Ground operators remotely control the drone to capture images, ensuring consistent shooting distance, angle, and lighting conditions for each image. Five images are captured at each measurement point, with the clearest image selected later. Electrochemical data acquisition employs a novel portable solid gel electrolyte probe. This probe is connected to a miniaturized electrochemical workstation. Testing personnel travel to the measurement point using an aerial work platform, press the probe firmly against the same location captured by the drone, and after one minute of stillness, perform single-frequency (0.1Hz) or multi-frequency impedance measurements to quickly obtain the equivalent low-frequency impedance modulus at that point. Data is collected monthly for six consecutive months. Meanwhile, standard plates of the same material and coating as the tower were placed at the bottom of the tower as natural exposure control samples. A batch was collected monthly for more comprehensive electrochemical testing and image analysis to calibrate the rapid on-site test data. In the end, a total of 120 sets of on-site "image-impedance" paired data were collected from 20 measurement points × 6 times, plus the data from the 3 plates collected each month for 6 months, for a total of approximately 1200 sets of samples.

[0048] Automatic data labeling and threshold adjustment based on on-site features.

[0049] First, based on the experience summarized in Example 1 and accelerated laboratory test data, the data of the coated films collected on-site were analyzed in depth. The analysis revealed that due to differences in the on-site environment (such as temperature and humidity fluctuations) and the coating system (thickness, formulation), absolute thresholds based solely on laboratory data are not entirely applicable. For example, a threshold of 10... 7 Ω•cm 2 The boundary may be slightly lower in the field when the coating is in good condition. Therefore, this embodiment first fine-tunes the corrosion state grading threshold based on the analysis of the first batch of samples and the initial state measurement points. The adjusted labeling rule is: Grade 0 (good) corresponds to |Z| 0.01Hz > 5×10 8 Ω•cm 2 Level 1 (Initial Infiltration) corresponds to 10 6 < |Z|0.01Hz ≤ 5×10 8 Ω•cm 2 Level 2 (Mid-term Failure) corresponds to 10 4 < |Z|0.01Hz ≤ 10 6 Ω•cm 2 Level 3 (Severe Corrosion) corresponds to |Z| 0.01Hz ≤ 10 4 Ω•cm 2Based on this adjusted rule, images of all 1200 data sets were automatically labeled. The labeling results clearly demonstrate the evolution of corrosion: in the initial state, the impedance of most measuring points was within 10... 8 -10 9 The magnitude was initially set at level 0; one month later, the impedance at some measuring points on the windward side dropped to 10. 7 The order of magnitude was changed, and the label was changed to level 1; three months later, the impedance at some measuring points dropped to 10. 5 The level was measured and labeled as level 2; tiny rust spots began to appear on the image. Six months later, the impedance in some severely affected areas dropped to 10. 3 The level is 3, and obvious rust spots and paint peeling are visible in the image.

[0050] Consistency verification and outlier data removal.

[0051] Consistency checks were performed on the automatically labeled data based on visual features. Image entropy and texture contrast based on the gray-level co-occurrence matrix were extracted as key visual features for all images. Analysis revealed that for the vast majority of samples, these two features showed a significant negative correlation with the logarithm of |Z| at 0.01 Hz, i.e., a decrease in impedance and an increase in image entropy and texture contrast, consistent with expectations. Specifically, the Spearman correlation coefficient ρ was approximately -0.78, indicating a strong negative correlation. However, the check process also identified approximately 96 sets (about 8%) of anomalous data. Analysis of these anomalous data revealed the following main sources: three measurement points had significant shadows of metal components mixed into the images due to drone shooting angle deviations, severely interfering with texture feature calculations and causing abnormally high image entropy; two measurement points had a thin layer of dry salt deposits on their surface after three months, but this was misjudged as blistering in the images, resulting in abnormal visual features; and a few data points had abnormally low measured impedance values ​​due to poor contact of the electrochemical probe in the field. These outlier samples all met the outlier condition of the 3σ criterion, and were therefore marked and removed from the final dataset.

[0052] Dataset construction and application effect evaluation.

[0053] Following the steps outlined above, a high-quality, highly consistent "Visual Dataset for Organic Coating Corrosion of Coastal Steel Structure Towers" was constructed, comprising approximately 1104 sets of data. This dataset covers data samples from different seasons, orientations, and corrosion stages, with labels possessing clear electrochemical and physical meanings. To verify its effectiveness, the project team trained a lightweight convolutional neural network classification model using this dataset and validated it during subsequent tower maintenance and inspection. Results show that the model trained on this dataset achieved an accuracy rate exceeding 90% in identifying corrosion levels in new on-site images, and its sensitivity to early-stage corrosion was significantly higher than that of manual visual inspection. More importantly, due to the removal of numerous outlier samples from the dataset, the model's robustness was significantly enhanced, exhibiting good tolerance to interference from changes in lighting and minor surface contaminants. This engineering application example fully demonstrates the feasibility and superiority of the method described in this invention in constructing high-quality corrosion visual datasets in real-world, complex environments, paving the way for intelligent corrosion recognition technology to move from the laboratory to engineering applications.

[0054] The above results are basically consistent with the actual corrosion situation. A certain error range is allowed, but the overall trend of change remains consistent.

[0055] Matters not covered in this invention are common knowledge.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing and annotating visual data on the corrosion of organic coatings based on electrochemical constraints, characterized in that, Includes the following steps: (1) Construct an organic coating corrosion test system, conduct corrosion tests under controlled environmental conditions, and synchronously collect coating surface image data and electrochemical impedance spectroscopy data according to time series; (2) Analyze the electrochemical impedance spectroscopy data to extract electrochemical parameters such as low-frequency impedance modulus, coating resistance and charge transfer resistance; (3) Based on the electrochemical parameters, a corrosion state classification threshold is set, and the image data at the corresponding time points are automatically labeled; the corrosion state classification is based on the order of magnitude change of the low-frequency impedance modulus. (4) Extract the visual feature parameters of the image and perform consistency verification on the annotation results, removing abnormal data that do not conform to the electrochemical evolution law. The consistency assessment is achieved by calculating the Spearman correlation coefficient between visual features and electrochemical parameters. The 3σ method based on standard deviation is used to remove abnormal data that exceed the mean ± 3 times the standard deviation. (5) The labeled image data is filtered and organized to construct a standardized visual dataset of organic coating corrosion.

2. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The automatic labeling includes assigning a corrosion level label and a corresponding electrochemical parameter label to each image.

3. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The consistency verification is based on the correlation between visual features and electrochemical parameters; the consistency assessment is achieved by calculating the Spearman correlation coefficient between visual features and electrochemical parameters.

4. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, When the trend of visual feature changes is inconsistent with the trend of electrochemical parameter changes, the corresponding data will be identified as abnormal samples and removed.

5. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The visual data mentioned above includes macroscopic images and local area images.

6. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The dataset contains data samples under different corrosion stages and various environmental conditions.

7. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The dataset constructed was used to train an intelligent identification model for organic coating corrosion.

8. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The specific implementation method of automatic labeling in step (3) is as follows: based on the preset electrochemical parameter threshold range, an automatic script is used to generate a corresponding corrosion level label for each image, and key electrochemical parameters such as low-frequency impedance modulus and coating resistance are stored as metadata in association with the image to form a labeling result containing multi-dimensional information.

9. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The consistency verification in step (4) includes a data screening calculation method, specifically: calculating the Spearman rank correlation coefficient between visual features and electrochemical parameters to determine whether the correlation direction conforms to the corrosion evolution law; simultaneously, using the 3σ criterion, samples with a fitting residual between visual features and electrochemical parameters exceeding three times the standard deviation are identified as abnormal data and removed. The consistency assessment is achieved by calculating the Spearman correlation coefficient between visual features and electrochemical parameters.

10. The method for constructing and annotating visual data of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The anomaly determination further includes: identifying samples whose electrochemical parameters decrease but whose visual characteristics do not deteriorate synchronously based on the evolution trend of visual features, marking them as potential anomalies, and finally confirming whether to remove them by combining manual review.

Images