MT-PINN-based corrosion monitoring system for earth-covered tanks

Abstract

This invention discloses a corrosion monitoring system for soil-covered storage tanks based on the MT-PINN model, comprising: an information acquisition unit (1) for acquiring service environment information of the environment in which the soil-covered storage tank is located; a corrosion feature preprocessing unit (2) for achieving attention-weighted fusion and spatiotemporal alignment of multi-source heterogeneous service environment information; and an online corrosion rate monitoring unit (3) for introducing corrosion feature information into the MT-PINN model. Furthermore, a physical information model describing the electrochemical corrosion process and corrosion kinetics of the soil-covered storage tank is constructed based on the Butler-Volmer equation; a multi-task physical information neural network is constructed to receive the corrosion feature information and output the corrosion rate monitoring results of the soil-covered storage tank. This system, based on a physical information neural network of corrosion kinetics, achieves dual data and physical drive through the construction of a multi-task branch network architecture and the design of a composite loss function, improving the prediction accuracy and generalization ability of the corrosion monitoring system for soil-covered storage tanks, and possessing good engineering adaptability and application value.

Description

Technical Field

[0001] This invention relates to the field of corrosion monitoring and safety assessment technology for soil-covered storage tanks, and more particularly, to a corrosion monitoring system for soil-covered storage tanks based on MT-PINN.

[0002] Physics-Informed Neural Networks (PINN) is a machine learning framework that uses known physical laws as constraints to train a neural network.

[0003] The combination of a Physical Information Neural Network (PINN), a Multi-Branch Neural Network (MBNN), and a composite loss function is called the MT-PINN model. The MT-PINN model features a multi-task parallel output mechanism. Background Technology

[0004] As a new type of storage equipment, soil-covered storage tanks are widely used in the chemical and energy fields due to their advantages such as saving land resources, improving the environment, and reducing safety risks. See "Discussion on Safety Technology and Application Prospect of Liquefied Hydrocarbon Soil-Covered Storage Tanks," Wang Huiqin, *Petrochemical Safety and Environmental Protection Technology*, Vol. 28, No. 4, 2012, which discloses the differences between soil-covered storage tanks and above-ground spherical tanks, as well as the corrosion prevention and regular safety inspection of soil-covered storage tanks. See "A Brief Discussion on the Current Status and Prospect of Soil-Covered Storage Containers," Du Chenyang, *China Special Equipment Safety*, Vol. 39, No. 2, 2023, which discloses the structure of soil-covered storage containers. See "Corrosion and Protection of Liquefied Hydrocarbon Soil-Covered Storage Tanks," Wan Zhang, *Corrosion and Protection*, Vol. 47, No. 2, February 2026, which discloses the corrosion damage modes of liquefied hydrocarbon soil-covered storage tanks, corrosion prevention measures, and dynamic monitoring of corrosion protection for liquefied hydrocarbon soil-covered storage tanks placed on the soil side. Because these storage tanks are exposed to soil cover for extended periods, they are susceptible to corrosion damage due to the combined effects of multiple factors, including temperature, moisture content, oxygen concentration, ionic composition, and structural stress. This corrosion is particularly pronounced under conditions of coating degradation and fluctuating cathodic protection, leading to localized corrosion and even perforation failure. Due to the enclosed and opaque nature of the soil cover environment, traditional corrosion monitoring methods struggle to provide continuous and accurate corrosion rate assessments, exhibiting issues such as monitoring lag and incomplete information. Existing corrosion prediction methods primarily include empirical model-based methods, purely data-driven machine learning methods, and physical mechanism-based model methods. Empirical models rely on single parameters, making it difficult to reflect the coupled effects of multiple factors; data-driven methods, while possessing strong fitting capabilities, lack physical constraints, resulting in poor prediction stability; and physical models, while having a clear mechanistic basis, face difficulties in parameter acquisition and are unsuitable for complex field environments.

[0005] As a pioneering method that integrates physical priors with deep learning, see "Physics Informed DeepLearning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations", RAISSI M, PERDIKARIS P, Ithaca: arXiv, October 28, 2017. PINN discloses that it can encode the physical laws of the research system into the neural network in the form of loss function constraints. This solves the problems of poor generalization and lack of physical meaning in prediction results in pure data-driven machine learning in small data scenarios, and overcomes the shortcomings of traditional physical models, such as difficulty in obtaining parameters and difficulty in adapting to complex field environments. Therefore, PINN shows broad application prospects in engineering system modeling, prediction of complex physical processes, and other fields.

[0006] Material corrosion is the phenomenon of damage, deterioration, or deterioration of materials and their properties under the chemical, electrochemical, and physical effects of the surrounding environment. Material corrosion is inseparable from the surrounding environment; they constitute a corrosion system. Therefore, preventing material corrosion requires consideration of both material and environmental factors. The methods and means adopted to achieve the purpose of material corrosion protection are called material corrosion protection technologies. All protection technologies can be approached from two aspects: (1) changing the material composition, surface treatment process, and engineering structure design; (2) changing the environment, including media treatment (drying, degassing, desalination, etc.), using corrosion inhibitors, and coatings that can isolate the environment. See Modern Material Corrosion and Protection, Huang Yongchang et al., Shanghai Jiaotong University Press, September 2012, 1st edition, pp. 3, 344.

[0007] To address the technical challenges faced by liquefied hydrocarbon (LNG) covered storage tanks operating in complex soil environments, including diverse corrosion mechanisms, heterogeneous monitoring data from multiple sources, and discrepancies in the compatibility between data and physical models, existing prediction methods have significant shortcomings. Traditional empirical models or purely data-driven models struggle to uniformly characterize the coupling effects of multiple physics fields, often resulting in poor physical interpretability of predictions, and their reliability significantly decreases in scenarios with scarce data or extrapolation of operating conditions. Meanwhile, first-principles-based mechanistic models suffer from high modeling difficulty and computational complexity due to the difficulty in accurately obtaining environmental parameters and complex boundary conditions, hindering their online application. Therefore, there is an urgent need for a corrosion rate prediction method that can integrate multi-source monitoring data and incorporate physical mechanism constraints to improve prediction accuracy and engineering applicability. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention designs a corrosion monitoring system for soil-covered storage tanks based on MT-PINN. This system enhances the physical information neural network based on the physical mechanism of corrosion kinetics and a composite loss function. Through closed-loop feedback of online monitoring and offline operation and maintenance data, it achieves self-optimization in predicting corrosion of soil-covered storage tanks, improving the physical interpretability and generalization ability of the corrosion prediction results. The system includes an information acquisition unit (1), a corrosion feature preprocessing unit (2), and an online corrosion rate monitoring unit based on the MT-PINN model (3). Specifically:

[0009] Information acquisition unit (1): used to acquire service environment information of the environment in which the soil-covered storage tank is located; this unit deploys various types of sensors to collect multi-source heterogeneous environmental data and build a service environment information sequence;

[0010] Corrosion feature preprocessing unit (2): Receives the multi-source heterogeneous service environment information sequence output by the information acquisition unit (1), and uses attention weighted fusion and spatiotemporal alignment filtering method to preprocess the data and extract corrosion-related feature information;

[0011] The corrosion rate online monitoring unit (3) based on the MT-PINN model is used to construct the MT-PINN model; it further includes:

[0012] Feature building module of corrosion kinetic physical mechanism (31): Based on the Butler-Volmer equation, a corrosion kinetic physical information model describing the electrochemical process and metal corrosion prevention behavior of the soil-covered storage tank is constructed;

[0013] Multi-task neural network prediction module (32): Constructs a multi-task physical information neural network and is used to receive the corrosion feature information, iteratively optimizes the MT-PINN model using a composite loss function, and outputs the corrosion rate monitoring results of the soil-covered storage tank.

[0014] Furthermore, training the MT-PINN model includes the following steps:

[0015] S101, the main network processes the training data through two hidden layers and then directly outputs the logarithmic value of the erosion depth. ;

[0016] S102, the first branch network extracts the logarithm of the corrosion current density from the training data. And expressed as a logarithm of corrosion current density Correcting the electrochemical terms in the MT-PINN model;

[0017] S103, the second branch network transforms the training data using the Sigmoid activation function and linearly maps it to the interval. The corrosion kinetic index n is obtained.

[0018] S104, the third branch network extracts the logarithmic value of the environmental coupling factor from the training data. .

[0019] Furthermore, the physical model is embedded through a composite loss function, and two corrosion depth prediction paths are learned simultaneously in the MT-PINN model: one path directly outputs the logarithm of the data-driven maximum corrosion depth prediction value from the main network. And the logarithm of the predicted maximum corrosion depth The value is assigned to the main network; the output of the other route's three branch networks is used to obtain the logarithm of the physics-driven maximum corrosion depth prediction through the physical model. The composite loss function consists of four parts, among which:

[0020] (1) Mean square error between predicted value and actual monitored value The erosion depth used to drive the network to fit the training samples;

[0021] (2) Physical consistency error Used to force the main network output to match the calculation results of the physical equations, so that the network learns the equivalent corrosion current. Time exponent n, environmental coupling factor It conforms to the laws of corrosion kinetics;

[0022] (3) Monotonicity constraint error This is used to ensure that the corrosion depth increases monotonically over time;

[0023] (4) L2 regularization term Used to prevent network overfitting;

[0024] Then we have the total loss function The mathematical expression is .

[0025] Furthermore, the multi-source data fusion edge preprocessing in the corrosion feature preprocessing unit (2) includes the following steps:

[0026] Step 23A: Divide the dataset into three layers: upper, middle, and lower.

[0027] Each layer of the dataset contains the following three types of parameters;

[0028] Size of soil environmental parameter series The subscript B indicates the total number of sensors used to measure environmental parameters of the environment where the soil-covered storage tank is located.

[0029] Size of the electrochemical parameter sequence The subscript C indicates the total number of all sensors used to measure the electrochemical parameters of the environment in which the soil-covered storage tank is located.

[0030] Size of the structural state parameter sequence The subscript D indicates the total number of sensors used to measure the structural state parameters of the environment where the soil-covered storage tank is located.

[0031] In addition, the service environment information is sampled in time sequence, if the first... The first type of sensor Each sensor node at the current sampling time The current sampling reading is represented by a single current sampling reading, which contains multi-source heterogeneous monitoring data of soil environmental parameter sequences, electrochemical parameter sequences, and structural state parameter sequences.

[0032] Step 23B: Align sensor data of the same type according to the time scale;

[0033] No. The first type of sensor The average value of each sensor node during the monitoring period is , This represents the total number of samples.

[0034] Step 23C: Calculation and weighted aggregation of attention weights within the same type of sensor;

[0035] No. The attention weight row vector of the information collected by the type of sensor is ; column vector The transpose of , where the superscript T is the transpose character. The weight matrix is ​​a learnable matrix. For biased row vectors, Temperature coefficient, used for regulation Output distribution sharpness hyperparameters;

[0036] Step 23D: Intra-layer weight generation;

[0037] Different weights are assigned to the information collected by different types of sensors to generate corrosion feature vectors deployed in different layers;

[0038] When there is sufficient training data, learnable adaptive weights can be used. , For row vectors, Let be a trainable coefficient matrix. These are trainable bias row vectors;

[0039] When there is sufficient training data, learnable adaptive weights can be used. , For row vectors, Let be a trainable coefficient matrix. These are trainable bias row vectors;

[0040] Step 23E, generation of the corrosion feature matrix;

[0041] No. Intra-layer weight row vector With the Feature row vectors of the layer The matrices have the same size, and by using the Hadamard product, we obtain the first... Corrosion characteristics of the layer , This is the element-wise product symbol.

[0042] Furthermore, the feature construction module (31) of the corrosion kinetic physical mechanism first selects the power law model as the basic physical model based on the engineering characteristics of the soil-covered storage tank; then, using the Butler-Volmer equation as the basic equation of electrochemical kinetics, the power law model is modified and characterized, and its modification is as follows: , The equivalent corrosion current is represented by the subscript *corr*, which indicates the corrosion process, and the subscript *eq*, which indicates the equivalent meaning. The term represents the service time, with the superscript n indicating the time index. As an environmental coupling factor, The corrosion depth is represented by the subscript max, which indicates the maximum depth. Finally, the equivalent corrosion current, time index, and environmental coupling factor are designed as physical intermediate parameters of MT-PINN, and their mapping relationship with the input features is clarified. That is, soil environmental parameters and electrochemical parameters mainly affect the equivalent corrosion current, while soil environmental parameters and structural state parameters jointly affect the time index, etc.

[0043] Furthermore, the multi-task neural network prediction module (32) first feeds the total feature matrix into a main prediction branch and three physical quantity constraint auxiliary branches simultaneously; then, it uses a multi-layer fully connected neural network as the basic network unit and differentiates the functional differences between the main and auxiliary branches; finally, it constructs a weighted fusion total loss function, combining data loss, physical loss, monotonic loss, and regularization term, resulting in a total loss function. The mathematical expression is .

[0044] Furthermore, three physical auxiliary constraint branches: these auxiliary branches are the core carriers into which the physical mechanisms of corrosion kinetics are embedded, including the electrochemical quantity branch. Environmental compensation factor branch Time exponential parameter branch The logarithmic equivalent corrosion current is calculated respectively. Logarithmic environmental coupling factor The three major physical intermediate parameters are: time index n;

[0045] Each branch adopts an independent network structure and a differentiated activation function to adapt to the value characteristics of different physical quantities. Furthermore:

[0046] (1) Electrochemical quantity branch With environmental compensation factor branch All of them use the tanh activation function to adapt to the continuous variation characteristics of electrochemical corrosion parameters and environmental material coupling factors;

[0047] (2) Time exponential parameter branch Using the sigmoid activation function and incorporating prior knowledge of corrosion kinetics, a linear transformation is employed to impose a strict constraint on the output, limiting the time exponent n to a specific value. Within a physically reasonable range, avoid unreasonable values ​​that exceed the physical boundaries when training the network.

[0048] Compared with the prior art, the technical advantages of the present invention are as follows:

[0049] (1) Deployment of different types of sensors in the environment of the soil-covered storage tank: First, based on the structural characteristics of the soil-covered storage tank and the physical characteristics of the external soil-covered environment, a multi-dimensional, hierarchical spatially differentiated monitoring system is established. In terms of sensor selection, three types of complementary sensor groups are arranged in a coordinated manner, including: soil environmental parameter sensors, electrochemical parameter sensors, and structural state sensors. In terms of spatial layout planning, given the non-uniform distribution of corrosion influencing factors on the outer wall of the tank along the height of the tank, this invention divides the monitoring area into three levels: upper, middle, and lower, and implements a differentiated deployment strategy. Through the multi-parameter fusion of different types of sensors and the differentiated deployment of the three-layer tank space, the problems of the closed and invisible soil-covered environment and the uneven distribution of corrosion influencing factors with burial depth are effectively solved.

[0050] (2) Combination of Physical Information Neural Network (PINN), Multi-Branch Neural Network (MBNN) and Composite Loss Function: Preprocessing of raw data based on an edge computing architecture is implemented, and a multi-dimensional data fusion MT-PINN model based on spatial-three-source-attention weighted fusion is constructed. This MT-PINN model aims to utilize edge computing to complete preliminary processing at the data source close to the monitoring site, thereby improving the computing speed of the soil-covered tank corrosion monitoring system and reducing the bandwidth pressure on cloud transmission. Further outlier removal is performed on the collected service environment information to eliminate environmental noise interference, and time-scale unification processing is executed to ensure temporal synchronization, ultimately generating a spatiotemporally aligned corrosion feature matrix with clear physical meaning.

[0051] (3) The MT-PINN model introduces physical constraints to make the learning process consistent with the underlying physical mechanism of the corrosion process, thereby solving the problem of insufficient prediction reliability of traditional data-driven methods in complex corrosion environments. Through multi-dimensional loss weighting constraints, the model retains the advantage of fitting complex multi-source data and has solid physical mechanism support. It can still maintain high-precision prediction under small sample conditions, which greatly improves the engineering practicality of online monitoring of corrosion rate of soil-covered storage tanks. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0053] Figure 1 This is a structural block diagram of a soil-covered storage tank corrosion monitoring system based on MT-PINN according to the present invention.

[0054] Figure 2 It is a cross-sectional schematic diagram of the soil-covered storage tank buried in the soil and the distribution of various sensors.

[0055] Figure 3 This is a structural diagram of the MT-PINN model of the present invention.

[0056] Figure 4 This is a flowchart of the corrosion feature processing method using attention-weighted fusion and spatiotemporal alignment screening in this invention.

[0057] Figure 5 These are schematic diagrams illustrating corrosion prediction results; (a) prediction results based on an artificial neural network model; (b) prediction results based on an MT-PINN model. The gray and red dots in the diagrams correspond to the training and test sets, respectively.

[0058] Figure labels: 1. Information acquisition unit; 2. Feature preprocessing unit; 3. Corrosion rate online monitoring unit based on MT-PINN model; 31. Feature construction module based on corrosion kinetic physical mechanism; 32. Multi-task neural network prediction module. Detailed Implementation

[0059] To enable those skilled in the art to better understand the technical solutions of this invention, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The listed parameters are merely examples of preferred embodiments of this invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

[0060] See Figure 1 As shown, this embodiment of the invention provides a corrosion monitoring system for soil-covered storage tanks based on MT-PINN, which specifically includes:

[0061] Information acquisition unit 1 is used to acquire service environment information of the environment where the soil-covered storage tank is located; further, the information acquisition unit 1 includes the deployment of multiple sensors of different types; and the assembly of service environment information sequence;

[0062] The corrosion feature preprocessing unit 2 uses attention-weighted fusion and spatiotemporal alignment to filter the multi-source heterogeneous service environment information sequence; and

[0063] The corrosion rate online monitoring unit 3 based on the MT-PINN model further includes a feature construction module 31 based on corrosion kinetic physical mechanisms and a multi-task neural network prediction module 32. These modules are used to monitor the corrosion rate of the corrosion feature information output by the corrosion feature preprocessing unit 2 using the MT-PINN model, thereby obtaining corrosion prediction results for the soil-covered storage tank.

[0064] Furthermore, the standard structure of a Physical Information Neural Network (PINN) is: input layer (spatial, temporal coordinates, etc.) – neural network layer (usually a fully connected network) – output layer (predicted physical field). The neural network layer employs a Multi-Branch Neural Network (MBNN) and has multiple hidden layers. In this invention, the combination of a Physical Information Neural Network (PINN), a Multi-Branch Neural Network (MBNN), and a composite loss function is referred to as the MT-PINN model. The loss function obtained by training the MT-PINN model includes the residuals of the corrosion kinetic physical information equation, boundary conditions, initial conditions, etc. In this invention, due to the use of different types of sensors to sample environmental information of the soil-covered storage tank, the sampled multi-source heterogeneous data exhibits different types. Based on the Butler-Volmer equation, a corrosion kinetic physical information model belonging to the electrochemical process of the soil-covered storage tank and its metal corrosion prevention is constructed.

[0065] Furthermore, a single network can be used to simultaneously predict multiple physical quantities (such as soil environmental parameters). Electrochemical parameters Structural state parameters The MT-PINN employs a multi-output head architecture to realize the calculation of corrosion rate of soil-covered storage tanks by processing multiple samples using a multi-branch MT-PINN.

[0066] Training the MT-PINN model

[0067] The present invention uses the following neural network structure to construct the training MT-PINN model. The following content is only a specific implementation case and does not constitute a limitation on the scope of protection of the present invention.

[0068] In this example, the Physical Information Neural Network (PINN) model based on the Multi-Branch Neural Network (MBNN) architecture adopts a parallel structure design of the main network and three branch networks. Training is performed by inputting training data into the MT-PINN model, and the training steps are as follows:

[0069] S101, the main network processes the training data through two hidden layers (8 neurons per layer, with tanh activation function) and then directly outputs the logarithm of the erosion depth. The logarithm of the corrosion depth reflects the master-major network relationship. ;

[0070] S102, the first branch network extracts the logarithm of the corrosion current density from the training data. And expressed as a logarithm of corrosion current density Correcting the electrochemical terms in the MT-PINN model;

[0071] S103, the second branch network transforms the training data using the Sigmoid activation function and linearly maps it to the interval. The corrosion kinetic index n is obtained; it is used to reflect the transition characteristics of the corrosion mechanism from diffusion control to electrochemical control.

[0072] S104, the third branch network extracts the logarithmic value of the environmental coupling factor from the training data. The environmental coupling factor reflects the coupled influence of factors such as soil moisture content, ion concentration, and coating condition on the corrosion rate.

[0073] This invention achieves the embedding of the physical model through a composite loss function. Specifically, the MT-PINN model simultaneously learns two paths of corrosion depth prediction: one path directly outputs the logarithm of the maximum corrosion depth prediction value driven by data from the main network. and will Assign to ,Right now The output of the other route's three branch networks is used to obtain the logarithm of the predicted maximum corrosion depth based on a physics-driven model. ,and . denoted as the logarithm of the corrosion current density. n represents the corrosion kinetic index. This is the logarithm of the service time. This is the logarithmic value of the environmental coupling factor.

[0074] In this invention, the total loss function It consists of four parts:

[0075] (1) Mean square error between predicted value and actual monitored value The erosion depth used to drive the network to fit the training samples. Further... . This is the measured value of the maximum corrosion depth.

[0076] (2) Physical consistency error Used to force the main network output to match the calculation results of the physical equations, so that the network learns the equivalent corrosion current. Time exponent n, environmental coupling factor It conforms to the laws of corrosion kinetics. Furthermore... .

[0077] (3) Monotonicity constraint error This is used to ensure that the corrosion depth increases monotonically over time. Further... mean indicates the sign of the average value. ReLU is the rectified activation function. This is the symbol for partial differentials.

[0078] (4) L2 regularization term Used to prevent network overfitting.

[0079] Then we have the total loss function The mathematical expression is .

[0080] This represents the mean square error between the predicted value and the actual monitored value.

[0081] This is an adjustable first weighting coefficient.

[0082] This represents the physical consistency error.

[0083] This is an adjustable second weighting coefficient.

[0084] This is a monotonicity constraint error.

[0085] This is an L2 regularization term.

[0086] Information collection unit 1 of the environment where the soil-covered storage tank is located

[0087] To address the technical challenges faced by liquefied hydrocarbon (LNG) covered storage tanks in complex soil environments, including diverse corrosion mechanisms, heterogeneous monitoring data from multiple sources, and discrepancies in the compatibility between data and physical models, this invention first divides the LNG covered storage tank into three sections along its height: an upper tank, a middle tank, and a lower tank. Figure 2 As shown. Then, different types of sensors are installed in different zones. These sensors are used to acquire environmental information (referred to as service environment information) about the location of the soil-covered storage tank.

[0088] 11. Deploy sensors

[0089] like Figure 2 The coordinate system of the soil-covered storage tank shown is denoted as The origin O is the center point of the soil-covered storage tank, and the coordinate system is established according to the right-hand rule. Different types of sensors deployed on the outer wall of the soil-covered storage tank are arranged in a grid, and the identification number and location of each sensor are recorded. All deployed sensors are described using mesh nodes, with any sensor node labeled as... The distance between the outer wall of the soil-covered storage tank and the sensor in the soil is denoted as q, and the sensor distance q is generally set to 1 meter.

[0090] For the partitioning of soil-covered storage tanks along the height of the tank, such as Figure 2 As shown, the tank is divided into three layers: the upper tank... Middle tank and lower tank Eight different types of sensors are deployed on each layer. These include sensors for measuring the ambient temperature of the soil-covered tank, the ambient humidity of the soil-covered tank, the soil resistivity of the soil in the soil-covered tank (such as resistance probes), the electrode potential sensors of the outer wall of the soil-covered tank (such as long-term buried reference electrodes, online corrosion potential sensors, etc.), the tank current density sensors of the outer wall of the soil-covered tank (such as coupled multi-electrode sensors (CMS), linear polarization resistance (LPR) sensors, corrosion current density monitors), the tank strain sensors of the outer wall of the soil-covered tank, the tank coating condition sensors of the outer wall of the soil-covered tank, and the sensors of the acoustic emission amplitude of the outer wall of the soil-covered tank.

[0091] Service environment information collected by different types of sensors Including soil environmental parameters Electrochemical parameters and structural state parameters That is, the service environment information is a three-element information group. or .

[0092] Soil environmental parameters are denoted as The Including but not limited to temperature ,humidity Soil resistivity Etc., that is, soil environmental parameters are a ternary information set. Based on the partitioning along the height of the tank, there is a set of soil environmental parameters. , These are the soil environmental parameters for the upper tank. These are the soil environmental parameters for the middle tank layer. These are the soil environmental parameters for the lower tank.

[0093] Electrochemical parameters The Corrosion potential acquired by electrode potential sensors, including but not limited to Current density collected by current density sensor Etc., that is, electrochemical parameters are a binary information set. Based on the partitioning along the height of the tank, there is a set of electrochemical parameters. , These are the electrochemical parameters of the upper tank. The electrochemical parameters of the middle tank are as follows: These are the electrochemical parameters of the lower tank.

[0094] Structural state parameters The Including but not limited to strain Coating condition Acoustic emission amplitude Etc., that is, the structural state parameters are a ternary information group. Based on the partitioning along the height of the tank, there is a set of structural state parameters. .

[0095] In this invention, the classification of coating states in training the MT-PINN model is shown in the table below.

[0096]

[0097] In training the MT-PINN model, the hierarchical scores of coating states are used to achieve semantic numericalization. Since text cannot be directly input into the model, mapping it to numerical values ​​can transform qualitative descriptions into quantitative features or label descriptions.

[0098] 12. Construction of Service Environment Information Sequence

[0099] Record the current sampling time as Located at the current sampling time The previous time point is recorded as the previous sampling time. Located at the current sampling time The time after that is recorded as the next sampling time. .

[0100] At the current sampling time The sampled service environment information sequence is .

[0101] Corrosion Feature Pretreatment Unit 2

[0102] See Figure 3 As shown, the corrosion feature preprocessing unit 2 performs attention-weighted fusion and spatiotemporal alignment processing on the received service environment information to obtain corrosion feature information, which is then output to the corrosion rate online monitoring unit 3 based on the MT-PINN model.

[0103] Multiple sensors of different types were deployed in the upper, middle, and lower sections of the outer wall of the soil-covered storage tank and in the surrounding soil, forming a sensor network. The sensor data was categorized into three types: soil environmental parameters, etc. Electrochemical parameters and structural state parameters .

[0104] The soil environmental parameters By influencing the electrochemical reaction kinetics and mass transfer conditions on metal surfaces, these factors exert fundamental constraints on corrosion behavior. For example, increased temperature accelerates anodic dissolution and cathodic reaction rates, altering the activation energy of the corrosion reaction; soil moisture affects electrolyte conductivity and the electrochemical reaction pathway on the metal surface; soil resistivity reflects the ion migration capacity of the medium and is significantly correlated with corrosion current density and corrosion morphology. These three factors, as fundamental environmental parameters, determine whether corrosion reactions occur and their potential severity, playing a fundamental role in the formation and variation of corrosion current density. Therefore, they primarily affect the equivalent corrosion current parameter and indirectly modulate the trend of the time exponent.

[0105] The electrochemical parameters This method directly characterizes the electrochemical state during metal corrosion. Electrode potential reflects whether the metal is in an activated, passivated, or accelerated corrosion state; current density characterizes the development of localized corrosion, pitting corrosion, or uniform corrosion. These parameters exhibit a highly nonlinear coupling relationship. Directly measuring the kinetic parameters of the corrosion reaction on the metal surface allows for the quantification of the corrosion rate at the current moment, serving as a primary determinant of the equivalent corrosion current parameter. Simultaneously, it participates in characterizing the stage-specific changes in the corrosion process and exerts a constraint on the value of the time exponent.

[0106] The structural state parameters By altering the exposure conditions and strain state of the metal surface, the corrosion process is coupled with other factors. For example, strain changes may induce stress corrosion susceptibility; coating integrity directly determines the degree of contact between the electrolyte and the metal substrate; acoustic emission characteristics can reflect the propagation of microcracks, coating failure, or the formation of localized active corrosion zones, affecting the spatial non-uniformity and evolution path of the corrosion process. Thus, these factors primarily act on the environmental-material coupling factor and regulate the evolution of the time exponent.

[0107] Despite soil environmental parameters While it can indicate corrosion tendency, relying solely on it to predict corrosion rate is clearly incomplete, because under the same environmental conditions, the protective state, electrochemical activity, and accumulated damage level of a metal surface can vary significantly. Electrochemical parameters While it can provide real-time data rates, it cannot reflect historical cumulative damage and overall distribution; structural state parameters While it can provide damage results, it cannot describe the continuous evolution of corrosion damage. Therefore, this invention integrates these three types of data, and their complementarity allows for the formation of a complete corrosion information chain from cause to process to result.

[0108] Information on service environment Corrosion feature preprocessing is performed through edge computing nodes. This preprocessing includes outlier removal, time scale unification, and multi-source data fusion. Corrosion feature preprocessing improves data quality, reduces transmission volume, and increases response speed.

[0109] 21. Outlier removal processing in corrosion characteristics

[0110] Data from the service environment information that includes over-range measurements, abnormal drift rates, or sensor outages will be removed. For example, data collected by the temperature sensor that is higher than [a certain value] will be removed. Temperature values ​​are discarded. Potential drift measured by the electrode potential sensor is greater than... Values ​​that are removed. Acoustic emission noise spikes exceeding 10 times the background amplitude are removed.

[0111] 22. Time alignment and frequency uniformity

[0112] Different types of sensors may have different data acquisition frequencies, which are unified through the following mechanism.

[0113] a. Set the time resolution to 1 minute;

[0114] b. High-frequency data (such as strain) are averaged within a window;

[0115] c. Low-frequency data (such as electrochemistry) are obtained using time-holding or linear interpolation.

[0116] 23. Multi-source data fusion edge preprocessing

[0117] Step 23A: Divide the dataset into three layers: upper, middle, and lower.

[0118] In the three sections of the tank body on the outer wall of the storage tank (such as...) Figure 2 As shown, eight different types of sensors are deployed. The sensors collect information including temperature, humidity, soil resistivity, electrode potential, current density, strain, coating condition, and acoustic emission amplitude. Any type of sensor is labeled as... ,and The aforementioned Also known as the sensor type label.

[0119] Numerical structuring was used to organize the monitoring data collected by different types of sensor nodes, constructing parameter matrices for multi-source heterogeneous monitoring data (soil environmental parameter sequences, electrochemical parameter sequences, and structural state parameter sequences). Each row of the matrix corresponds to a type of environmental physical parameter, and each column corresponds to the monitoring value of a single sensor node. Each layer of the dataset contains all three types of parameters. The specific format is as follows: [Size of soil environmental parameter sequence]. The subscript B indicates the total number of sensors used to measure environmental parameters of the environment surrounding the soil-covered storage tank. The size of the electrochemical parameter sequence... The subscript C indicates the total number of sensors used to measure the electrochemical parameters of the environment surrounding the soil-covered storage tank. The size of the structural state parameter sequence. The subscript D indicates the total number of sensors used to measure the structural state parameters of the environment in which the soil-covered storage tank is located.

[0120] In this invention, service environment information To sample sequentially, if the first... The first type of sensor Each sensor node at the current sampling time The reading is denoted as This is simply referred to as the current sampled reading. A current sampled reading... It contains multi-source heterogeneous monitoring data of soil environmental parameter sequences, electrochemical parameter sequences, and structural state parameter sequences.

[0121] Step 23B: Align sensor data of the same type according to the time scale;

[0122] Because the sampling frequencies of different types of sensors monitoring various states of the storage tank are inconsistent, the average value over a single sensor time period is used. For a certain layer (taking any layer as an example, omitting the superscript), the total sampling time is... within, no. The first type of sensor The time-period data of each sensor node is taken as an arithmetic mean (e.g., daily or weekly average). , This represents the total number of samples. It is simply referred to as the average value (scalar) within the monitoring period.

[0123] The first The average values ​​of the various types of sensors are grouped into a column vector, denoted as... For example, a temperature column vector formed by concatenating information collected by temperature sensors. Humidity column vector acoustic emission amplitude column vector .

[0124] Step 23C: Calculation and weighted aggregation of attention weights within the same type of sensor;

[0125] In the monitoring of liquefied hydrocarbon covered storage tanks, multiple sensor nodes are distributed in different spatial locations within the same layer of the tank, and their numerical values ​​directly reflect the state differences in that local area. The core purpose of assigning weights based on numerical values ​​is to ensure that sensors with higher values ​​(corresponding to areas of abnormal temperature increases) contribute more during fusion, thereby focusing attention on sensor nodes that may have leaks, overheating, or other safety hazards, rather than averaging all service environment information. This mechanism utilizes nonlinear functions (such as...) This approach can both amplify the weight difference between high and low values, making abnormal signals stand out during feature extraction, and maintain the normalization characteristics of the weights, ensuring that the fused corrosion features still have clear physical meaning.

[0126] No. The attention weight row vector of the information collected by the type of sensor is . column vector The transpose of , where the superscript T is the transpose character. The weight matrix is ​​a learnable matrix. For biased row vectors, Temperature coefficient, used for regulation Output the hyperparameters of the distribution sharpness.

[0127] The scoring function in the further attention mechanism is: Its core function is to calculate the average value within the monitoring period. Convert it into a comparable attention score, and then... The final weight distribution is generated using the temperature coefficient.

[0128] use The function normalizes the attention weights of the collected information, ensuring that the sum of the weights is 1 and maintaining the weighted average meaning of the physical quantities. It also amplifies the differences between different sensor values ​​through an exponential function, allowing sensor nodes experiencing extremely high temperatures or sudden pressure changes to receive significantly higher weights, thus enabling focused attention on important sensor values. Simultaneously, The differentiability of the function makes the weight matrix... and bias row vector It can perform end-to-end optimization through backpropagation of neural networks, enabling attention weights to adaptively learn the combination of sensor nodes most relevant to tank safety.

[0129] Weighted aggregation of data collected by sensors of the same type yields sampled data of the same type. ,Right now Deployed in the upper tank. The upper-level row vector is composed of eight types of scalars aggregated from all sensors. ,and Deployed in the middle tank The aggregated data of all sensors, consisting of eight types of scalars, forms the middle-level row vector. ,and Deployed in the lower tank. The lower-level row vector is composed of eight types of scalars aggregated from all sensors. ,and .

[0130] Step 23D: Intra-layer weight generation;

[0131] Different weights are assigned to information collected by different types of sensors to generate corrosion feature vectors deployed in different layers. This is based on physical quantities (soil environmental parameters). Electrochemical parameters and structural state parameters The importance of tank safety monitoring within the same layer is considered, and quantities are selectively integrated to avoid treating all physical quantities equally. Key information is highlighted while secondary or interfering information is suppressed. For example, upper tanks are closer to the ground surface, experiencing drastic temperature fluctuations and being greatly affected by environmental climate; therefore, temperature should have a higher weight. Lower tanks, closer to the sand bed, experience high humidity and bear the weight of the tanks and the pressure of the medium; therefore, pressure, strain, and humidity should have higher weights. Thus, the generation of weights within each layer includes both fixed prior weights and learnable adaptive weights.

[0132] (1) Fixed prior weights

[0133] Based on engineering experience, physical mechanisms, or design specifications, fixed prior weights for each physical quantity are pre-defined. subscript The markings indicate the upper, middle, and lower layers of the tank, and subscript The label indicating the sensor type, and And remain constant throughout the prediction process. The sum of the prior weights is fixed. The normalization constraint ensures the probabilistic meaning of the weights.

[0134] Fixed prior weights of the upper tank The mathematical expression is ,and .

[0135] Fixed prior weights of the middle tank The mathematical expression is ,and .

[0136] Fixed prior weights of the lower tank The mathematical expression is ,and .

[0137] If fixed weights cannot adapt to changes in operating conditions, such as seasonal humidity variations that may temporarily increase the importance of humidity, then learnable adaptive weight generation can be used to achieve dynamic changes in the weights of sensor-collected information.

[0138] (2) Learnable adaptive weights

[0139] Learnable adaptive weights refer to weights that are automatically learned from service environment information via artificial neural networks. The optimal weights for each physical quantity are learned, and these weights are dynamically adjusted based on data distribution and task objectives. When training within an artificial neural network architecture, trainable bias row vectors need to be introduced. (dimension is) and trainable coefficient matrix (dimension is) By iterating through the bias row vectors and coefficient matrices, the dynamic characteristics of the service environment information can be adapted, ultimately achieving a fit to the changing process. The learnable adaptive weights are... The learnable adaptive weight model is ; Subscript The markings indicate the upper, middle, and lower layers of the tank, and , For the first Learnable adaptive weights of a layer (also known as in-layer weight row vectors). For the first The row vector of the layer. (Used) It replaces the row vectors of the upper, middle, or lower layers, i.e. , and .

[0140] Learnable adaptive weights of the upper tank The mathematical expression is .

[0141] Learnable adaptive weights for the middle tank The mathematical expression is .

[0142] Learnable adaptive weights of the lower tank The mathematical expression is .

[0143] In this invention, when the amount of training data is small, fixed prior weight values ​​are selected. When the amount of training data is sufficient, the artificial neural network can automatically learn the importance of physical quantities under specific operating conditions and discover the correlations between various parameters (e.g., the coupling between humidity and strain at a specific time). The operating status of the storage tank changes over time (e.g., seasonal changes, medium changes), and the adaptive weights can be adjusted in real time. Adaptive weights typically achieve higher prediction accuracy than fixed weights, especially in complex, nonlinear scenarios. Weight learning is jointly trained with the final prediction target, allowing the weights to directly serve the final task rather than being set in isolation.

[0144] Step 23E, generation of the corrosion feature matrix;

[0145] No. Intra-layer weight row vector With the Feature row vectors of the layer The matrices have the same size, and the Hadamard product (element-by-element product) is used to obtain the first... Corrosion characteristics of the layer , The symbol is for element-wise product. The total corrosion characteristic information is obtained by summing the corrosion characteristic information of each layer in the tank section. Total characteristic matrix The dimension is .

[0146] In this invention, the corrosion feature preprocessing unit 2 first compresses the continuous sampling of each sensor into the mean value, then fuses multiple sensors of the same type into a single representative value through an attention mechanism, and assigns weights to each layer of feature matrices, finally compressing the three layers of multi-physical quantity information into a feature matrix. This supervised dimensionality reduction process significantly reduces the number of input parameters for corrosion feature preprocessing while retaining key information, thereby improving generalization ability.

[0147] Furthermore, a numerical magnitude-based approach is introduced during the corrosion feature extraction process. Attention mechanism: When a sensor reading is significantly high, its weight is amplified exponentially, making the final feature more sensitive to abnormal hotspots and preventing the signal from being overwhelmed by averaging. The attention weight itself also provides interpretable evidence, allowing for the tracing of contributions from key sensors.

[0148] The feature matrix obtained after processing by the corrosion feature preprocessing unit 2 has a high signal-to-noise ratio, and the neural network is trained faster and requires fewer samples. Since the physical meaning of the features is clear, the network optimization is more stable, effectively avoiding the gradient vanishing or exploding problem.

[0149] Through the above mechanism, this embodiment achieves a unified time reference for heterogeneous sensors, high-quality fusion of multi-source data, and data offloading and preprocessing at the edge.

[0150] Corrosion Rate Online Monitoring Unit 3 Based on MT-PINN Model

[0151] See Figure 1 , Figure 3 , Figure 4 As shown, the corrosion rate online monitoring unit 3 based on the MT-PINN model includes a feature construction module 31 based on the physical mechanism of corrosion kinetics and a prediction module 32 for constructing a multi-task neural network.

[0152] In this invention, the total feature matrix output by the corrosion feature preprocessing unit 2 (containing environmental, electrochemical, and structural features collected and fused from multiple sensors) serves as input information for training or monitoring the MT-PINN model. This input information is used in the online corrosion rate monitoring unit 3 based on the MT-PINN model for training. After training, the MT-PINN model can directly output a predicted corrosion depth of the soil-covered storage tank, or its logarithmic form, for the currently input corrosion feature matrix, and can be further converted into a corrosion rate. This output result possesses both data-driven accuracy and interpretability based on physical mechanisms.

[0153] Feature construction module 31 based on corrosion kinetics physical mechanism

[0154] The core of the feature construction module 31 based on corrosion kinetic physical mechanisms lies in the construction of the corrosion kinetic physical model and the design of intermediate parameters. It converts the preprocessed multi-source heterogeneous corrosion feature data from the corrosion feature preprocessing unit 2 into intermediate parameters that can be directly substituted into the MT-PINN model. These intermediate parameters must have clear physical meaning and be directly used as input parameters for the corrosion kinetic model, thereby ensuring that the MT-PINN model possesses clear physical mechanism characteristics and avoiding becoming a purely black-box model without physical meaning.

[0155] Further design of intermediate parameters and construction of physical models can effectively constrain multi-branch neural networks using physical mechanisms, ultimately improving prediction accuracy under small sample conditions and enhancing the generalization ability of the MT-PINN model.

[0156] In this invention, the feature construction module 31 based on the physical mechanism of corrosion kinetics completes the process of decoupling the input corrosion features into three intermediate parameter targets with clear physical meaning through the following steps.

[0157] Step 31A: Selection of corrosion kinetic physical mechanism and adaptation to working conditions;

[0158] The selection of corrosion kinetic physical mechanisms and the adaptation to operating conditions are the foundational steps for physical modeling. The core function is to combine the engineering characteristics of soil corrosion in covered storage tanks to complete the screening of basic physical models and the analysis of the defects of traditional models. This provides a theoretical basis and technical foundation for the correction of the MT-PINN model and mechanism coupling, ensuring the rationality and adaptability of the overall physical model.

[0159] The following section describes the method for constructing corrosion characteristics using a typical physical model. Power-law distribution is one of the most prevalent laws in nature and engineering (Research on Multivariate Econometric Aggregation of Scientific Knowledge, Dong Ke, Wuhan University Press, May 2017, 1st edition, pp. 54-55). Therefore, in this embodiment, for the long-term corrosion evolution law of soil-covered storage tanks in complex soil environments, this invention selects a power-law model as the physical model driven by a Physical Information Neural Network (PINN).

[0160] Step 31B: Correction of the physical model based on the electrochemical corrosion mechanism;

[0161] To achieve a deep integration of empirical evolution models and corrosion kinetic mechanisms, the physical model based on electrochemical corrosion mechanisms is modified by mechanistic empowerment of traditional power-law models. This results in the construction of a corrosion kinetic physical model adapted to the soil of covered storage tanks, solving the core problems of traditional models lacking physical mechanisms and having poor adaptability to operating conditions.

[0162] Furthermore, although the power-law model can characterize the change of corrosion rate over time, it has two inherent defects that make it unsuitable for the engineering scenario of the liquefied hydrocarbon covered storage tank of this invention: First, the power-law model only considers the single-dimensional influence of time, completely ignoring the coupling factors of complex service environments such as soil environment and storage tank structure, which is seriously inconsistent with the actual multi-field coupled corrosion conditions of soil; Second, the traditional power-law model is a purely empirical statistical model, which does not combine the physical mechanism of corrosion kinetics and has no physical essence support. It can only achieve data fitting, resulting in a large deviation in the prediction results. It cannot output corrosion depth and corrosion rate parameters that conform to the actual working conditions of the liquefied hydrocarbon covered storage tank, and it is difficult to meet the multi-objective engineering monitoring needs. Therefore, after the selection of the basic corrosion kinetic model and the adaptation of the working conditions, a physical model correction and compensation based on the electrochemical corrosion mechanism was set up.

[0163] The Butler-Volmer equation, as the fundamental equation of electrochemical kinetics, reveals the exponential response relationship between electrode potential and current density. Corrosion current density is the physical essence of corrosion rate, directly reflecting the real-time corrosion intensity. To imbue the power-law model with corrosion kinetics, this invention deeply couples the scaling factor of the power-law model with the corrosion current density in the Butler-Volmer equation, defining the scaling factor as a function of the equivalent corrosion current density. Based on this, this invention modifies and characterizes the fundamental power-law model, resulting in the following modification: , The equivalent corrosion current is represented by the subscript *corr*, which indicates the corrosion process, and the subscript *eq*, which indicates the equivalent meaning. The term represents the service time, with the superscript n indicating the time index. As an environmental coupling factor, The value represents the corrosion depth, with the subscript "max" indicating the maximum depth.

[0164] The core innovation of the physical model correction based on electrochemical corrosion mechanism lies in the coupling of empirical evolution law (power law distribution) with corrosion kinetic mechanism, upgrading the traditional black box empirical model into a mechanism-driven model, giving the model a clear physical essence of corrosion, breaking through the limitation of traditional models that only rely on time as a single variable, and incorporating multiple corrosion influencing factors such as time, electrochemistry and environmental factors into the physical model.

[0165] Step 31C: Construct physical intermediate parameter decoupling and mapping for MT-PINN;

[0166] To achieve deep adaptation between the physical model and the MT-PINN multi-task neural network, this invention uses three core parameters from the physical model modified based on the electrochemical corrosion mechanism—equivalent corrosion current density, corrosion time exponent, and environmental coupling factor—as intermediate physical parameters of the MT-PINN model. All parameters possess independent and clearly defined physical meanings related to corrosion, eliminating the "black box" problem of neural network parameters lacking physical meaning. Simultaneously, the multi-source input parameters and intermediate parameters in the corrosion feature preprocessing unit 2 form the following mapping relationship: soil environmental parameters... and electrochemical parameters The combined effects affect the equivalent corrosion current parameter; soil environmental parameters With structural state parameters The time index and soil environmental parameters work together. The environmental coupling factor is the main influencing factor, and it is related to the structural state parameters. This creates a coupling effect.

[0167] This design ensures that the physical model, while maintaining its versatility, is deeply compatible with the service environment information of soil-covered storage tanks, thereby improving the accuracy of corrosion rate prediction and the model's generalization ability. The modified power-law model not only compensates for the shortcomings of traditional time-based univariate models that do not consider the environment and structure, but also achieves an effective combination of corrosion data and corrosion mechanisms, demonstrating the innovation of this invention.

[0168] Constructing a multi-task neural network prediction module 32

[0169] The multi-task neural network prediction module 32 serves as the intelligent prediction core of the MT-PINN model, inheriting the physical model and intermediate parameter system of the aforementioned feature construction module 31 based on corrosion kinetic physical mechanisms. Addressing the industry shortcomings of traditional pure data-driven neural networks—single output, lack of physical constraints, strong black-box nature, poor generalization ability with small samples, and weak interpretability—this module innovatively designs a master-slave parallel multi-task physical information neural network structure. The corrosion kinetic physical model constructed by the feature construction module 31 based on corrosion kinetic physical mechanisms is rigidly embedded into the entire network training process. This retains the neural network's nonlinear fitting ability for complex, multi-source heterogeneous data while standardizing the network learning logic through physical constraints, ultimately achieving high accuracy and high generalization of corrosion rate or corrosion depth under small sample conditions.

[0170] The present invention uses the following neural network structure to achieve the above prediction modeling. The following content is only a specific implementation case and does not constitute a limitation on the scope of protection of the present invention.

[0171] Step 32A: Construct a multi-task network;

[0172] In this invention, the core function of constructing a multi-task network is to address the inherent defects of traditional corrosion prediction neural networks by building an overall architecture of the MT-PINN multi-task network with backbone prediction and multi-branch physical constraints, clarifying the functional positioning and collaborative logic of each network branch, and realizing the integration of data-driven and physical mechanisms.

[0173] Existing corrosion prediction neural networks mostly adopt a pure data-driven structure with a single output, relying solely on sample data to fit mapping relationships without incorporating any prior knowledge of corrosion dynamics and physics. This has several technical drawbacks: First, the network training lacks physical constraints, making it prone to learning false features that do not conform to the corrosion mechanism, resulting in poor physical consistency of prediction results. Second, it cannot adapt to the corrosion evolution law coupled with multiple physical parameters, nor can it connect to multi-dimensional corrosion intermediate parameters. Third, under small sample conditions, the data features are insufficient, making the model prone to overfitting, significantly reducing generalization ability and prediction accuracy. Furthermore, the overall model is a pure black box structure with extremely poor interpretability, failing to meet the reliability requirements of engineering monitoring.

[0174] To address the aforementioned shortcomings, this invention constructs a physically constrained multi-task neural network (MT-PINN). The network input uniformly receives the total feature matrix of multi-source data preprocessed by the erosion feature preprocessing unit 2. The feature vector contains environmental, material, temporal, and electrochemical parameters related to the corrosion evolution of soil-covered storage tanks, collected by multiple sensors. The network adopts a parallel, multi-task structure with master-slave and multiple outputs, consisting of one main prediction branch and three physical quantity constraint auxiliary branches. The main prediction branch is responsible for the accurate prediction of the final corrosion target quantity, while the three auxiliary branches are responsible for solving the three core physical quantities output by module 31: logarithmic equivalent corrosion current. Time index Intermediate parameters and logarithmic environment coupling factor Through multi-task parallel learning, bidirectional learning of data feature fitting and physical mechanism restoration is achieved.

[0175] Step 32B: Main and auxiliary branch network structure and physical boundary constraints;

[0176] The main-auxiliary branch network structure and physical boundary constraints are designed to meet the different functional requirements of the main prediction branch and the three physical constraint branches. Different network levels, neuron configurations, activation functions and physical boundary constraints are designed to achieve accurate solutions for each physical parameter and ensure that the network output results fully conform to the objective laws of corrosion dynamics.

[0177] The main and auxiliary branch network structure and physical boundary constraints adopt a multi-layer fully connected neural network as the basic network unit. Differentiated designs are made for the functional differences between the main and auxiliary branches. The specific structural configuration and physical constraints are as follows:

[0178] First, the main prediction branch: The main branch is the core output unit of the MT-PINN model, primarily implementing the nonlinear mapping of input features and outputting the logarithmic form of the corrosion target response value, which is used as the core predictor of the tank corrosion depth. In this preferred embodiment, the main prediction branch has two hidden layers, each with 8 neurons, and uniformly uses the hyperbolic tangent function tanh as the activation function. The tanh activation function can effectively capture the nonlinear correlation features of corrosion data, while suppressing gradient oscillation and gradient vanishing problems during network training, improving the fitting accuracy of complex working condition data. Simultaneously, an L2 regularization term is introduced into the hidden layer to constrain the network weight parameters, effectively suppressing overfitting in small sample training scenarios and improving model stability.

[0179] Second, there are three physical auxiliary constraint branches: these auxiliary branches are the core carriers embedded in the physical mechanisms of corrosion kinetics, including the electrochemical quantity branch. Environmental compensation factor branch Time exponential parameter branch The logarithmic equivalent corrosion current is calculated respectively. Logarithmic environmental coupling factor The three major physical intermediate parameters are time exponent n, and each branch adopts an independent network structure and differentiated activation function to adapt to the value characteristics of different physical quantities. Furthermore:

[0180] (1) Electrochemical quantity branch With environmental compensation factor branch The tanh activation function is used to adapt to the continuous variation characteristics of electrochemical corrosion parameters and environmental material coupling factors, accurately extracting continuous features related to soil electrochemical corrosion and material-environment coupling, and ensuring the continuity and accuracy of physical parameter solutions.

[0181] (2) Time exponential parameter branch Using the sigmoid activation function and incorporating prior knowledge of corrosion kinetics, a linear transformation is employed to impose a strict constraint on the output, limiting the time exponent n to a specific value. Within a physically reasonable range, unreasonable values ​​that exceed physical boundaries in network training output are avoided, thus ensuring the physical validity of the network output from a structural perspective.

[0182] The core innovation of implementing the main-auxiliary branch network structure and physical boundary constraints lies in the differentiated design of the network structure and hard constraints of the physical boundary for different physical quantities. This is different from the free training mode of traditional neural networks that is irregular and without boundaries. It ensures that the output of each physical intermediate parameter fits the corrosion dynamic mechanism, providing accurate and reliable parameter support for subsequent physical consistency loss calculation, and greatly improving the rationality of the model prediction results.

[0183] Step 32C: Constructing physical consistency loss and training the model through multi-dimensional fusion;

[0184] The multi-dimensional fusion of physical consistency loss construction and model training is aimed at the loss function generation stage of the multi-task neural network prediction module 32. It constructs a multi-dimensional composite loss function that integrates physical consistency loss, monotonicity prior loss, and regularization loss by using the corrosion kinetic physical formula modified by the feature construction module 31 based on the corrosion kinetic physical mechanism. This realizes the full constraint of physical mechanism on network training, completes the closed-loop training of the MT-PINN model, and ultimately achieves the simultaneous improvement of model accuracy, generalization and interpretability.

[0185] Based on the revised core formula of corrosion kinetics Construct explicit physical constraints for The physical constraint paradigm converted to logarithmic form is .

[0186] During the forward propagation of the MT-PINN network, the backbone outputs the predicted logarithmic value of the corrosion. The three auxiliary branches synchronously output the theoretical physics prediction values ​​of the corresponding intermediate physical parameters. Data-driven predictions Compared with the theoretical value of physical mechanism We construct a physical consistency loss term to force the network learning process to always conform to the corrosion dynamics, thereby achieving explicit embedding of the physical mechanism.

[0187] Furthermore, this invention introduces a monotonicity constraint mechanism for the prior laws of corrosion evolution. Based on the long-term corrosion mechanism of soil-covered storage tanks, the corrosion depth monotonically increases with service time (ET), exhibiting no physical phenomenon of decay over time. Therefore, this module calculates the main output using automatic differentiation technology. Relative to the gradient of the service time ET, when a negative gradient is detected, a penalty constraint is automatically applied, and a monotonic loss term is constructed. This completely eliminates abnormal results from network outputs that violate the evolutionary laws of erosion time, and strengthens the physical rationality of model predictions.

[0188] To balance data fitting ability, physical constraint effectiveness, and model generalization performance, this invention constructs a weighted fusion total loss function, combining data loss, physical loss, monotonicity loss, and regularization term. The mathematical expression is .in, To ensure the accuracy of the model's data fitting, the mean square error between the predicted and actual monitored values ​​is calculated. The first weighting coefficient is adjustable. To address physical consistency errors, a mechanism constraint is implemented. The second weighting coefficient is adjustable. To constrain the monotonicity of the error and ensure compliance with the corrosion evolution trend; This is an L2 regularization term to suppress model overfitting. In this invention, a tolerance is introduced. As the total loss function The convergence criterion, if The predicted result is output if the expected result is not found; otherwise, the multi-branch neural network is returned. During training, a relative convergence criterion is used: when the change in the total loss function is less than the tolerance over a consecutive number of training cycles, the model is considered to have reached a stable convergence state. The tolerance is set based on the order of magnitude of the physical residuals. A tolerance is introduced to define the stopping condition of the training process: when the loss function value or its relative decrease is below a given threshold, the optimization is considered to have entered the effective convergence interval, and the iteration is terminated to avoid invalid computation or overfitting. By setting adjustable weight coefficients, the strength of physical constraints and monotonicity constraints can be adaptively adjusted according to different working conditions, adapting to corrosion monitoring scenarios with different soil environments and different sample numbers.

[0189] During the model training phase, a gradient descent-type optimization algorithm is used to minimize the total loss function. The network weights and bias parameters are continuously updated iteratively through the backpropagation mechanism until the loss function converges or the preset training termination condition is met, thus completing the model training.

[0190] The innovative value of multi-dimensional fusion of physical consistency loss construction and model training lies in the construction of a composite constraint system that integrates data-driven approaches, physical mechanisms, prior laws, and overfitting prevention. This solves the shortcomings of traditional data-driven methods, such as single constraints, low physical fit, and susceptibility to failure in small samples. Through multi-dimensional loss weighted constraints, the model retains its advantage in fitting complex multi-source data while possessing solid physical mechanism support. It can still maintain high-accuracy predictions under small sample conditions, greatly improving the engineering practicality of online monitoring of corrosion rates in soil-covered storage tanks.

[0191] Performance testing of MT-PINN application

[0192] Using PyCharm 2025.1 software tools Corrosion rate monitoring is performed on liquefied hydrocarbon tanks with soil covering. Twenty different types of sensors are deployed in each layer.

[0193] First, the corrosion feature information is used as input to the MT-PINN model and divided into training and testing sets in an 8:2 ratio. Model training employs the Adam optimizer to minimize the composite loss function, which balances the supervisory loss and physical consistency residuals through weighting factors. The convergence criterion is that the loss values ​​from both the data-driven and physical constraint paths are below a preset threshold and tend to stabilize.

[0194] The predictive performance of the MT-PINN model is measured by the coefficient of determination ( To evaluate. For example... Figure 5 As shown in the figure, the horizontal axis, Actual Depth (mm), represents the measured corrosion depth, and the vertical axis, Predicted Depth (mm), represents the corrosion depth predicted by the model. Traditional artificial neural network models exhibit significant overfitting and insufficient generalization ability, with determination coefficients of 0.711 and 0.351 on the training and test sets, respectively. In contrast, the MT-PINN model proposed in this invention achieves determination coefficients of 0.945 and 0.935 on the training and test sets, respectively. Furthermore, the gray (training set) and red (test set) scatter points are densely distributed along the diagonal in the figure, indicating a high degree of consistency between the predicted and actual values.

[0195] The above results demonstrate the advantages of the model in this invention: MT-PINN, by introducing physical constraint terms, aligns the learning process with the underlying physical mechanism of the corrosion process, achieving not only extremely high prediction accuracy but also effectively suppressing overfitting. This allows the model to maintain stable and excellent generalization performance even when faced with unseen data, thus solving the problem of insufficient prediction reliability of traditional data-driven methods in complex corrosion environments.

Claims

1. A corrosion monitoring system for soil-covered storage tanks based on MT-PINN, characterized in that... Including: Information acquisition unit (1): used to acquire service environment information of the environment in which the soil-covered storage tank is located; this unit deploys various types of sensors to collect multi-source heterogeneous environmental data and build a service environment information sequence; Corrosion feature preprocessing unit (2): Receives the multi-source heterogeneous service environment information sequence output by the information acquisition unit (1), and uses attention weighted fusion and spatiotemporal alignment filtering method to preprocess the data and extract corrosion-related feature information; The corrosion rate online monitoring unit (3) based on the MT-PINN model is used to construct the MT-PINN model; it further includes: Feature building module of corrosion kinetic physical mechanism (31): Based on the Butler-Volmer equation, a corrosion kinetic physical information model describing the electrochemical process and metal corrosion prevention behavior of the soil-covered storage tank is constructed; Multi-task neural network prediction module (32): Constructs a multi-task physical information neural network and is used to receive the corrosion feature information, iteratively optimizes the MT-PINN model using a composite loss function, and outputs the corrosion rate monitoring results of the soil-covered storage tank.

2. The corrosion monitoring system for soil-covered storage tanks based on MT-PINN according to claim 1, characterized in that... Training the MT-PINN model involves the following steps: S101, the main network processes the training data through two hidden layers and then directly outputs the logarithm of the erosion depth; S102, the first branch network extracts the logarithm of the corrosion current density from the training data and uses the logarithm of the corrosion current density to correct the electrochemical term in the MT-PINN model. S103, the second branch network transforms the training data using the Sigmoid activation function and linearly maps it to the interval. The corrosion kinetic index n is obtained. S104, the third branch network extracts the logarithmic value of the environmental coupling factor from the training data.

3. The corrosion monitoring system for soil-covered storage tanks based on MT-PINN according to claim 1, characterized in that: The physical model is embedded through a composite loss function, and two corrosion depth predictions are learned simultaneously in the MT-PINN model: one route directly outputs the logarithm of the maximum corrosion depth prediction value based on data-driven methods from the main network, and assigns the logarithm of the maximum corrosion depth prediction value to the main network; the other route obtains the logarithm of the maximum corrosion depth prediction value based on physical driving from the outputs of the three branch networks through the physical model.

4. The MT-PINN-based corrosion monitoring system for soil-covered storage tanks according to claim 3, characterized in that: The composite loss function consists of four parts, among which: (1) Mean square error between predicted value and actual monitored value The erosion depth used to drive the network to fit the training samples; (2) Physical consistency error Used to force the main network output to match the calculation results of the physical equations, so that the network learns the equivalent corrosion current. Time exponent n, environmental coupling factor It conforms to the laws of corrosion kinetics; (3) Monotonicity constraint error This is used to ensure that the corrosion depth increases monotonically over time; (4) L2 regularization term Used to prevent network overfitting; Then we have the total loss function The mathematical expression is . This represents the mean square error between the predicted value and the actual monitored value. This is an adjustable first weighting coefficient. This represents the physical consistency error. This is an adjustable second weighting coefficient. This is a monotonicity constraint error. This is an L2 regularization term.

5. The corrosion monitoring system for soil-covered storage tanks based on MT-PINN according to claim 1, characterized in that: The soil-covered storage tank is divided into upper, middle and lower layers for sensor deployment; there are 8 types of sensors.

6. The corrosion monitoring system for soil-covered storage tanks based on MT-PINN according to claim 1, characterized in that... The multi-source data fusion edge preprocessing in the corrosion feature preprocessing unit (2) includes the following steps: Step 23A: Divide the dataset into three layers: upper, middle, and lower. Each layer of the dataset contains the following three types of parameters; Size of soil environmental parameter series The subscript B indicates the total number of sensors used to measure environmental parameters of the environment where the soil-covered storage tank is located. Size of the electrochemical parameter sequence The subscript C indicates the total number of all sensors used to measure the electrochemical parameters of the environment in which the soil-covered storage tank is located. Size of the structural state parameter sequence The subscript D indicates the total number of sensors used to measure the structural state parameters of the environment where the soil-covered storage tank is located. In addition, the service environment information is sampled in time sequence, if the first... The first type of sensor Each sensor node at the current sampling time The current sampling reading is represented by a single current sampling reading, which contains multi-source heterogeneous monitoring data of soil environmental parameter sequences, electrochemical parameter sequences, and structural state parameter sequences. Step 23B: Align sensor data of the same type according to the time scale; No. The first type of sensor The average value of each sensor node during the monitoring period is , This represents the total number of samples. Step 23C: Calculation and weighted aggregation of attention weights within the same type of sensor; No. The attention weight row vector of the information collected by the type of sensor is ; column vector The transpose of , where the superscript T is the transpose character. The weight matrix is ​​a learnable matrix. For biased row vectors, Temperature coefficient, used for regulation Output distribution sharpness hyperparameters; Step 23D: Intra-layer weight generation; Different weights are assigned to the information collected by different types of sensors to generate corrosion feature vectors deployed in different layers; When there is sufficient training data, learnable adaptive weights can be used. , For row vectors, Let be a trainable coefficient matrix. These are trainable bias row vectors; When there is sufficient training data, learnable adaptive weights can be used. , For row vectors, Let be a trainable coefficient matrix. These are trainable bias row vectors; Step 23E, generation of the corrosion feature matrix; No. Intra-layer weight row vector With the Feature row vectors of the layer The matrices have the same size, and by using the Hadamard product, we obtain the first... Corrosion characteristics of the layer , This is the element-wise product symbol.

7. The MT-PINN-based corrosion monitoring system for soil-covered storage tanks according to claim 1, characterized in that: The characteristic construction module of the corrosion kinetic physical mechanism (31) first selects the power law model as the basic physical model based on the engineering characteristics of the soil-covered storage tank; then, using the Butler-Volmer equation as the basic equation of electrochemical kinetics, the power law model is modified and characterized, and its modification is as follows: , The equivalent corrosion current is represented by the subscript *corr*, which indicates the corrosion process, and the subscript *eq*, which indicates the equivalent meaning. The term represents the service time, with the superscript n indicating the time index. As an environmental coupling factor, The corrosion depth is represented by the subscript max, indicating the maximum depth. Finally, the equivalent corrosion current, time exponent, and environmental coupling factor are designed as the physical intermediate parameters of MT-PINN.

8. The corrosion monitoring system for soil-covered storage tanks based on MT-PINN according to claim 1, characterized in that: The multi-task neural network prediction module (32) first feeds the total feature matrix into a main prediction branch and three physical quantity constraint auxiliary branches simultaneously; then, it uses a multi-layer fully connected neural network as the basic network unit and designs differentiated functions for the main and auxiliary branches; finally, it constructs a weighted fusion total loss function, combining data loss, physical loss, monotonic loss, and regularization term. The mathematical expression is .

9. The MT-PINN-based corrosion monitoring system for soil-covered storage tanks according to claim 1, characterized in that: The main branch prediction has two hidden layers, each with 8 neurons, and the hyperbolic tangent function tanh is used as the activation function.

10. The MT-PINN-based corrosion monitoring system for soil-covered storage tanks according to claim 1, characterized in that: Three physical auxiliary constraint branches: These auxiliary branches are the core carriers embedded in the physical mechanisms of corrosion kinetics, including the electrochemical quantity branch. Environmental compensation factor branch Time exponential parameter branch The logarithmic equivalent corrosion current is calculated respectively. Logarithmic environmental coupling factor The three major physical intermediate parameters are: time index n; Each branch adopts an independent network structure and a differentiated activation function to adapt to the value characteristics of different physical quantities. Furthermore: (1) Electrochemical quantity branch With environmental compensation factor branch All of them use the tanh activation function to adapt to the continuous variation characteristics of electrochemical corrosion parameters and environmental material coupling factors; (2) Time exponential parameter branch Using the sigmoid activation function and incorporating prior knowledge of corrosion kinetics, a linear transformation is employed to impose a strict constraint on the output, limiting the time exponent n to a specific value. Within a physically reasonable range, avoid unreasonable values ​​that exceed the physical boundaries when training the network.

Images