An artificial intelligence-based hot galvanized steel pipe corrosion performance prediction system

Abstract

The application relates to the technical field of artificial intelligence, and discloses a hot galvanized steel pipe corrosion performance prediction system based on artificial intelligence. The system collects environment, electrochemistry, thickness and surface image data and the like through a multi-source sensor, carries out time sequence alignment, denoising and completion through an edge computing unit, and obtains high-quality monitoring data. A cloud end adopts a causal consistency constraint time sequence convolution network model (TCN), combines a self-attention mechanism and a causal coordinate mapping, realizes joint modeling and physical causal alignment of multi-modal features, and outputs an annual corrosion rate, a residual zinc layer thickness, a red rust initial appearance time and a life interval. The system realizes intelligent identification and prediction of a corrosion state, improves the precision, stability and interpretability of a model, and provides efficient and reliable technical support for corrosion operation and maintenance and life management of the hot galvanized steel pipe.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based system for predicting the corrosion performance of hot-dip galvanized steel pipes. Background Technology

[0002] Hot-dip galvanized steel pipes are widely used in outdoor structures in industries such as power, communications, transportation, municipal engineering, and petrochemicals due to their excellent corrosion resistance, high mechanical strength, and cost-effectiveness. Their protective mechanism primarily relies on the shielding effect of the zinc layer and the sacrificial anode protection effect. However, under long-term atmospheric exposure, the steel pipe surface is affected by a combination of factors, including humidity, temperature, salt spray, pollutants (SO2, NO2), and ultraviolet radiation. This leads to the gradual consumption of the zinc layer, resulting in deterioration phenomena such as red rust spots, coating damage, and thinning, ultimately affecting structural safety and service life. While existing intelligent corrosion monitoring and prediction technologies have incorporated sensor networks and machine learning methods, they are mostly limited to shallow feature modeling or single-modal data analysis. They lack the ability to jointly model multi-source features such as environmental, electrochemical, and geometric changes, resulting in insufficient interpretability and cross-scenario stability of the models, making it difficult to accurately reflect the true corrosion patterns and lifespan changes of hot-dip galvanized steel pipes under complex atmospheric environments. Summary of the Invention

[0003] This invention proposes an artificial intelligence-based corrosion performance prediction system for hot-dip galvanized steel pipes. Addressing the shortcomings of existing intelligent corrosion prediction technologies, such as insufficient multimodal fusion and poor model generalization and interpretability, the system constructs a complete framework integrating multi-source sensing monitoring, edge computing, and causal-constrained deep learning inference. The system deploys multi-source sensors in the steel pipe and its surrounding environment to acquire real-time multi-dimensional information such as temperature and humidity, contaminant concentration, surface wetting time, electrochemical parameters, and visual images. In the edge computing unit, time synchronization, anomaly detection, extended Kalman filtering for denoising, and data completion are performed to ensure the integrity and reliability of the input data. In the cloud, a temporal convolutional network (TCN) model with causal consistency constraints is used. This model embeds a self-attention mechanism and introduces causal coordinate systems and subspace consistency constraints in temporal modeling, enabling the model to capture multimodal feature coupling relationships while maintaining physical causal consistency. This allows for accurate prediction of corrosion rate, zinc layer thickness decay, and the initial appearance time of red rust. The innovation of this system lies in combining data-driven learning with corrosion mechanism modeling through the "causal consistency constraint + multi-source intelligent fusion" mechanism, which significantly improves the stability, interpretability and cross-environment adaptability of the model, and provides a new technical approach for life prediction and protection strategy optimization of hot-dip galvanized steel pipes.

[0004] This invention proposes an artificial intelligence-based corrosion performance prediction system for hot-dip galvanized steel pipes, applicable to online monitoring and lifespan assessment of hot-dip galvanized steel pipes exposed to outdoor atmospheric conditions. The system includes: The sensing and acquisition layer is used to collect inspection data of hot-dip galvanized steel pipes; The edge computing and gateway unit adopts an integrated industrial-grade embedded computer with a built-in GNSS timing module to unify the timestamps of the hot-dip galvanized steel pipe inspection data. It also performs outlier detection and extended Kalman filtering for noise reduction, packet loss completion and data integrity verification, and locally encrypts and caches the breakpoint data. Finally, it obtains high-quality inspection data that is time-aligned, noise-reduced and smoothed, and complete and continuous. The cloud service platform includes feature engineering and AI inference services. The AI ​​inference service has a built-in causal consistency constraint TCN model. Feature engineering processes and extracts features from high-quality detection data to construct multi-dimensional feature data. The multi-dimensional feature data is input into the causal consistency constraint TCN model, which outputs the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining life range, and triggers alarms and maintenance suggestions according to thresholds. The causal consistency constraint TCN model is constructed as follows: based on the TCN model, a self-attention layer is embedded in the main structure of the TCN model to realize global interaction modeling between multimodal input features. At the output of the TCN model, a causal coordinate system and a subspace selection matrix are introduced to construct the CL subspace loss function of causal consistency constraint, which is used to optimize the consistency between the latent space representation and the physical causal space of the TCN model, ensuring the prediction stability and interpretability of the model under different working conditions, thus constructing the causal consistency constraint TCN model. Communication and security module: used for data interaction between edge computing and gateway units and cloud service platforms; Power supply and installation components: Provide independent power support for the sensing and acquisition layer, edge computing and gateway unit, and communication and security module. It adopts a combination of solar panels and lithium iron phosphate batteries for power supply, and works with power management and low temperature heating control modules to ensure long-term stable operation of the system without external power supply.

[0005] Furthermore, the process of processing and extracting features from high-quality detection data through feature engineering to construct multidimensional feature data specifically includes the following steps: Step B1: Based on the temperature and humidity parameters in the high-quality detection data, calculate the wetting time (TOW) and combine it with the time distribution characteristics to form wetting exposure characteristic data, which is used to characterize the time period characteristics of the steel pipe surface in an electrochemically active state. Step B2: Based on the atmospheric SO2 and NO2 concentrations and wind speed parameters in the high-quality detection data, calculate the pollutant deposition rate and total flux to obtain pollution load characteristics, which are used to reflect the external driving force of environmental pollution on the corrosion process. Step B3: Based on the LPR, ER and ultrasonic thickness measurement parameters in the high-quality test data, calculate the polarization resistance, instantaneous corrosion rate and thickness loss rate to form corrosion behavior and geometric degradation characteristics, which are used to describe the material's electrochemical reaction and morphological degradation process. Step B4: Based on the steel pipe surface image in the high-quality inspection data, a lightweight convolutional neural network is used to extract texture, color and morphological features to generate visual representation features for identifying red rust distribution, coating damage and localized corrosion areas; the visual representation features include red rust area ratio visual features, coating damage ratio visual features and localized corrosion area visual features. Step B5: Integrate wet exposure characteristic data, pollution load characteristics, corrosion behavior and geometric degradation characteristics, and visual representation characteristics to construct multidimensional feature data, providing a unified input for the causal consistency constraint TCN model.

[0006] Furthermore, the multidimensional feature data is input into the causal consistency constraint TCN model, which outputs the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range. This process includes the following steps: Step S1: Normalize and time-series aligned multidimensional feature data, construct a multidimensional causal feature matrix according to the category dimensions of physical environment features, corrosion electrochemical features and visual image features, and define causal variable labels in the multidimensional causal feature matrix to form a set of causal variables; Step S2: Input the multidimensional causal feature matrix into the TCN model to extract local time-dependent features; capture the global interaction relationship between cross-dimensional features through the self-attention layer to obtain intermediate hidden representations; and then generate preliminary prediction results through output layer mapping. Step S3: Establish a causal coordinate system based on the multidimensional causal feature matrix and the causal variable set; define an alignment function to map the intermediate hidden representations to the causal coordinate system and obtain the alignment result, which includes the aligned intermediate representations and causal variables; define the corresponding selection matrix for each causal variable in the causal variable set and construct the subspace selection matrix. Step S4: During multiple batches of training, continuously record the alignment results and construct a natural representation library to store the natural distribution samples of the causal consistency constraint TCN model under different working conditions; the causal consistency constraint TCN model generates new intervention intermediate representation data during training, and retrieves a set of samples that match the causal variables corresponding to the intervention intermediate representation data from the natural representation library to obtain the hidden representation of the samples; and performs a weighted average on the hidden representation of the samples to generate a counterfactual latent template. Step S5: Based on the subspace selection matrix, perform subspace projection on the intervention intermediate representation data and the representation data of the counterfactual potential template. All projection results together constitute a causal subspace set. In each causal subspace of the causal subspace set, define the CL subspace loss function of causal consistency constraints. According to the CL subspace loss function of causal consistency constraints, sum the losses of all causal subspaces in the causal subspace set to form the overall CL loss function. Introduce the task loss function and combine the overall CL loss function and the task loss function according to the weight coefficients to construct the total loss function. Step S6: Train the causal consistency constraint TCN model using the total loss function to obtain the trained causal consistency constraint TCN model. The trained causal consistency constraint TCN model outputs the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range.

[0007] By adopting the above solution, the beneficial effects achieved by the present invention are as follows: This invention constructs a multi-source sensing monitoring system to achieve full-element data acquisition and time-series synchronization of hot-dip galvanized steel pipes under real atmospheric exposure conditions. It can simultaneously acquire environmental parameters, electrochemical characteristics, geometric thickness, and surface visual information, forming a structured, multimodal corrosion monitoring data foundation. By performing extended Kalman filtering for noise reduction and time alignment through edge computing units, the integrity and timeliness of the data are improved. This solves problems such as data loss, time-series asynchrony, and insufficient representativeness of local sampling in traditional corrosion monitoring, providing high-quality input support for subsequent intelligent prediction and enhancing the system's stable operation under long-term outdoor conditions.

[0008] This invention introduces a temporal convolutional network model with causal consistency constraints, combining self-attention mechanisms with causal coordinate mapping to achieve global interaction between multimodal features and joint modeling under physical causal constraints. This enhances the model's ability to identify complex nonlinear relationships between environmental drivers and corrosion responses. This mechanism effectively addresses the problem of existing intelligent prediction models relying solely on data fitting and failing to adequately characterize physical laws, resulting in predictions that are not only more accurate but also interpretable and stable across different scenarios. Through causal subspace selection and alignment constraints, the model can maintain prediction consistency under various climatic and pollution conditions, achieving high-confidence inference of corrosion rate, remaining zinc layer thickness, and time to first red rust (TTRR).

[0009] This invention enhances the real-time performance and adaptability of the system through an integrated architecture of "edge computing—cloud inference—intelligent assessment," enabling online assessment of corrosion status and dynamic lifespan prediction. The system can automatically trigger maintenance or replacement recommendations based on the risk level output by the model, forming a closed-loop management mechanism that significantly improves the operational efficiency and safety assurance level of hot-dip galvanized steel pipes in complex environments. This technology achieves end-to-end innovation from multi-source sensing to intelligent decision-making, driving the transformation of metal protective materials from static detection to intelligent prediction and proactive protection, and providing high reliability and engineering application value for long-term corrosion management of infrastructure. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall structure of an artificial intelligence-based corrosion performance prediction system for hot-dip galvanized steel pipes proposed in this invention. Figure 2 This is a comparison of the training convergence curves of the TCN model with causal consistency constraints proposed in Example 3 and the TCN model.

[0011] Figure 2 In the diagram, the horizontal axis (X) represents the number of training epochs, and the vertical axis (Y) represents the model training loss. The blue dashed line represents the TCN model, which has a slower curve descent and significant fluctuations in the mid-to-late stages, resulting in less smooth convergence. The red solid line represents the TCN model with causal consistency constraints, which has a significantly faster curve descent and significantly smaller oscillations than the TCN model. Detailed Implementation

[0012] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0013] Example 1, according to Figure 1 This invention proposes an artificial intelligence-based corrosion performance prediction system for hot-dip galvanized steel pipes, applicable to online monitoring and life assessment of hot-dip galvanized steel pipes exposed to outdoor atmospheric conditions. The system includes: The sensing and acquisition layer is used to collect test data of hot-dip galvanized steel pipes. This layer is fixedly installed on the outer surface of the steel pipe under test and its near-field environment, and is equipped with a multi-source sensor assembly, including: an air temperature and humidity sensor (operating range of −30℃ to +70℃, humidity range of 0% to 100%RH), and atmospheric sulfur dioxide (SO2) and nitrogen dioxide (NO2) electrochemical gas sensors (range of 0 to 20). The system includes: ppm), chloride ion deposition and salt spray deposition acquisition units (using a conductivity deposition meter for detection), surface wetting and condensation sensors (for calculating wetting time to death (TOW)), open circuit potential sensors and linear polarization resistance (LPR) corrosion probes, resistive corrosion (ER) probes, coating thickness eddy current probes, and ultrasonic wall thickness probes; an optional industrial camera with IP67 protection is used to acquire visualized images of red rust spots, coating damage, and localized corrosion areas on the steel pipe surface, thereby enabling automatic identification and quantitative assessment of red rust area ratio and surface degradation status; through the collaborative work of the above multi-source sensors, complete hot-dip galvanized steel pipe inspection data covering environmental parameters, electrochemical characteristics, geometric thickness, and appearance status is formed, providing a high-precision input basis for subsequent corrosion performance prediction and life assessment; The edge computing and gateway unit adopts an integrated industrial-grade embedded computer with an IP65 protective housing. It operates in a temperature range of −20℃ to +55℃ and is equipped with 4–20 mA and 0–10 V analog input interfaces, RS-485 / Modbus communication interfaces, Ethernet interfaces, and digital input / output interfaces. The built-in GNSS timing module unifies the timestamps of hot-dip galvanized steel pipe inspection data, performs outlier detection and extended Kalman filtering for noise reduction, packet loss completion, and data integrity verification, and locally encrypts and caches breakpoint data. Finally, it obtains high-quality inspection data that is time-aligned, noise-reduced, smoothed, complete, and continuous. The cloud service platform includes feature engineering and AI inference services. The AI ​​inference service has a built-in causal consistency constraint TCN model. Feature engineering processes and extracts features from high-quality detection data to construct multi-dimensional feature data. The multi-dimensional feature data is input into the causal consistency constraint TCN model, which outputs the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining life range, and triggers alarms and maintenance suggestions according to thresholds. The causal consistency constraint TCN model is constructed as follows: based on the TCN model, a self-attention layer is embedded in the main structure of the TCN model to realize global interaction modeling between multimodal input features. At the output of the TCN model, a causal coordinate system and a subspace selection matrix are introduced to construct the CL subspace loss function of causal consistency constraint, which is used to optimize the consistency between the latent space representation and the physical causal space of the TCN model, ensuring the prediction stability and interpretability of the model under different working conditions, thus constructing the causal consistency constraint TCN model. Communication and security module: Used for data interaction between edge computing and gateway units and cloud service platforms, supporting LoRaWAN, NB-IoT and 5G communication methods, and transmitting information with the cloud service platform via HTTPS protocol; adopts TLS1.3 encryption mechanism and device certificate authentication to ensure transmission security and identity trust; optional explosion-proof housing (Ex d IIC T4) to meet explosion-proof requirements; Power supply and installation components: Provide independent power support for the sensing and acquisition layer, edge computing and gateway unit, and communication and security module. It adopts a combination of solar panels and lithium iron phosphate batteries for power supply, and works with power management and low temperature heating control modules to ensure long-term stable operation of the system without external power supply.

[0014] Example 2, based on Example 1, describes the process of processing and extracting features from high-quality detection data through feature engineering to construct multi-dimensional feature data. The specific steps include: Step B1: Based on the temperature and humidity parameters in the high-quality detection data, calculate the wetting time (TOW) and combine it with the time distribution characteristics to form wetting exposure characteristic data, which is used to characterize the time period characteristics of the steel pipe surface in an electrochemically active state. Step B2: Based on the atmospheric SO2 and NO2 concentrations and wind speed parameters in the high-quality detection data, calculate the pollutant deposition rate and total flux to obtain pollution load characteristics, which are used to reflect the external driving force of environmental pollution on the corrosion process. Step B3: Based on the LPR, ER and ultrasonic thickness measurement parameters in the high-quality test data, calculate the polarization resistance, instantaneous corrosion rate and thickness loss rate to form corrosion behavior and geometric degradation characteristics, which are used to describe the material's electrochemical reaction and morphological degradation process. Step B4: Based on the steel pipe surface image in the high-quality inspection data, a lightweight convolutional neural network is used to extract texture, color and morphological features to generate visual representation features for identifying red rust distribution, coating damage and localized corrosion areas; the visual representation features include red rust area ratio visual features, coating damage ratio visual features and localized corrosion area visual features. Step B5: Integrate wet exposure characteristic data, pollution load characteristics, corrosion behavior and geometric degradation characteristics, and visual representation characteristics to construct multidimensional feature data, providing a unified input for the causal consistency constraint TCN model.

[0015] Example 3, according to Figure 2 This embodiment is based on Embodiment 2. In this invention, multidimensional feature data is input into the causal consistency constraint TCN model, which outputs the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range. Specifically, the following steps are included: Step S1: Normalize and time-series aligned multidimensional feature data. Construct a multidimensional causal feature matrix according to the categories of physical environment features, corrosion electrochemical features, and visual image features. Define six categories of causal variable labels in the multidimensional causal feature matrix, including wetting time (TOW), pollution load, polarization resistance (Rp), thickness loss rate, red rust area ratio, and coating damage ratio, to form a set of causal variables. Step S2: Input the multidimensional causal feature matrix into the TCN model to extract local time-dependent features; capture the global interaction relationship between cross-dimensional features through the self-attention layer to obtain the intermediate hidden representation; then generate the preliminary prediction result through the output layer mapping; the intermediate representation comprehensively carries the multi-source feature information of physical environment, corrosion dynamics and visual degradation state at different times; Step S3: Establish a causal coordinate system based on the multidimensional causal feature matrix and the set of causal variables to describe the structural relationships between six causal dimensions: wetting time (TOW), pollution load, polarization resistance (Rp), thickness loss rate, red rust area ratio, and coating damage ratio; define an alignment function to map intermediate hidden representations to the causal coordinate system to obtain the alignment result, which includes the aligned intermediate representations and causal variables; achieve structural alignment between the model latent space and the physical causal space; define a corresponding selection matrix for each causal variable in the set of causal variables, and construct a subspace selection matrix to extract the subspace coordinates corresponding to each causal dimension from the alignment result; Step S4: During multiple training sessions, continuously record the alignment results and construct a natural representation library to store the natural distribution samples of the causal consistency constraint TCN model under different operating conditions. The causal consistency constraint TCN model generates new intervention intermediate representation data during training. Based on the causal variables corresponding to the intervention intermediate representation data, retrieve a matching sample set from the natural representation library to obtain the hidden representation of the samples. Then, perform a weighted average on the hidden representations to generate a counterfactual latent template. By comparing the current alignment results with the counterfactual latent template, determine the distribution differences of the model representation in the causal dimension, providing a reference for subsequent subspace constraints. Step S5: Based on the subspace selection matrix, the intervention intermediate representation data and the representation data of the counterfactual potential template are subspaced and projected to the low-dimensional subspaces corresponding to each causal variable. Each subspace selection matrix corresponds to a physical causal variable. After projection, the low-dimensional representation under that variable is obtained. All projection results together constitute a causal subspace set. In each causal subspace of the causal subspace set, a causal consistency constraint CL subspace loss function is defined to minimize the Euclidean distance (L2 term) and direction difference (cosine term) between the intervention representation and the counterfactual template in the same causal dimension, so as to maintain the consistency of amplitude and direction, thereby achieving the alignment and distribution convergence of the representation in the causal space. According to the causal consistency constraint CL subspace loss function, the losses of all causal subspaces in the causal subspace set are summed to form the overall CL loss function. The task loss function is introduced, and the overall CL loss function and the task loss function are combined according to the weight coefficients to construct the total loss function. Define the CL subspace loss function for causal consistency constraints: , , ; in, The loss function of the CL subspace representing the causal consistency constraint. This represents the set of causal dimensions, which is the local summation index within a single subspace; The intermediate representation of the intervention is in the first Representation (projected coordinates) in a causal subspace. Represents the subspace selection matrix. This represents the intermediate characteristics of the intervention. Indicates the alignment function; The representation data in the counterfactual potential template is in the first... Representation (projected coordinates) in a causal subspace; This represents the representational data in the counterfactual potential template; This represents the Euclidean distance term (L2 term), used to constrain the magnitude consistency of the two representations; The cosine term represents the similarity of the angle between the directions of two vectors. Overall CL loss function: ; in, This represents a global summary index across subspaces. Represents the overall CL loss function; ; in, Represents the total loss function. This represents the task loss function. This represents a hyperparameter used to adjust the weight of causal constraints in the total loss, achieving a stable anti-divergence balance. Step S6: Perform joint optimization training on the causal consistency constraint TCN model using the total loss function to obtain the trained causal consistency constraint TCN model. Output the final prediction results through the trained causal consistency constraint TCN model, including annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range.

[0016] In conventional technical fields, multidimensional feature data is input into the TCN model to output annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range. This process specifically includes the following steps: Step R1: Input multidimensional feature data into the TCN model, extract local dependency features in the time series through a one-dimensional dilated convolutional structure, and model the erosion evolution trend using a fixed-depth convolutional layer to output intermediate time series representations. Step R2: Input the intermediate time series representation directly into the fully connected layer to generate preliminary prediction results; Step R3: Using mean squared error (MSE) as the training loss function, optimize the model parameters using gradient descent to obtain the trained TCN model. The trained TCN model outputs the final prediction results, including annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining lifetime range.

[0017] Example 4, this example is based on Example 3. The cloud service platform includes feature engineering and artificial intelligence inference services. The artificial intelligence inference service has a built-in causal consistency constraint TCN model. Feature engineering processes and extracts features from high-quality detection data to construct multi-dimensional feature data. The multi-dimensional feature data is input into the causal consistency constraint TCN model to output the annual corrosion rate, remaining zinc layer thickness, time to first appearance of red rust (TTRR), and remaining life range, and triggers alarms and maintenance suggestions according to thresholds. In this embodiment, the corrosion performance of hot-dip galvanized steel pipes in area A is monitored online and their lifespan is predicted. The installation site is about 500 m from the coast, with an average annual temperature of 16.8 ℃ and an average annual relative humidity of 75%, and is constantly affected by industrial pollution. The inference output of the causal consistency-consistent TCN model is as follows: Annual average corrosion rate: 2.12 µm / a; Remaining zinc layer thickness: It is predicted that approximately 71 µm will remain after 3 years of current service and 52 µm will remain after 6 years. Time to first appearance of red rust (TTRR): estimated at 3.2 years; Remaining lifetime range: 8.5 to 9.8 years when the zinc layer fails completely (thickness less than 15 µm).

[0018] Warnings and maintenance recommendations: Short-term measures: Conduct localized zinc spraying operations within the next 6 months to repair approximately 0.8 m of the red rusted section; Mid-term plan: During the third year's re-inspection, the area and thickness of red rust on the entire pipe will be re-measured; Long-term management: After a cumulative service life of 6 years, the bracket should be completely replaced or re-galvanized.

[0019] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A system for predicting the corrosion performance of hot-dip galvanized steel pipes based on artificial intelligence, characterized in that, The system includes: The sensing and acquisition layer collects inspection data of hot-dip galvanized steel pipes; The edge computing and gateway unit is an integrated industrial-grade embedded computer with a built-in GNSS timing module to unify the timestamps of hot-dip galvanized steel pipe inspection data, and to perform outlier detection, extended Kalman filtering for noise reduction, packet loss completion, and data integrity verification to obtain high-quality inspection data. The cloud service platform includes feature engineering and artificial intelligence inference services. The artificial intelligence inference service has a built-in causal consistency constraint TCN model. Through feature engineering, high-quality detection data is processed and features are extracted to construct multi-dimensional feature data. The multi-dimensional feature data is input into the causal consistency constraint TCN model to output the annual corrosion rate, remaining zinc layer thickness, time of initial appearance of red rust, and remaining life range. The communication and security module is used for data interaction between the edge computing and gateway unit and the cloud service platform; Power supply and installation components provide independent power support for the sensing and acquisition layer, edge computing and gateway units, and communication and security modules; The process of inputting multidimensional feature data into the causal consistency constraint TCN model and outputting the annual corrosion rate, remaining zinc layer thickness, initial appearance time of red rust, and remaining lifetime range includes the following steps: Step S1: Normalize and time-series aligned multidimensional feature data to construct a multidimensional causal feature matrix, and define causal variable labels in the multidimensional causal feature matrix to form a set of causal variables; Step S2: Input the multidimensional causal feature matrix into the TCN model to extract local time-dependent features; capture the global interaction relationship between cross-dimensional features through a self-attention layer to obtain intermediate hidden representations; Step S3: Establish a causal coordinate system based on the multidimensional causal feature matrix and the set of causal variables; define an alignment function to map the intermediate hidden representations to the causal coordinate system and obtain the alignment result; establish a subspace selection matrix based on the set of causal variables; Step S4: Continuously record the alignment results, construct a natural representation library, and store the natural distribution samples of the causal consistency constraint TCN model under different working conditions; the causal consistency constraint TCN model generates intervention intermediate representation data during training, and retrieves the matching sample set from the natural representation library based on the intervention intermediate representation data to obtain the sample hidden representation; and performs a weighted average on the sample hidden representation to generate a counterfactual latent template. Step S5: Based on the subspace selection matrix, perform subspace projection on the intervention intermediate representation data and the representation data in the counterfactual potential template. All projection results together constitute a causal subspace set. In each causal subspace of the causal subspace set, define the CL subspace loss function of causal consistency constraints. According to the CL subspace loss function of causal consistency constraints, sum the losses of all causal subspaces in the causal subspace set to form the overall CL loss function. Introduce the task loss function and combine the overall CL loss function and the task loss function according to the weight coefficients to construct the total loss function. Step S6: Jointly optimize and train the causal consistency constraint TCN model using the total loss function to obtain the trained causal consistency constraint TCN model. Output the annual corrosion rate, remaining zinc layer thickness, initial appearance time of red rust, and remaining lifespan range using the trained causal consistency constraint TCN model. The alignment results include aligned intermediate representations and causal variables.

2. The corrosion performance prediction system for hot-dip galvanized steel pipes based on artificial intelligence according to claim 1, characterized in that: The construction method of the causal consistency constraint TCN model is as follows: based on the TCN model, a self-attention layer is embedded in the main structure of the TCN model to model the global interaction between multimodal input features; at the output of the TCN model, a causal coordinate system and a subspace selection matrix are introduced to construct the CL subspace loss function of causal consistency constraint, optimize the consistency between the latent space representation of the TCN model and the physical causal space, and obtain the causal consistency constraint TCN model.

3. The corrosion performance prediction system for hot-dip galvanized steel pipes based on artificial intelligence according to claim 1, characterized in that: The process of processing and extracting features from high-quality detection data through feature engineering to construct multidimensional feature data includes the following steps: Step B1: Based on the temperature and humidity parameters in the high-quality detection data, calculate the wetting time and obtain the wetting exposure characteristic data; Step B2: Based on the atmospheric SO2 and NO2 concentrations and wind speed parameters in the high-quality detection data, calculate the pollutant deposition rate and total flux to obtain the pollution load characteristics; Step B3: Based on the LPR, ER and ultrasonic thickness measurement parameters in the high-quality detection data, calculate the polarization resistance, corrosion rate and thickness loss rate to form corrosion behavior and geometric degradation characteristics; Step B4: Based on the steel pipe surface image in the high-quality inspection data, use a lightweight convolutional neural network to extract texture, color and morphological features to generate visual representation features; Step B5: Integrate wet exposure characteristic data, pollution load characteristics, corrosion behavior and geometric degradation characteristics, and visual characterization characteristics to construct multidimensional characteristic data.

4. The corrosion performance prediction system for hot-dip galvanized steel pipes based on artificial intelligence according to claim 3, characterized in that: Visual characteristics include the visual characteristics of red rust area ratio, the visual characteristics of coating damage ratio, and the visual characteristics of localized corrosion areas.

Images