Method and system for predicting service life of stainless steel pipe

Abstract

The invention discloses a stainless steel pipe service life prediction method and system, and the method comprises the following steps: collecting and preprocessing sensor data of a stainless steel pipe, and generating a preprocessing data set; inputting the preprocessed data set into an improved SimCLR model, and generating feature representation; inputting the feature representation into an improved Mamba model, and carrying out time sequence modeling and residual life prediction; minimizing contrast loss for the improved SimCLR model, and minimizing regression loss and classification loss for the improved Mamba model; sensor data from the stainless steel pipe are received in real time, residual life prediction is carried out, and the model is updated and optimized regularly; health management and maintenance decision making are conducted on the stainless steel pipe according to the real-time prediction result, and a life prediction report is generated. The method can accurately predict the remaining service life of the stainless steel pipe, and improves the intelligent level of pipeline maintenance decision.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and sensor technology, and in particular to a method and system for predicting the service life of stainless steel pipes. Background Technology

[0002] With the continuous improvement of industrial automation and intelligence, health monitoring and life prediction technologies for pipeline equipment have gradually received widespread attention. Most existing stainless steel pipe health management systems rely on traditional periodic inspections and experience-based judgments, or simple sensor data collection and analysis. However, in practical applications, these methods generally suffer from the following problems:

[0003] First, traditional pipeline health monitoring systems often rely solely on single sensor data, such as pressure or temperature, neglecting the combined influence of multiple operating conditions, resulting in limited prediction accuracy. Second, most existing life prediction models employ linear methods or simple models based on statistical analysis, failing to fully capture the complex nonlinear relationships within sensor data, leading to inaccurate predictions under extreme conditions. Furthermore, current prediction methods are highly sensitive to data noise and outliers, failing to effectively address common issues in long-term monitoring such as sensor failures and data gaps, further reducing prediction reliability. Finally, significant challenges remain in real-time data processing and model updates; traditional methods struggle to implement real-time feedback mechanisms, preventing models from adjusting promptly based on the latest data, thus impacting the real-time performance and accuracy of pipeline life predictions.

[0004] Therefore, how to provide a method and system for predicting the service life of stainless steel pipes is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method and system for predicting the service life of stainless steel pipes. This invention makes full use of sensor data acquisition, improved SimCLR and improved Mamba models, and describes in detail the pipeline remaining service life prediction algorithm based on contrastive learning, time series modeling and regression analysis, which has the advantages of real-time prediction, dynamic optimization and maintenance decision support.

[0006] A method for predicting the service life of a stainless steel pipe according to an embodiment of the present invention includes the following steps:

[0007] Sensor data from stainless steel pipes is collected and preprocessed to generate a preprocessed dataset.

[0008] The preprocessed dataset is fed into the improved SimCLR model, and feature representations are generated through data augmentation, feature extraction, and contrastive learning.

[0009] The feature representations are input into the improved Mamba model to perform time series modeling and remaining lifetime prediction, generating prediction results;

[0010] Minimize the contrastive loss for the improved SimCLR model, and minimize the regression and classification losses for the improved Mamba model;

[0011] It receives sensor data from stainless steel pipes in real time, performs remaining life prediction, generates real-time prediction results, and regularly updates and optimizes the improved SimCLR model and the improved Mamba model through real-time data feedback.

[0012] Based on real-time forecast results, health management and maintenance decisions are made for stainless steel pipes, providing specific lifespan predictions or health status assessment reports.

[0013] Optionally, the process of collecting sensor data from the stainless steel pipe and preprocessing it to generate a preprocessed dataset specifically includes:

[0014] Sensor data of stainless steel pipes are collected, including temperature data, pressure data, corrosion data, vibration data, and image data of the stainless steel pipe surface, to generate a raw dataset;

[0015] The original dataset is preprocessed to generate a preprocessed dataset. The preprocessing includes removing noise from temperature data, pressure data, corrosion data, and vibration data using filters or smoothing algorithms to generate a denoised dataset. Various types of data in the denoised dataset are normalized to generate a normalized dataset. Time series interpolation is performed on the missing time series data in the normalized dataset to fill in the missing values ​​and generate an interpolated dataset. Image data is sized and color standardized to generate a standardized image dataset.

[0016] The interpolated dataset and the standardized image dataset are aligned according to timestamps or data correspondences, and then stitched or fused based on corresponding features to generate a unified preprocessed dataset.

[0017] Optionally, the step of inputting the preprocessed dataset into the improved SimCLR model to generate feature representations through data augmentation, feature extraction, and contrastive learning specifically includes:

[0018] The preprocessed dataset is input into the improved SimCLR model, which includes an adaptive multimodal data augmentation module, a feature extraction module, a nonlinear projection head, and a contrastive learning module.

[0019] The adaptive multimodal data augmentation module generates multiple perturbation samples by adding noise, time shifting, and time scale transformation to the data, thus obtaining an augmented time series dataset. It performs joint spatial-frequency domain augmentation on the image data, including performing Fourier transform on the image data, scaling, cropping, and filtering operations on the image in the frequency domain, and generating an augmented image dataset by combining geometric transformations. Finally, it merges the augmented time series dataset and the augmented image dataset to generate an augmented dataset.

[0020] The augmented dataset is input into the feature extraction module, which uses a multi-scale self-attention convolutional neural network to learn features and generate intermediate feature representations. The multi-scale self-attention convolutional neural network uses convolutional kernels of different scales in different convolutional layers and adaptively allocates attention weights in combination with a self-attention mechanism.

[0021] The intermediate feature representation is input into a nonlinear projection head, which includes a multilayer perceptron and a trainable gating mechanism. The intermediate feature representation is mapped to a low-dimensional space through a multilayer ReLU or GELU activation function to generate a low-dimensional feature representation. The trainable gating mechanism is used to dynamically adjust the activation intensity of each feature channel.

[0022] An adaptive weighted contrastive loss and a dynamic temperature adjustment mechanism are used to optimize feature representation. The adaptive weighted contrastive loss dynamically adjusts the weights of positive and negative sample pairs based on the difficulty of the samples. The difficulty of the samples is the distance or similarity between the positive and negative sample pairs in the feature space. The dynamic temperature adjustment mechanism adjusts the temperature parameter in the loss function through an adaptive algorithm to generate feature representation.

[0023] Optionally, the step of inputting the feature representation into the improved Mamba model for time series modeling and remaining lifetime prediction to generate prediction results specifically includes:

[0024] The feature representation is input into an improved Mamba model, which includes a time series modeling module and a lifetime prediction module;

[0025] The time series modeling module uses a selective state space network and a self-attention mechanism to generate time series feature representations. The selective state space network evaluates each time step in the feature representation and dynamically selects the most critical time step for modeling. The input of each time step is weighted and combined with information from past time steps to evaluate the contribution of the current time step to the remaining lifetime prediction, which is used as the importance of the current time step.

[0026] The self-attention mechanism is used to calculate and weight the relative importance of each time step in the temporal feature representation to generate a weighted temporal feature representation. The self-attention mechanism evaluates the contribution of each time step to the whole sequence by calculating the similarity between each time step and other time steps. The attention score of each time step is used to calculate the influence of the time step on the weighted temporal feature representation. The calculation process includes performing a dot product operation on each pair of time steps, normalizing the score, and using it to generate the weight of each time step.

[0027] The weighted temporal feature representation is input into the lifetime prediction module. The lifetime prediction module processes the weighted temporal feature representation through a regression network to generate the remaining lifetime prediction value as the prediction result output. The regression network adopts a multilayer perceptron structure, which passes the weighted temporal feature representation layer by layer. Each layer includes a fully connected layer and a ReLU activation function.

[0028] Optionally, minimizing the contrastive loss for the improved SimCLR model and minimizing the regression loss for the improved Mamba model specifically includes:

[0029] When training the improved SimCLR model, the contrastive loss is minimized. The contrastive loss is calculated by comparing the similarity between positive and negative sample pairs, so that the distance between positive sample pairs in the feature space is as small as possible and the distance between negative sample pairs is as large as possible.

[0030] When training the improved Mamba model, the regression loss is minimized by calculating and minimizing the difference between the predicted remaining lifetime value and the actual remaining lifetime value of the improved Mamba model.

[0031] By minimizing both regression and contrastive losses simultaneously through a joint loss function, which weights and sums the regression and contrastive losses, the improved SimCLR model optimizes the contrastive loss during feature learning, while the improved Mamba model optimizes remaining lifetime prediction by minimizing the regression loss.

[0032] Optionally, the real-time reception of sensor data from the stainless steel pipe, the prediction of remaining life, the generation of real-time prediction results, and the periodic updating and optimization of the improved SimCLR model and the improved Mamba model through real-time data feedback specifically include:

[0033] The system receives sensor data from stainless steel pipes in real time, preprocesses the data to generate a real-time dataset, and inputs it into an improved SimCLR model for feature extraction to generate feature representations. The feature representations are then input into an improved Mamba model, which uses a regression network to predict the remaining lifetime and generate real-time prediction results.

[0034] Through a real-time data feedback mechanism, the real-time prediction results are compared with the actual monitoring values ​​to generate a feedback error, which is used to evaluate the accuracy of the real-time prediction results.

[0035] In the real-time data feedback mechanism, the improved SimCLR model and the improved Mamba model are updated and optimized regularly based on the feedback error. The optimization process includes dynamically adjusting the contrast loss in the improved SimCLR model according to the feedback error, and updating the regression network and parameters in the improved Mamba model according to real-time data.

[0036] Optionally, the step of making health management and maintenance decisions for stainless steel pipes based on real-time prediction results and providing specific lifespan predictions or health status assessment reports includes:

[0037] Based on the real-time prediction results, a life prediction report is generated, which includes the remaining life of the pipeline, health status assessment and potential risks. The current prediction is compared with historical data to assess the health status of the stainless steel pipe and whether there are any potential risks.

[0038] Based on the life prediction report and health status assessment report, and according to the remaining life prediction value, it is determined whether preventive maintenance or equipment replacement is required. If the predicted remaining life is lower than the set threshold, the health management plan will prompt maintenance or replacement. If the health status assessment shows that the pipeline is in normal working condition, maintenance will be delayed.

[0039] In the real-time data feedback mechanism, the health status assessment report is dynamically adjusted by combining new real-time data and real-time prediction results. Whenever new sensor data is input into the system, the remaining lifespan and health status are reassessed, and the lifespan prediction report is updated and the health status assessment report is adjusted based on the latest assessment results.

[0040] Life expectancy prediction reports or health status assessment reports are presented in digital form.

[0041] A service life prediction system for stainless steel pipes according to an embodiment of the present invention includes the following modules:

[0042] The data acquisition and preprocessing module is used to acquire sensor data from stainless steel pipes and perform data preprocessing to generate a preprocessed dataset.

[0043] The feature extraction and generation module is used to input the preprocessed dataset into the improved SimCLR model and generate feature representations through data augmentation, feature extraction, and contrastive learning.

[0044] The remaining lifetime prediction module is used to input feature representations into the improved Mamba model for time series modeling and remaining lifetime prediction.

[0045] The loss optimization module is used to optimize the improved SimCLR model and the improved Mamba model;

[0046] The feedback optimization module is used to receive sensor data from the stainless steel pipe in real time and regularly update and optimize the improved SimCLR model and the improved Mamba model through a real-time data feedback mechanism.

[0047] The health management and maintenance decision-making module is used to generate lifespan prediction reports and health status assessment reports based on real-time prediction results.

[0048] The beneficial effects of this invention are:

[0049] This invention effectively improves the accuracy of remaining life prediction for stainless steel pipes by combining multi-source sensor data and introducing improved SimCLR and Mamba models. First, by preprocessing the sensor data to eliminate noise and fill in missing values, the data quality is improved, providing a reliable foundation for subsequent analysis. Second, the improved SimCLR model, through data augmentation, feature extraction, and contrastive learning to generate feature representations, can deeply explore the state characteristics of stainless steel pipes under various operating conditions, improving prediction accuracy. Third, combining time-series modeling and remaining life prediction with the improved Mamba model allows for dynamic assessment of pipeline health based on historical data, providing more accurate prediction results. Furthermore, through a real-time data feedback mechanism, the system can periodically optimize and update the model, ensuring that the prediction results always reflect the current pipeline condition, greatly improving the model's real-time performance and robustness. Finally, the life prediction report and health status assessment report generated based on real-time prediction results provide a scientific basis for maintenance decisions, avoiding the drawbacks of relying on experience in traditional methods, helping users to perform timely pipeline maintenance or replacement, effectively extending equipment life and reducing maintenance costs. In summary, this invention has significant advantages in improving prediction accuracy, real-time performance, and reliability. It can be widely applied to health management and maintenance decisions for stainless steel pipes, and has high engineering application value and social benefits. Attached Figure Description

[0050] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0051] Figure 1 This is a flowchart of a method for predicting the service life of stainless steel pipes proposed in this invention.

[0052] Figure 2 This is a schematic diagram of the improved SimCLR model in the stainless steel pipe service life prediction method proposed in this invention.

[0053] Figure 3 This is a framework diagram of a stainless steel pipe service life prediction system proposed in this invention. Detailed Implementation

[0054] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0055] refer to Figure 1-2 A method for predicting the service life of stainless steel pipes includes the following steps:

[0056] Sensor data from stainless steel pipes is collected and preprocessed to generate a preprocessed dataset.

[0057] The preprocessed dataset is fed into the improved SimCLR model, and feature representations are generated through data augmentation, feature extraction, and contrastive learning.

[0058] The feature representations are input into the improved Mamba model to perform time series modeling and remaining lifetime prediction, generating prediction results;

[0059] Minimize the contrastive loss for the improved SimCLR model, and minimize the regression and classification losses for the improved Mamba model;

[0060] It receives sensor data from stainless steel pipes in real time, performs remaining life prediction, generates real-time prediction results, and regularly updates and optimizes the improved SimCLR model and the improved Mamba model through real-time data feedback.

[0061] Based on real-time forecast results, health management and maintenance decisions are made for stainless steel pipes, providing specific lifespan predictions or health status assessment reports.

[0062] In this embodiment, the process of collecting sensor data from the stainless steel pipe and preprocessing it to generate a preprocessed dataset specifically includes:

[0063] Sensor data of stainless steel pipes are collected, including temperature data, pressure data, corrosion data, vibration data, and image data of the stainless steel pipe surface, to generate a raw dataset;

[0064] The original dataset is preprocessed to generate a preprocessed dataset. The preprocessing includes removing noise from temperature data, pressure data, corrosion data, and vibration data using filters or smoothing algorithms to generate a denoised dataset. The various types of data in the denoised dataset are normalized to ensure that data of different dimensions have the same weight to generate a normalized dataset. Time series interpolation is performed on the missing time series data in the normalized dataset to fill in the missing values ​​in the data to generate an interpolated dataset. The image data is sized and color standardized to generate a standardized image dataset.

[0065] The interpolated dataset and the standardized image dataset are aligned according to timestamps or data correspondences, and then stitched or fused based on corresponding features to generate a unified preprocessed dataset.

[0066] In this embodiment, the step of inputting the preprocessed dataset into the improved SimCLR model and generating feature representations through data augmentation, feature extraction, and contrastive learning specifically includes:

[0067] The preprocessed dataset is input into the improved SimCLR model, which includes an adaptive multimodal data augmentation module, a feature extraction module, a nonlinear projection head, and a contrastive learning module.

[0068] The adaptive multimodal data augmentation module generates multiple perturbation samples by adding noise, time shifting, and time scale transformation to the data, thus obtaining an augmented time series dataset. It performs joint spatial-frequency domain augmentation on the image data, including performing Fourier transform on the image data, scaling, cropping, and filtering operations on the image in the frequency domain, and generating an augmented image dataset by combining geometric transformations. Finally, it merges the augmented time series dataset and the augmented image dataset to generate an augmented dataset.

[0069] The augmented dataset is input into the feature extraction module, which uses a multi-scale self-attention convolutional neural network to learn features and generate intermediate feature representations. The multi-scale self-attention convolutional neural network can effectively capture local and global information by using convolutional kernels of different scales in different convolutional layers. Combined with the self-attention mechanism, it adaptively allocates attention weights, enabling the network to focus on key features in the data and improve the ability of the improved SimCLR model to express complex information.

[0070] The intermediate feature representation is input into a nonlinear projection head, which includes a multilayer perceptron and a trainable gating mechanism. The intermediate feature representation is mapped to a low-dimensional space through multilayer ReLU or GELU activation functions to generate a low-dimensional feature representation. The trainable gating mechanism is used to dynamically adjust the activation intensity of each feature channel, thereby selectively highlighting important features and improving the sensitivity of the improved SimCLR model to key features.

[0071] An adaptive weighted contrastive loss and a dynamic temperature adjustment mechanism are employed to optimize feature representation. The adaptive weighted contrastive loss dynamically adjusts the weights of positive and negative sample pairs based on the difficulty of the samples, where the difficulty is defined as the distance or similarity between the positive and negative sample pairs in the feature space. This results in higher weights for difficult-to-distinguish negative sample pairs, thereby enhancing the model's learning ability for difficult samples. The dynamic temperature adjustment mechanism adjusts the temperature parameter in the loss function through an adaptive algorithm, enabling the improved SimCLR model to flexibly adjust at different training stages, thereby improving optimization efficiency and effectiveness, and generating feature representations.

[0072] In this embodiment, the step of inputting the feature representation into the improved Mamba model for time series modeling and remaining lifetime prediction, and generating prediction results, specifically includes:

[0073] The feature representation is input into an improved Mamba model, which includes a time series modeling module and a lifetime prediction module;

[0074] The time series modeling module employs a selective state-space network and a self-attention mechanism to capture temporal dependencies in feature representations and generate temporal feature representations. The selective state-space network evaluates each time step in the feature representation and dynamically selects the most critical time step for modeling. The input of each time step is weighted and combined with information from past time steps to evaluate the contribution of the current time step to the remaining lifetime prediction, which serves as the importance of the current time step. This evaluation process enables the improved Mamba model to automatically select the most relevant time steps for modeling, thereby reducing computational complexity and retaining the most influential time series data.

[0075] The self-attention mechanism is used to calculate and weight the relative importance of each time step in the temporal feature representation, generating a weighted temporal feature representation. The self-attention mechanism evaluates the contribution of each time step to the entire sequence by calculating the similarity between each time step and other time steps. The attention score of each time step is used to calculate the influence of the time step on the weighted temporal feature representation. The calculation process includes performing a dot product operation on each pair of time steps, normalizing the score, and using it to generate the weight of each time step. The self-attention mechanism enables the improved Mamba model to assign higher weights to time steps that have strong dependencies on other time steps, thereby strengthening the feature representation of key time steps and ensuring that the improved Mamba model focuses on the moments most important for lifetime prediction.

[0076] The weighted time-series feature representation is input into the life prediction module. The life prediction module processes the weighted time-series feature representation through a regression network to generate a remaining life prediction value as the prediction result output. The regression network adopts a multilayer perceptron structure, which passes the weighted time-series feature representation layer by layer. Each layer includes a fully connected layer and a ReLU activation function. In this way, the regression network can capture the complex nonlinear relationship in the weighted time-series feature representation and finally output the remaining life prediction value of the stainless steel pipe, which represents the remaining time that the pipe can continue to be used.

[0077] In this embodiment, minimizing the contrastive loss of the improved SimCLR model and minimizing the regression loss of the improved Mamba model specifically includes:

[0078] When training the improved SimCLR model, the contrastive loss is minimized. The contrastive loss is calculated by comparing the similarity between positive and negative sample pairs, so that the distance between positive sample pairs in the feature space is as small as possible and the distance between negative sample pairs is as large as possible. Through this process, the improved SimCLR model can optimize the learned feature representation, thereby improving the model's ability to distinguish different samples and enhancing the expression effect of features.

[0079] When training the improved Mamba model, the regression loss is minimized. The regression loss is calculated by the difference between the predicted remaining life value and the actual remaining life value of the improved Mamba model. By minimizing the difference, the improved Mamba model can improve the accuracy of remaining life prediction, enabling it to accurately predict the remaining service life of stainless steel pipes, thus providing an effective basis for maintenance decisions.

[0080] By simultaneously minimizing the regression loss and contrastive loss through a joint loss function, the overall performance of the improved SimCLR model and the improved Mamba model is optimized. The joint loss function is a weighted sum of the regression loss and the contrastive loss, which enables the improved SimCLR model to optimize the contrastive loss when performing feature learning, while the improved Mamba model optimizes the remaining lifetime prediction by minimizing the regression loss. This joint optimization strategy can handle both feature learning and remaining lifetime prediction tasks at the same time, thereby achieving an effective balance between the two tasks and improving the overall prediction capability.

[0081] In this embodiment, the real-time reception of sensor data from the stainless steel pipe, the prediction of remaining life, the generation of real-time prediction results, and the periodic updating and optimization of the improved SimCLR model and the improved Mamba model through real-time data feedback specifically include:

[0082] The system receives sensor data from stainless steel pipes in real time, preprocesses the data to generate a real-time dataset, and inputs it into an improved SimCLR model for feature extraction to generate feature representations. The feature representations are then input into an improved Mamba model, which uses a regression network to predict the remaining lifetime and generate real-time prediction results.

[0083] Through a real-time data feedback mechanism, the real-time prediction results are compared with the actual monitoring values ​​to generate feedback errors. These feedback errors are used to evaluate the accuracy of the real-time prediction results. This mechanism ensures that the improved SimCLR model and the improved Mamba model can be adjusted in a timely manner and maintain high prediction accuracy as data continues to flow in and the device status changes.

[0084] In the real-time data feedback mechanism, the improved SimCLR model and the improved Mamba model are periodically updated and optimized based on the feedback error. The optimization process includes dynamically adjusting the contrast loss in the improved SimCLR model according to the feedback error to improve the accuracy of feature representation, and updating the regression network and parameters in the improved Mamba model according to real-time data so that it can more accurately predict the remaining lifetime. In this way, the model can self-adjust and optimize after receiving new sensor data each time to adapt to data changes.

[0085] In this embodiment, the step of making health management and maintenance decisions for stainless steel pipes based on real-time prediction results and providing specific lifespan predictions or health status assessment reports specifically includes:

[0086] Based on the real-time prediction results, a life prediction report is generated, which includes the remaining life of the pipeline, health status assessment and potential risks. The current prediction is compared with historical data to assess the health status of the stainless steel pipe and whether there are any potential risks.

[0087] Based on the life prediction report and health status assessment report, the remaining life prediction value is used to determine whether preventive maintenance or equipment replacement is required. If the predicted remaining life is lower than the set threshold, the health management plan will prompt maintenance or replacement. If the health status assessment shows that the pipeline is in normal working condition, maintenance will be delayed. This health management plan can serve as the basis for decisions on daily maintenance and emergency repairs.

[0088] In the real-time data feedback mechanism, the health status assessment report is dynamically adjusted by combining new real-time data and real-time prediction results. Whenever new sensor data is input into the system, the remaining lifespan and health status are reassessed, and the lifespan prediction report is updated and the health status assessment report is adjusted based on the latest assessment results.

[0089] Life prediction reports or health status assessment reports are presented in digital form. These reports can be updated in real time and adjusted according to changes in sensor data, ensuring that the maintenance team always makes decisions based on the latest pipeline status, thereby improving the management and maintenance efficiency of the equipment and extending the service life of stainless steel pipes.

[0090] refer to Figure 3 A service life prediction system for stainless steel pipes includes the following modules:

[0091] The data acquisition and preprocessing module is used to acquire sensor data from stainless steel pipes and perform data preprocessing to generate a preprocessed dataset.

[0092] The feature extraction and generation module is used to input the preprocessed dataset into the improved SimCLR model and generate feature representations through data augmentation, feature extraction, and contrastive learning.

[0093] The remaining lifetime prediction module is used to input feature representations into the improved Mamba model for time series modeling and remaining lifetime prediction.

[0094] The loss optimization module is used to optimize the improved SimCLR model and the improved Mamba model;

[0095] The feedback optimization module is used to receive sensor data from the stainless steel pipe in real time and regularly update and optimize the improved SimCLR model and the improved Mamba model through a real-time data feedback mechanism.

[0096] The health management and maintenance decision-making module is used to generate lifespan prediction reports and health status assessment reports based on real-time prediction results.

[0097] Example 1:

[0098] To verify the feasibility of this invention in practice, it was applied to the pipeline health management system of a petrochemical company, specifically for predicting the remaining service life and making maintenance decisions for stainless steel pipes within the company. This company uses a large number of stainless steel pipes in its production process. These pipes are subjected to long-term exposure to high temperatures, high pressures, and corrosive substances; therefore, accurately predicting their service life and health status is crucial. Traditional pipeline management methods rely on periodic inspections and manual judgment, but these methods often depend on experience and are highly subjective, making it difficult to guarantee accurate pipeline health assessments and timely maintenance.

[0099] The company's pipeline health management system is equipped with various sensors to monitor the pipeline's operating status in real time, including temperature sensors, pressure sensors, vibration sensors, corrosion sensors, and cameras to monitor corrosion on the pipeline surface. These sensors collect a large amount of real-time data daily, serving as the basis for predicting the remaining lifespan and assessing the health status of the stainless steel pipes. To verify the effectiveness of the method, several stainless steel pipelines that had been in operation for more than 5 years were specifically selected for data collection and analysis.

[0100] These data mainly include temperature data (unit: °C), pressure data (unit: MPa), vibration data (unit: m / s²), corrosion data (unit: mm), and surface image data. The system collects data from these sensors once per hour, with the entire experimental period lasting 30 days. Table 1 below shows a sample of some of the collected data:

[0101] Table 1 Sample Table of Stainless Steel Pipe Sensor Data Acquisition

[0102] date Temperature (°C) Pressure (MPa) Vibration (m / s²) Corrosion degree (mm) Surface image data (encoded) 2025 / 9 / 1 95.2 8.5 0.12 0.15 Image_001 2025 / 9 / 2 96.1 8.7 0.13 0.16 Image_002 2025 / 9 / 3 94.8 8.4 0.14 0.18 Image_003 2025 / 9 / 4 97 8.9 0.16 0.2 Image_004 2025 / 9 / 5 95.5 8.6 0.15 0.22 Image_005

[0103] To ensure data quality, preprocessing was performed after data acquisition, including denoising, standardization, and temporal interpolation. Data on temperature, pressure, vibration, and corrosion were filtered and denoised to eliminate abnormal data caused by sensor noise or external interference. For image data, size and color were standardized to ensure consistent format across images acquired at different times, facilitating subsequent processing.

[0104] Subsequently, this preprocessed data is fed into the improved SimCLR model, where feature representations are generated through data augmentation, feature extraction, and contrastive learning. Data augmentation mainly includes time shifting, time-scale transformation, and joint spatial-frequency domain augmentation after Fourier transform of the image data. Through these processes, an augmented dataset containing rich features is generated, and feature representations representing the pipeline health status are extracted through an adaptive multimodal data augmentation module and a feature extraction module.

[0105] After feature representations are generated, these features are fed into an improved Mamba model for time series modeling and remaining lifetime prediction. The improved Mamba model employs a selective state-space network and a self-attention mechanism, evaluating the importance of each time step through a weighted mechanism and combining information from past time steps to generate a weighted time series feature representation. Finally, a regression network is used to predict the pipeline's remaining lifetime.

[0106] For example, data collected on September 5, 2025, after model processing, predicted the remaining lifespan of the stainless steel pipe to be 2.3 years, with an error within 5%, while the actual predicted lifespan of the pipe was 2.4 years. This accuracy verifies the efficiency and precision of the method of this invention.

[0107] Because the working environment and condition of pipelines may change in actual production, real-time data feedback and model updates are crucial. The system of this invention can receive sensor data from stainless steel pipes in real time and periodically update and optimize the improved SimCLR and improved Mamba models based on a real-time data feedback mechanism. For example, when the pressure in a pipeline exceeds the normal operating range, the system automatically adjusts the model weights and updates the feature representation to improve the accuracy of remaining life prediction.

[0108] Based on real-time predictions, the system generates pipeline life prediction reports and health status assessment reports. The reports include the pipeline's remaining lifespan, health status assessment, potential risks, and whether preventative maintenance or equipment replacement is necessary. For example, in the report dated September 7, 2025, the system predicts a pipeline's remaining lifespan to be 1.5 years, with increasing corrosion indicating a high potential risk. Based on this result, the system recommends preventative maintenance to avoid future equipment failures or safety incidents.

[0109] Table 2. Stainless Steel Pipe Remaining Life Prediction Report and Health Status Assessment Form

[0110] date Remaining life expectancy (years) Health status assessment Potential risks Recommended maintenance measures 2025 / 9 / 7 1.5 healthy High corrosion risk Preventive maintenance 2025 / 9 / 8 1.3 healthy medium corrosion risk Delayed maintenance, monitoring 2025 / 9 / 9 1 Minor damage High corrosion risk Replace immediately

[0111] As shown in Table 2, the remaining lifespan of the pipeline gradually decreases over time, while the accuracy of the health status assessment continuously improves. Through real-time monitoring and prediction, the system can promptly identify potential risks and take corresponding maintenance measures, avoiding production downtime and safety accidents caused by pipeline aging or damage. Furthermore, the introduction of a data feedback mechanism allows the model to be dynamically optimized and adjusted in practical applications, improving the real-time performance and accuracy of the prediction results.

[0112] By applying this invention to a stainless steel pipe lifespan prediction system, the remaining lifespan of the pipeline can be accurately predicted and dynamically optimized, significantly improving the efficiency and accuracy of pipeline health management. Compared with traditional periodic inspections and manual judgment, this invention not only reduces labor costs but also improves the timeliness of pipeline maintenance and replacement, avoiding safety risks and production downtime caused by pipeline failures. Simultaneously, the real-time data feedback mechanism and optimization update process ensure that the prediction results are more closely aligned with actual operating conditions, providing enterprises with a scientific and accurate basis for production management decisions.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the service life of stainless steel pipes, characterized in that, Includes the following steps: Sensor data from stainless steel pipes is collected and preprocessed to generate a preprocessed dataset. The preprocessed dataset is fed into the improved SimCLR model, and feature representations are generated through data augmentation, feature extraction, and contrastive learning. The feature representations are input into the improved Mamba model to perform time series modeling and remaining lifetime prediction, generating prediction results; Minimize the contrastive loss for the improved SimCLR model, and minimize the regression and classification losses for the improved Mamba model; It receives sensor data from stainless steel pipes in real time, performs remaining life prediction, generates real-time prediction results, and regularly updates and optimizes the improved SimCLR model and the improved Mamba model through real-time data feedback. Based on real-time prediction results, health management and maintenance decisions are made for stainless steel pipes, providing specific lifespan predictions or health status assessment reports.

2. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The process of collecting sensor data from the stainless steel pipe and preprocessing it to generate a preprocessed dataset specifically includes: Sensor data of stainless steel pipes are collected, including temperature data, pressure data, corrosion data, vibration data, and image data of the stainless steel pipe surface, to generate a raw dataset; The original dataset is preprocessed to generate a preprocessed dataset. The preprocessing includes removing noise from temperature data, pressure data, corrosion data, and vibration data using filters or smoothing algorithms to generate a denoised dataset. Various types of data in the denoised dataset are normalized to generate a normalized dataset. Time series interpolation is performed on the missing time series data in the normalized dataset to fill in the missing values ​​and generate an interpolated dataset. Image data is sized and color standardized to generate a standardized image dataset. The interpolated dataset and the standardized image dataset are aligned according to timestamps or data correspondences, and then stitched or fused based on corresponding features to generate a unified preprocessed dataset.

3. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The process of inputting the preprocessed dataset into the improved SimCLR model and generating feature representations through data augmentation, feature extraction, and contrastive learning specifically includes: The preprocessed dataset is input into an improved SimCLR model, which includes an adaptive multimodal data augmentation module, a feature extraction module, a nonlinear projection head, and a contrastive learning module. The adaptive multimodal data augmentation module generates multiple perturbation samples by adding noise, time shifting, and time scale transformation to the data, thus obtaining an augmented time series dataset. It performs joint spatial-frequency domain augmentation on the image data, including performing Fourier transform on the image data, scaling, cropping, and filtering operations on the image in the frequency domain, and generating an augmented image dataset by combining geometric transformations. Finally, it merges the augmented time series dataset and the augmented image dataset to generate an augmented dataset. The augmented dataset is input into the feature extraction module, which uses a multi-scale self-attention convolutional neural network to learn features and generate intermediate feature representations. The multi-scale self-attention convolutional neural network uses convolutional kernels of different scales in different convolutional layers and adaptively allocates attention weights in combination with a self-attention mechanism. The intermediate feature representation is input into a nonlinear projection head, which includes a multilayer perceptron and a trainable gating mechanism. The intermediate feature representation is mapped to a low-dimensional space through a multilayer ReLU or GELU activation function to generate a low-dimensional feature representation. The trainable gating mechanism is used to dynamically adjust the activation intensity of each feature channel. An adaptive weighted contrastive loss and a dynamic temperature adjustment mechanism are used to optimize feature representation. The adaptive weighted contrastive loss dynamically adjusts the weights of positive and negative sample pairs based on the difficulty of the samples. The difficulty of the samples is the distance or similarity between the positive and negative sample pairs in the feature space. The dynamic temperature adjustment mechanism adjusts the temperature parameter in the loss function through an adaptive algorithm to generate feature representation.

4. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The step of inputting the feature representation into the improved Mamba model for time series modeling and remaining lifetime prediction, and generating prediction results, specifically includes: The feature representation is input into an improved Mamba model, which includes a time series modeling module and a lifetime prediction module; The time series modeling module uses a selective state space network and a self-attention mechanism to generate time series feature representations. The selective state space network evaluates each time step in the feature representation and dynamically selects the most critical time step for modeling. The input of each time step is weighted and combined with information from past time steps to evaluate the contribution of the current time step to the remaining lifetime prediction, which is used as the importance of the current time step. The self-attention mechanism is used to calculate and weight the relative importance of each time step in the temporal feature representation to generate a weighted temporal feature representation. The self-attention mechanism evaluates the contribution of each time step to the whole sequence by calculating the similarity between each time step and other time steps. The attention score of each time step is used to calculate the influence of the time step on the weighted temporal feature representation. The calculation process includes performing a dot product operation on each pair of time steps, normalizing the score, and using it to generate the weight of each time step. The weighted temporal feature representation is input into the lifetime prediction module. The lifetime prediction module processes the weighted temporal feature representation through a regression network to generate the remaining lifetime prediction value as the prediction result output. The regression network adopts a multilayer perceptron structure, which passes the weighted temporal feature representation layer by layer. Each layer includes a fully connected layer and a ReLU activation function.

5. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The minimization of contrastive loss for the improved SimCLR model and the minimization of regression loss for the improved Mamba model specifically include: When training the improved SimCLR model, the contrastive loss is minimized. The contrastive loss is calculated by comparing the similarity between positive and negative sample pairs, so that the distance between positive sample pairs in the feature space is as small as possible and the distance between negative sample pairs is as large as possible. When training the improved Mamba model, the regression loss is minimized by calculating and minimizing the difference between the predicted remaining lifetime value and the actual remaining lifetime value of the improved Mamba model. By minimizing both regression and contrastive losses simultaneously through a joint loss function, which weights and sums the regression and contrastive losses, the improved SimCLR model optimizes the contrastive loss during feature learning, while the improved Mamba model optimizes remaining lifetime prediction by minimizing the regression loss.

6. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The process of receiving sensor data from the stainless steel tube in real time, predicting the remaining lifespan, generating real-time prediction results, and periodically updating and optimizing the improved SimCLR model and the improved Mamba model through real-time data feedback specifically includes: The system receives sensor data from stainless steel pipes in real time, preprocesses the data to generate a real-time dataset, and inputs it into an improved SimCLR model for feature extraction to generate feature representations. The feature representations are then input into an improved Mamba model, which uses a regression network to predict the remaining lifetime and generate real-time prediction results. Through a real-time data feedback mechanism, the real-time prediction results are compared with the actual monitoring values ​​to generate a feedback error, which is used to evaluate the accuracy of the real-time prediction results. In the real-time data feedback mechanism, the improved SimCLR model and the improved Mamba model are updated and optimized regularly based on the feedback error. The optimization process includes dynamically adjusting the contrast loss in the improved SimCLR model according to the feedback error, and updating the regression network and parameters in the improved Mamba model according to real-time data.

7. The method for predicting the service life of a stainless steel pipe according to claim 1, characterized in that, The method of making health management and maintenance decisions for stainless steel pipes based on real-time prediction results, and providing specific lifespan predictions or health status assessment reports, specifically includes: Based on the real-time prediction results, a life prediction report is generated, which includes the remaining life of the pipeline, health status assessment and potential risks. The current prediction is compared with historical data to assess the health status of the stainless steel pipe and whether there are any potential risks. Based on the life prediction report and health status assessment report, and according to the remaining life prediction value, it is determined whether preventive maintenance or equipment replacement is required. If the predicted remaining life is lower than the set threshold, the health management plan will prompt maintenance or replacement. If the health status assessment shows that the pipeline is in normal working condition, maintenance will be delayed. In the real-time data feedback mechanism, the health status assessment report is dynamically adjusted by combining new real-time data and real-time prediction results. Whenever new sensor data is input into the system, the remaining lifespan and health status are reassessed, and the lifespan prediction report is updated and the health status assessment report is adjusted based on the latest assessment results. Life expectancy prediction reports or health status assessment reports are presented in digital form.

8. A service life prediction system for stainless steel pipes, comprising the service life prediction method for stainless steel pipes as described in any one of claims 1 to 7, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to acquire sensor data from stainless steel pipes and perform data preprocessing to generate a preprocessed dataset. The feature extraction and generation module is used to input the preprocessed dataset into the improved SimCLR model and generate feature representations through data augmentation, feature extraction, and contrastive learning. The remaining lifetime prediction module is used to input feature representations into the improved Mamba model for time series modeling and remaining lifetime prediction. The loss optimization module is used to optimize the improved SimCLR model and the improved Mamba model; The feedback optimization module is used to receive sensor data from the stainless steel pipe in real time and regularly update and optimize the improved SimCLR model and the improved Mamba model through a real-time data feedback mechanism. The health management and maintenance decision-making module is used to generate lifespan prediction reports and health status assessment reports based on real-time prediction results.

Images