Method and system for detecting running state of high-temperature steam pipeline based on failure mode recognition

Abstract

The invention provides a high-temperature steam pipeline running state detection method and system based on failure mode recognition. The method comprises the steps that real-time state data of a steam pipeline are acquired; performing feature extraction on the real-time state data to obtain real-time state features of the steam pipeline; inputting the real-time state characteristics into a preset failure mode recognition model, so that the failure mode recognition model recognizes a dominant failure mode of the steam pipeline according to the real-time state characteristics; extracting corresponding target features from a plurality of historical state data of the steam pipeline according to the type of the dominant failure mode; and inputting stress distribution characteristics in the real-time state characteristics and the target characteristics into a preset operation state detection model, so that the operation state detection model generates the current operation state and the residual life of the steam pipeline according to input data, and the accuracy of detecting the operation state of the high-temperature steam pipeline is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment state monitoring, and particularly relates to a data transmission control method and system for a power system. BACKGROUND

[0002] High-temperature steam pipelines are lifelines in the industrial fields of thermal power generation, petrochemical industry, nuclear power, etc., and bear the core task of transporting high-temperature and high-pressure steam. These pipelines are operated under extreme conditions for a long time and face the synergistic effect of multiple degradation mechanisms such as creep, fatigue, corrosion and oxidation. The sudden failure of the pipeline will not only cause great economic losses due to unplanned shutdown, but also may cause catastrophic accidents such as steam leakage and explosion, which seriously threaten personnel safety and the environment. In view of the importance of safe operation and the severity of failure consequences, it is necessary to realize real-time and accurate monitoring of the operating state of the high-temperature steam pipeline and scientific evaluation of the remaining life, which has become a core technical requirement for ensuring the intrinsic safety and predictive maintenance of key industrial facilities.

[0003] The current mainstream detection methods have significant limitations: on the one hand, conventional methods such as periodic manual visual inspection, single-point ultrasonic thickness measurement, and offline metallographic analysis are discrete, lagging and subjective. Manual inspection cannot capture dynamic changes in real time and is not sensitive to internal defects and small cracks; a single-point sensor can only reflect the local state and is difficult to fully grasp the temperature / stress distribution gradient of the entire pipeline, making it easy to miss early damage in critical areas. On the other hand, although some advanced factories use online monitoring systems, they are usually limited to single-physical-quantity monitoring and lack comprehensive analysis capabilities for multi-field coupling effects. The failure modes of high-temperature pipelines are complex and often coupled with each other, and single-dimensional data cannot accurately distinguish the dominant failure mechanism, resulting in a large deviation in the prediction of the remaining life. For example, relying only on temperature fluctuation data may completely ignore the risk of stress corrosion cracking (SCC), and monitoring only the wall thickness reduction cannot warn of brittle fracture.

[0004] In addition, existing evaluation methods highly depend on expert experience and static industry standard thresholds. Engineers often judge the pipeline state based on historical experience or general standards, which makes it difficult to quantify the competition between different failure modes and their dynamic evolution trends. With the improvement of unit parameters and the extension of material service life, the traditional "pass / fail" binary judgment mode based on fixed thresholds cannot meet the needs of precision and predictive maintenance. The industry urgently needs an intelligent method that can deeply integrate multi-source heterogeneous data, automatically identify the dominant failure mechanism, and accurately evaluate the state and life based on the failure mechanism, to realize the fundamental change from "after-the-fact remediation" to "pre-incident warning", and to ensure the safe, reliable and economic operation of key energy infrastructure. SUMMARY

[0005] In view of the above technical problems, the application provides a high-temperature steam pipeline operation state detection method and system based on failure mode identification, which improves the accuracy of high-temperature steam pipeline operation state detection.

[0006] In a first aspect, the application provides a high-temperature steam pipeline operation state detection method based on failure mode identification, which comprises the following steps: Obtaining real-time state data of the steam pipeline; Extracting features from the real-time state data to obtain real-time state features of the steam pipeline; Inputting the real-time state features into a preset failure mode identification model, so that the failure mode identification model identifies the dominant failure mode of the steam pipeline according to the real-time state features; According to the type of the dominant failure mode, corresponding target features are extracted from a plurality of historical state data of the steam pipeline, wherein if the dominant failure mode is creep-dominant, temperature gradient evolution features and thermal fluctuation features are extracted from a plurality of historical temperature distribution data; if the dominant failure mode is fatigue crack dominant, geometric morphology evolution features and texture evolution features are extracted from a plurality of historical image data; if the dominant failure mode is corrosion dominant, corrosion kinetics evolution features and ion migration features are extracted from a plurality of historical chemical composition data; Inputting stress distribution features in the real-time state features and the target features into a preset operation state detection model, so that the operation state detection model generates the current operation state and the remaining life of the steam pipeline according to the input data; The failure mode identification model is jointly constructed based on a convolutional neural network and a Transformer network and is trained using a plurality of historical state data, and the operation state detection model is jointly constructed based on an ensemble learning model and a time series deep learning model and is trained using a plurality of historical state data.

[0007] The embodiment of the application provides a kind of based on failure mode identification high temperature steam pipeline operating state detection method, by extracting the real-time state feature of steam pipeline and using failure mode identification model to determine the current dominant failure mode, then according to the dominant failure mode targetedly constructs target feature, finally using stress distribution feature and target feature predict the current operating state and residual life of steam pipeline, realize the deep fusion of multi-source heterogeneous data and progressive state detection.In the embodiment of the application, based on CNN-Transformer fusion model, the dominant failure mode is autonomously identified according to the input feature, which significantly improves the reliability of failure mechanism identification under complex conditions.Then innovatively dynamically correlate the dominant failure mode with historical data features, adaptively extract the core time sequence features from the historical data according to the dominant failure mode, focus on temperature gradient evolution when creep is dominant;When fatigue is dominant, analyze the crack geometry expansion trend;When corrosion is dominant, track the ion migration dynamics process.This strategy avoids redundant calculation of irrelevant historical data, improves feature extraction efficiency;At the same time, it ensures that the extracted features strictly match the current failure mechanism, provides high correlation evolution law input for the life prediction model, reduces the evaluation bias from the source, provides a mechanism basis for residual life assessment, forms a closed-loop framework of "data acquisition-failure mode identification-state detection", and essentially improves the accuracy of high temperature steam pipeline operating state detection and reduces the risk of unplanned shutdown.

[0008] Further, the real-time state data of the steam pipeline is acquired, including: The real-time state data of the steam pipeline is acquired by a plurality of sensors at each preset position; The temperature distribution data of the surface of the steam pipeline is acquired by a fiber-optic temperature sensor array; The stress distribution data of the steam pipeline is acquired by a strain gauge array; The image data of the surface of the steam pipeline is acquired by an infrared thermal imager and a visible light camera; The chemical composition data of the surface of the steam pipeline and the surrounding environment is acquired by a chemical sensor.

[0009] The embodiment of the application provides a method for acquiring real-time state data, which integrates a fiber-optic sensor array, a strain gauge array, an infrared / visible light dual-mode imaging, and a chemical sensor, establishes a spatially fully-covered in-situ monitoring network, and provides sufficient data support for subsequent pattern recognition and state detection.Temperature distribution data captures local overheating anomalies, strain data locates stress concentration areas, dual-mode images identify surface micro-crack morphology, and chemical sensors quantify corrosion environment parameters, realizing the synchronous acquisition of multi-source heterogeneous data, overcoming the visual blind area of traditional single-point detection, and improving the accuracy of high temperature steam pipeline operating state detection.

[0010] Further, the feature extraction on the real-time state data obtains real-time state features of the steam pipeline, including: calculating temperature gradient data of the steam pipeline according to the temperature distribution data, and further constructing temperature distribution features; performing strain data analysis on the stress distribution data to determine maximum stress points and stress concentration areas of the steam pipeline, and further constructing stress distribution features; combining infrared images and visible light images in the image data, using a computer vision algorithm to identify a plurality of existing surface cracks of the steam pipeline, and determining crack lengths and crack directions of each surface crack, and further constructing crack morphology features; determining real-time concentrations of corrosive ions and environmental pH values at each preset position of the steam pipeline according to the chemical composition data, calculating corrosion intensity indexes and environmental erosion degrees at each preset position of the steam pipeline according to the real-time concentrations of each corrosive ion and each environmental pH value, and further constructing chemical corrosion features according to each corrosion intensity index and each environmental erosion degree; combining the temperature distribution features, the stress distribution features, the crack morphology features, and the chemical corrosion features to obtain the real-time state features.

[0011] The feature extraction method provided in the embodiments of the present application designs special feature extraction logic for different data types, wherein the temperature gradient data is used to reveal the evolution trend of local thermal anomalies, the maximum stress points and the stress concentration areas are used to quantify the degree of mechanical load concentration, the crack lengths and the crack directions of each surface crack are used to represent the geometric state of the surface micro-cracks of the pipeline, and the real-time concentrations of corrosive ions and the environmental pH values are used to comprehensively reflect the erosion intensity at each position. The embodiments of the present application convert the original data into interpretable engineering features, eliminate the subjective interference of artificial experience, provide high-discrimination input information for failure mode recognition, and at the same time, enhance the relevance of features and actual physical mechanisms, thereby further improving the accuracy of the operation state detection of the high-temperature steam pipeline.

[0012] In a possible implementation manner, the failure mode recognition model recognizes a dominant failure mode of the steam pipeline according to the real-time state features, including: performing, by a first convolutional neural network branch, a plurality of convolution operations and MaxPooling dimension reduction operations on the temperature distribution features in the real-time state features in sequence to generate a temperature feature vector of a preset dimension; performing, by a second convolutional neural network branch, spatial position coding on the stress distribution features in the real-time state features to obtain corresponding position embedding data, and then performing full connection mapping on the embedding data to generate a stress distribution feature vector of a preset dimension; The crack morphology feature in the real-time state feature is sequentially subjected to multi-layer residual connection, global average pooling and feature normalization operation by a third convolutional neural network branch to generate a crack morphology feature vector of a preset dimension; The chemical corrosion feature in the real-time state feature is sequentially subjected to Min-Max normalization and feature encoding operation by a fourth convolutional neural network branch to generate a chemical corrosion feature vector of a preset dimension; The temperature feature vector, stress distribution feature vector, crack morphology feature vector and chemical corrosion feature vector are input into a Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on an attention fusion mechanism to obtain a context perception feature of a preset dimension, and then generates the dominant failure mode of the steam pipeline according to the context perception feature through a Softmax classifier.

[0013] The embodiments of the present application provide a method for identifying the dominant failure mode of a steam pipeline, which uses different convolutional neural network branches to process different feature data. Specifically, the temperature branch retains key information of heat distribution through data dimension reduction, the stress branch establishes spatial position relationship between stresses through position encoding, the crack branch enhances feature expression capability through residual connection, and the chemical branch guarantees data comparability through normalization. Then, the Transformer network is used to dynamically weight the contribution of each modality, solving the mechanical problem of traditional models in multi-source data fusion. This architecture significantly improves the model's ability to analyze the temperature-stress-chemical multi-field coupling effect, especially in the competitive failure scenario, enhancing the classification robustness and improving the accuracy of high-temperature steam pipeline operation state detection.

[0014] In one possible implementation, the operation state detection model generates the current operation state and the remaining life of the steam pipeline according to input data, including: The stress distribution feature and the target feature are subjected to feature fusion based on a cross-modal attention mechanism to generate a fusion feature vector of a preset dimension; According to the fusion feature vector, the gradient boosting decision tree in the operation state detection model is used to predict the probability distribution corresponding to each operation state of the steam pipeline, and then the current operation state of the steam pipeline is output through probability interval mapping; According to the fusion feature vector, the bidirectional LSTM network in the operation state detection model is used to generate a corresponding fusion prediction value, and the fusion prediction value is subjected to Bayesian prediction to generate the remaining life of the steam pipeline.

[0015] The embodiment of the application provides a method for predicting the current operation state and the residual life of a steam pipeline, which fuses stress distribution features and historical target features through a cross-modal attention mechanism, eliminates data dimension differences, generates a unified high-dimensional representation, and the high-dimensional representation implicitly contains dominant failure mode information, thereby providing useful information for subsequent prediction. Then, according to the fused feature vector, a gradient boosting decision tree (GBDT) is used to output an operation state probability distribution, multi-level risk early warning is realized, the rigidity of traditional binary judgment is broken through, and the comprehensiveness of high-temperature steam pipeline operation state detection is improved. On the other hand, a bidirectional LSTM is used to capture the time sequence dependency relationship from the fused feature vector, the uncertainty range of the residual life is quantified by combining Bayesian prediction, and the residual life of the steam pipeline is generated. Based on the above technical means, the embodiment of the application can maintain prediction stability in a complex failure scenario, realize cooperative optimization of state grading and life assessment.

[0016] Further, the method further comprises the following steps: According to the dominant failure mode, a corresponding physical constraint equation is determined, wherein the creep dominant type corresponds to an Arrhenius temperature compensation equation, the fatigue crack dominant type corresponds to a stress intensity factor compensation equation, and the corrosion dominant type corresponds to an ion concentration compensation equation; The fused feature vector is numerically corrected according to the physical constraint equation to obtain a physical constraint feature vector; The physical constraint feature vector is time series formatted based on a preset time sliding window to obtain a formatted time sequence matrix; The formatted time sequence matrix is input into the bidirectional LSTM network, so that the bidirectional LSTM network generates a corresponding fused prediction value; Based on a Monte Carlo sampling method, a plurality of prediction values are randomly generated according to the fused prediction value; According to each prediction value, Bayesian prediction is performed to generate the residual life of the steam pipeline.

[0017] In the embodiments of the present application, further physical equation constraints are introduced during the service life, wherein the Arrhenius equation is applicable to the creep dominant type for quantifying the exponential acceleration effect of high temperature on material damage; the stress intensity factor is used to capture the power law relationship between the crack propagation rate and the stress intensity factor, highlighting the instability risk; and the ion concentration equation is used to directly embed the decisive influence of acidity on the electrochemical corrosion rate into the model. The setting of the above three physical equation constraints can ensure that the model prediction is constrained by physical laws, avoid unreasonable and absurd results caused by pure data-driven, and at the same time, adaptively switch the corresponding physical equation constraints according to the dominant failure mode, thereby improving the prediction accuracy. Then, after generating a probability distribution through Monte Carlo sampling, the Bayesian framework is used to fuse the physical constraints and data prediction results, forming a mechanism and data dual-driven evaluation system. The embodiments of the present application effectively suppress the overfitting risk of pure data models, and significantly improve the reliability and generalization ability of life prediction under extreme working conditions.

[0018] In a possible implementation manner, the running state detection model is constructed based on an integrated learning model and a time series deep learning model, and is obtained by training a plurality of historical state data. A plurality of historical state data sets of the steam pipeline in each period under different dominant failure modes are collected. Each of the historical state data sets is added with a corresponding real label in an artificial labeling manner, and the real label includes a real running state and a residual life of the steam pipeline in the corresponding period. A training data set is constructed according to each of the historical state data sets and the corresponding real label. An initial running state detection model is constructed based on an integrated learning model, a time series deep learning model and a plurality of physical constraint equations, and the initial running state detection model includes a state detection branch and a life prediction branch. The initial running state detection model is trained based on a supervised learning manner and by using the training data set, and the running state detection model is obtained, wherein the state detection branch is trained by using a cross-entropy loss, and the life prediction branch is trained by using a Huber loss.

[0019] The embodiment of the application provides a model training method of an operation state detection model, an initial model is constructed based on an integrated learning model and a time series deep learning model, the integrated learning model efficiently mines the nonlinear relationship between multiple source features, the time series model accurately captures the evolution trend, and the collaborative optimization of state detection and life prediction is realized. In the training process, a physical constraint equation is introduced, the physical law of material failure is directly embedded into the model, the prediction result that violates the scientific principle generated by pure data driving is avoided, and the generalization ability under extreme working conditions is significantly improved. At the same time, through the real label artificially labeled, and combining the cross entropy loss to train the state detection branch and the Huber loss to train the life prediction branch, the model training convergence stability and prediction robustness are guaranteed, and the accuracy of subsequent operation state detection of high-temperature steam pipelines is improved.

[0020] In a second aspect, the embodiment of the application provides a high-temperature steam pipeline operation state detection system based on failure mode identification, comprising an acquisition module, a first feature extraction module, a mode identification module, a second feature extraction module, and an operation state detection module. The acquisition module is configured to acquire real-time state data of a steam pipeline. The first feature extraction module is configured to perform feature extraction on the real-time state data to obtain real-time state features of the steam pipeline. The mode identification module is configured to input the real-time state features into a preset failure mode identification model, so that the failure mode identification model identifies a dominant failure mode of the steam pipeline according to the real-time state features. The second feature extraction module is configured to extract corresponding target features from a plurality of historical state data of the steam pipeline according to the type of the dominant failure mode, wherein if the dominant failure mode is creep-dominant, temperature gradient evolution features and thermal fluctuation features are extracted from a plurality of historical temperature distribution data; if the dominant failure mode is fatigue crack dominant, geometric morphology evolution features and texture evolution features are extracted from a plurality of historical image data; and if the dominant failure mode is corrosion dominant, corrosion kinetics evolution features and ion migration features are extracted from a plurality of historical chemical composition data. The operation state detection module is configured to input stress distribution features in the real-time state features and the target features into a preset operation state detection model, so that the operation state detection model generates a current operation state and a remaining life of the steam pipeline according to the input data. The failure mode identification model is constructed based on a convolutional neural network and a Transformer network and is trained based on a plurality of historical state data, and the operation state detection model is constructed based on an integrated learning model and a time series deep learning model and is trained based on a plurality of historical state data.

[0021] In a possible implementation, the failure mode identification model identifies a dominant failure mode of the steam pipeline according to the real-time state feature, and the identification includes: The temperature distribution feature in the real-time state feature is sequentially subjected to a multiple convolution operation and a MaxPooling dimension reduction operation by the first convolutional neural network branch, to generate a temperature feature vector of a preset dimension; The stress distribution feature in the real-time state feature is subjected to spatial position coding by the second convolutional neural network branch, to obtain corresponding position embedding data, and then the embedding data is subjected to full connection mapping, to generate a stress distribution feature vector of a preset dimension; The crack morphology feature in the real-time state feature is sequentially subjected to a multiple residual connection, a global average pooling, and a feature normalization operation by the third convolutional neural network branch, to generate a crack morphology feature vector of a preset dimension; The chemical corrosion feature in the real-time state feature is sequentially subjected to Min-Max normalization and feature coding by the fourth convolutional neural network branch, to generate a chemical corrosion feature vector of a preset dimension; The temperature feature vector, the stress distribution feature vector, the crack morphology feature vector, and the chemical corrosion feature vector are input into a Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on an attention fusion mechanism, to obtain a context perception feature of a preset dimension, and then the dominant failure mode of the steam pipeline is generated by a Softmax classifier based on the context perception feature.

[0022] In a possible implementation, the running state detection model generates a current running state and a remaining life of the steam pipeline according to input data, and the generation includes: The stress distribution feature and the target feature are subjected to feature fusion based on a cross-modal attention mechanism, to generate a fusion feature vector of a preset dimension; According to the fusion feature vector, a probability distribution corresponding to each running state of the steam pipeline is predicted by a gradient boosting decision tree in the running state detection model, and then a current running state of the steam pipeline is output by probability interval mapping; According to the fusion feature vector, a fusion prediction value is generated by a bidirectional LSTM network in the running state detection model, and the fusion prediction value is subjected to Bayesian prediction, to generate a remaining life of the steam pipeline. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a running state detection method of a high-temperature steam pipeline based on failure mode identification provided by an embodiment of the present application; Figure 2 A feature processing flowchart of a failure mode identification model in a running state detection method of a high-temperature steam pipeline based on failure mode identification provided by an embodiment of the present application is shown in the figure. Figure 3 A feature fusion classification flowchart of a failure mode identification model in a running state detection method of a high-temperature steam pipeline based on failure mode identification provided by an embodiment of the present application is shown in the figure. Figure 4 A model architecture diagram of a running state detection model in a running state detection method of a high-temperature steam pipeline based on failure mode identification provided by an embodiment of the present application is shown in the figure. Figure 5 A structure diagram of a running state detection system of a high-temperature steam pipeline based on failure mode identification provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0025] It should be noted that the step numbers in the text are only for the convenience of explaining the specific embodiments, and do not serve as the function of limiting the execution sequence of the steps. In the description of the present application, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features.

[0026] Embodiment One: As shown in the figure, embodiment one provides a running state detection method of a high-temperature steam pipeline based on failure mode identification, including steps S1-S5: Figure 1 Step S1, acquiring real-time state data of a steam pipeline; Step S2, performing feature extraction on the real-time state data to obtain real-time state features of the steam pipeline; Step S3, inputting the real-time state features into a preset failure mode identification model, so that the failure mode identification model identifies the dominant failure mode of the steam pipeline according to the real-time state features; Step S4, determining the running state of the steam pipeline according to the dominant failure mode; and Step S4, extracting corresponding target features from the historical state data of the steam pipeline according to the type of the dominant failure mode, wherein if the dominant failure mode is creep-dominant, temperature gradient evolution features and thermal fluctuation features are extracted from the historical temperature distribution data; if the dominant failure mode is fatigue crack-dominant, geometric morphology evolution features and texture evolution features are extracted from the historical image data; if the dominant failure mode is corrosion-dominant, corrosion kinetics evolution features and ion migration features are extracted from the historical chemical composition data; Step S5, inputting the stress distribution feature in the real-time state feature and the target feature into a preset operation state detection model, so that the operation state detection model generates the current operation state and the remaining life of the steam pipeline according to the input data; The failure mode recognition model is constructed based on a convolutional neural network and a Transformer network and is trained using historical state data, and the operation state detection model is constructed based on an ensemble learning model and a time series deep learning model and is trained using historical state data.

[0027] The embodiment of the present application provides a kind of operation state detection method of high temperature steam pipeline based on failure mode identification, by extracting the real-time state feature of steam pipeline and using failure mode identification model to determine current dominant failure mode, then according to dominant failure mode targetedly constructs target feature, finally using stress distribution feature and target feature to predict the current operation state and the remaining life of steam pipeline, realizes the deep fusion of multi-source heterogeneous data and progressive state detection.In the embodiment of the present application, based on CNN-Transformer fusion model, dominant failure mode is autonomously identified according to input feature, which significantly improves the reliability of failure mechanism identification under complex conditions.Then innovatively dynamically correlate dominant failure mode with historical data features, and adaptively extract core time series features from historical data according to dominant failure mode.When creep is dominant, focus on temperature gradient evolution law;When fatigue is dominant, analyze crack geometric morphology expansion trend;When corrosion is dominant, track ion migration kinetics process.This strategy avoids redundant calculation of irrelevant historical data, improves feature extraction efficiency;At the same time, it ensures that the extracted features strictly match the current failure mechanism, provides high correlation evolution law input for life prediction model, reduces evaluation bias from the source, provides mechanism basis for remaining life assessment, forms a closed-loop framework of "data acquisition-failure mode identification-state detection", and essentially improves the accuracy of operation state detection of high temperature steam pipeline, reduces unplanned downtime risk.

[0028] Further, in step S1, the real-time state data of the steam pipeline is obtained, including: Real-time state data of the steam pipeline is acquired through a plurality of sensors at each preset position; Among them, temperature distribution data of the surface of the steam pipeline is acquired through an optical fiber temperature sensor array; Stress distribution data of the steam pipeline is acquired through a strain gauge array; Image data of the surface of the steam pipeline is acquired through an infrared thermal imager and a visible light camera; Chemical component data of the surface of the steam pipeline and the surrounding environment is acquired through a chemical sensor.

[0029] The embodiment of the application provides a method for acquiring real-time state data, integrates an optical fiber sensor array, a strain gauge array, an infrared / visible light dual-mode imaging and a chemical sensor, establishes an in-situ monitoring network with full spatial coverage, and provides sufficient data support for subsequent pattern recognition and state detection. Among them, the temperature distribution data captures local overheating abnormalities, the strain data locates stress concentration areas, the dual-mode image identifies the morphology of surface micro-cracks, and the chemical sensor quantifies corrosion environment parameters, so that synchronous acquisition of multi-source heterogeneous data is realized, the visual blind area of traditional single-point detection is overcome, and the accuracy of operation state detection of high-temperature steam pipelines is improved.

[0030] In a preferred embodiment, the multi-type sensor array deployed at the key position of the steam pipeline is combined as follows: 1. Temperature monitoring A distributed optical fiber temperature sensor (DTS) is adopted, is spirally wound (pitch ≤ 50 cm) along the surface of the pipeline, and real-time axial / circumferential temperature distribution is collected. The sampling frequency of the temperature sensor is ≥ 2 Hz, the temperature measurement range is 0-600°C, the accuracy is ± 0.5°C, and a three-dimensional thermal map of the temperature field is generated.

[0031] 2. Stress monitoring A resistance strain gauge array (120Ω constant current power supply, Wheatstone bridge full-bridge configuration) is pasted at the weld, elbow and support to synchronously collect axial stress, circumferential stress and shear stress.

[0032] 3. Surface imaging Infrared thermal imager: resolution 640×480, thermal sensitivity 0.03°C, frame rate 30fps, capture temperature abnormal area. Visible light camera: 20 million pixels, equipped with a ring-shaped LED fill light, crack recognition resolution better than 0.1mm.

[0033] 4. Chemical monitoring An electrochemical sensor group is installed on the outer surface of the pipe wall: Cl⁻ / SO4 2- Ion selective electrode (detection limit 0.1ppm); pH sensor (range 0-14, accuracy ± 0.2). Data acquisition interval 5 minutes, transmitted through Modbus protocol.

[0034] Further, in step S2, the real-time state data is feature extracted to obtain real-time state features of the steam pipeline, including: temperature gradient data of the steam pipeline is calculated according to the temperature distribution data, and a temperature distribution feature is constructed; strain data analysis is performed on the stress distribution data to determine a maximum stress point and a stress concentration area of the steam pipeline, and a stress distribution feature is constructed; infrared images and visible light images in the image data are combined, computer vision algorithms are used to identify a plurality of surface cracks already existing in the steam pipeline, and crack lengths and crack directions of the surface cracks are determined, and a crack morphology feature is constructed; real-time concentrations of corrosive ions and environmental pH values at each preset position of the steam pipeline are determined according to the chemical composition data, corrosion intensity indexes and environmental corrosion degrees at each preset position of the steam pipeline are calculated according to the real-time concentrations of the corrosive ions and the environmental pH values, and a chemical corrosion feature is constructed according to the corrosion intensity indexes and the environmental corrosion degrees; The real-time state features are obtained by combining the temperature distribution feature, the stress distribution feature, the crack morphology feature, and the chemical corrosion feature.

[0035] The embodiment of the application provides a feature extraction method, special feature extraction logic is designed for different data types, wherein the temperature gradient data is used to reveal the evolution trend of local thermal anomalies, the maximum stress point and the stress concentration area are used to quantify the concentration degree of mechanical load, the crack lengths and the crack directions of the surface cracks are used to represent the geometric state of the surface micro-cracks of the pipeline, and the real-time concentrations of the corrosive ions and the environmental pH values are used to comprehensively reflect the corrosion intensity of each position. The original data is converted into an interpretable engineering feature in the embodiment of the application, the subjective interference of artificial experience is eliminated, high-discrimination input information is provided for failure mode recognition, the relevance of the features and the actual physical mechanism is enhanced, and the accuracy of the operation state detection of the high-temperature steam pipeline is further improved.

[0036] In a preferred embodiment, the acquired real-time state data is feature extracted according to data types, and temperature distribution features, stress distribution features, crack morphology features and chemical corrosion features are obtained respectively. Among them, the temperature gradient matrix is generated by calculating the temperature difference of adjacent temperature sensors, and then the maximum value and standard deviation of the gradient are extracted from the temperature gradient matrix to obtain the temperature distribution features. The strain energy density at each location is calculated by the strain energy density formula and compared with the preset threshold value to locate the stress concentration area, wherein the threshold value can be set as 80% of the yield strength of the material. For crack features, on the one hand, based on visible light images, the U-Net network is used to segment the crack area, and the length and branch angle of each crack are calculated to construct the morphological skeleton of the initial crack feature. On the other hand, based on infrared images, each crack is further screened, and the crack edge temperature anomaly area is identified as an active crack, and the final crack morphology feature is constructed according to the active crack. For chemical corrosion features, the corrosion intensity index of each location is calculated by the concentrations of chloride ions and potassium ions at each location, and the environmental erosion degree of each location is calculated by the environmental pH value and ion migration rate at each location, and finally the corrosion intensity index and environmental erosion degree are weighted and fused to construct the chemical corrosion feature of each location.

[0037] In a possible implementation manner, in step S3, the failure mode identification model identifies the dominant failure mode of the steam pipeline according to the real-time state features, including: the temperature distribution features in the real-time state features are sequentially subjected to multiple convolution operations and MaxPooling dimension reduction operations by the first convolutional neural network branch to generate a temperature feature vector of a preset dimension; the stress distribution features in the real-time state features are subjected to spatial position coding by the second convolutional neural network branch to obtain corresponding position embedding data, and then the embedding data is subjected to full connection mapping to generate a stress distribution feature vector of a preset dimension; the crack morphology features in the real-time state features are sequentially subjected to multi-layer residual connection, global average pooling and feature normalization operations by the third convolutional neural network branch to generate a crack morphology feature vector of a preset dimension; the chemical corrosion features in the real-time state features are sequentially subjected to Min-Max normalization and feature coding operations by the fourth convolutional neural network branch to generate a chemical corrosion feature vector of a preset dimension; The temperature feature vector, stress distribution feature vector, crack morphology feature vector, and chemical corrosion feature vector are input into a Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on an attention fusion mechanism, to obtain context awareness features of a preset dimension, and then a Softmax classifier is used to generate the dominant failure mode of the steam pipeline according to the context awareness features.

[0038] The embodiment of the present application provides a method for identifying a dominant failure mode of a steam pipeline, which uses different convolutional neural network branches to process different feature data. The temperature branch retains key information of heat distribution through data dimension reduction, the stress branch establishes spatial position relationship between stresses through position coding, the crack branch enhances feature expression capability through residual connection, and the chemical branch guarantees data comparability through normalization. Then, the Transformer network is used to dynamically weight the contribution of each modality, so as to solve the mechanical problem of traditional models in multi-source data fusion. This architecture significantly improves the analysis capability of the model on temperature-stress-chemical multi-field coupling effect, especially enhances the classification robustness in the competitive failure scenario, and thus improves the accuracy of high-temperature steam pipeline operation state detection.

[0039] In a preferred embodiment, the failure mode identification process mainly includes an input feature parallel processing flow and a feature fusion classification flow. As shown in Figure 2 After the input data enters the failure mode identification model, the input feature parallel processing flow is first processed in parallel through different convolutional neural network branches, including a temperature branch, a stress branch, a crack branch, and a corrosion branch. In the temperature branch, 3 layers of 1D convolution are first performed on the input temperature distribution feature, and the number of convolution channels is 32, 64, and 128 respectively, the kernel size is 3, and the step size is 1. Then, MaxPooling dimension reduction operation is performed on the convolved data, and the pooling size in the dimension reduction process is set to 2, and finally a 64-dimensional temperature feature vector is obtained. In the stress branch, spatial position coding is first performed on the input stress distribution feature, position embedding is performed based on the Cartesian coordinate system, and then double mapping is performed through a fully connected layer, and finally a 128-dimensional stress distribution feature vector is output. In the crack branch, a ResNet-18 network is used to extract features from a 512-dimensional ResNet feature map, multi-layer residual connection operation is performed through the layer4 residual block of the ResNet-18, global average pooling (GAP) compression is performed on the spatial dimension, and finally feature normalization is performed on the pooled features to generate a 512-dimensional high-dimensional semantic feature, i.e., the crack morphology feature vector. In the corrosion branch, dynamic range normalization is performed on the input 30-dimensional chemical corrosion feature to obtain a 64-dimensional corrosion feature code, i.e., the chemical corrosion feature vector. The feature fusion classification flow is as shown in Figure 3The processed feature vectors are fused, and the dominant failure mode of the steam pipeline is determined according to the fused vector, including multi-head attention calculation, feature compression, Dropout operation and Softmax classification.

[0040] In a possible implementation, in step S4, if the dominant failure mode is creep-dominant, a temperature distribution field containing a time dimension is established according to a plurality of historical temperature distribution data, and then the temperature distribution field at a continuous time point is spatially differentiated, a temperature gradient evolution feature is constructed by calculating a gradient direction change rate; the temperature data of the key detection point are subjected to wavelet packet decomposition (Db6 wavelet basis) to extract 0.01-1Hz frequency band data, a thermal cycle parameter is calculated according to the extracted data, and a thermal fluctuation feature is constructed. If the dominant failure mode is fatigue crack-dominant, a geometric morphology evolution feature is constructed by calculating the length growth rate and the fractal dimension change rate of each crack; according to each historical image data, a crack tip ROI region is cropped and a gray level co-occurrence matrix (GLCM) analysis is performed, a texture change amount of each crack is quantified, and a texture evolution feature is constructed. If the dominant failure mode is corrosion-dominant, a corrosion kinetics evolution feature is constructed by calculating a pitting index change rate and a corrosion potential drift amount according to a plurality of historical chemical composition data; a spatial chemical concentration field is established according to a plurality of historical chemical composition data, a concentration gradient change rate and an ion migration flux are calculated based on the spatial chemical concentration field, and an ion migration feature is constructed.

[0041] In a possible implementation, in step S5, the running state detection model generates the current running state and the remaining life of the steam pipeline according to the input data, including: fusing the stress distribution feature and the target feature based on a cross-modal attention mechanism to generate a fused feature vector of a preset dimension; predicting, according to the fused feature vector, a probability distribution corresponding to each running state of the steam pipeline by a gradient boosting decision tree in the running state detection model, and then outputting the current running state of the steam pipeline by probability interval mapping; generating, according to the fused feature vector, a corresponding fused prediction value by a bidirectional LSTM network in the running state detection model, and generating the remaining life of the steam pipeline by Bayesian prediction on the fused prediction value.

[0042] The embodiment of the application provides a method for predicting the current operation state and the residual life of a steam pipeline, which fuses stress distribution features and historical target features through a cross-modal attention mechanism, eliminates data dimension differences, generates a unified high-dimensional representation, and the high-dimensional representation implicitly contains dominant failure mode information, providing useful information for subsequent prediction. Then, according to the fusion feature vector, a gradient boosting decision tree (GBDT) is used to output an operation state probability distribution, realizing multi-level risk early warning and breaking through the rigidity of traditional binary judgment, thereby improving the comprehensiveness of the operation state detection of the high-temperature steam pipeline. On the other hand, by using a bidirectional LSTM to capture the time sequence dependence from the fusion feature vector, the uncertainty range of the residual life is quantified by combining Bayesian prediction, and the residual life of the steam pipeline is generated. Based on the above technical means, the embodiment of the application can maintain prediction stability in complex failure scenarios and realize collaborative optimization of state classification and life assessment.

[0043] In a preferred embodiment, the model architecture of the operation state detection model is as shown in Figure 4 The feature engineering layer, the state classification module and the life prediction module. In the feature engineering layer, the stress distribution features and the target features are preliminarily spliced, and the mean, slope and variance of the spliced features are calculated through a sliding window to construct a 120-dimensional fusion feature vector. In the state classification module, the fusion feature vector is input into a pre-trained gradient boosting decision tree, which outputs a four-class probability (normal / attention / abnormal / danger), and the class with the highest probability is taken as the predicted current operation state of the steam pipeline. The number of trees of the gradient boosting decision tree is set to 200, and the maximum depth is set to 6. In the life prediction module, according to the current dominant failure mode, the corresponding physical constraint layer is selected, and the corresponding fusion prediction value is generated by combining the pre-trained bidirectional LSTM network, and then the residual life mean ± 3σ confidence interval is output in the Bayesian prediction layer. The synchronous prediction of the operation state and the residual life of the steam pipeline is completed.

[0044] Further, according to the fusion feature vector, the corresponding fusion prediction value is generated by the bidirectional LSTM network in the operation state detection model, and the fusion prediction value is subjected to Bayesian prediction to generate the residual life of the steam pipeline, comprising: According to the dominant failure mode, the corresponding physical constraint equation is determined, wherein the creep dominant type corresponds to the Arrhenius temperature compensation equation, the fatigue crack dominant type corresponds to the stress intensity factor compensation equation, and the corrosion dominant type corresponds to the ion concentration compensation equation; According to the physical constraint equation, the fusion feature vector is numerically modified to obtain a physical constraint feature vector; time series format the physical constraint feature vector based on a preset time sliding window to obtain a formatted time sequence matrix; input the formatted time sequence matrix into the bidirectional LSTM network to enable the bidirectional LSTM network to generate a corresponding fusion prediction value; randomly generate a plurality of prediction values according to the fusion prediction value based on a Monte Carlo sampling method; perform Bayesian prediction according to each prediction value to generate a remaining life of the steam pipeline.

[0045] In the embodiments of the present application, physical equation constraints are further introduced in the life process, wherein the Arrhenius equation is applicable to the creep dominant type and is used to quantify the exponential acceleration effect of high temperature on material damage; the stress intensity factor is used to capture the power law relationship between the crack propagation rate and the stress intensity factor, highlighting the instability risk; the ion concentration equation is used to directly embed the decisive influence of acidity on the electrochemical corrosion rate into the model, and the setting of the above three physical equation constraints can ensure that the model prediction is constrained by physical laws, avoid unreasonable and absurd results caused by pure data driving, and at the same time, adaptively switch the corresponding physical equation constraints according to the dominant failure mode, thereby improving the prediction accuracy. Then, after generating a probability distribution through Monte Carlo sampling, the Bayesian framework is used to fuse the physical constraint and the data prediction result, forming a mechanism and data double-driven evaluation system. The embodiments of the present application effectively suppress the overfitting risk of a pure data model and significantly improve the reliability and generalization ability of life prediction under extreme working conditions.

[0046] In a preferred embodiment, the specific formula of the Arrhenius equation is as follows:

[0047] wherein, is the fusion feature vector, Q is the creep activation energy, indicating the energy barrier that needs to be overcome for the material to occur creep deformation, R is the gas constant, T represents the current pipeline working temperature, represents the Hadamard product, is the modified physical constraint feature vector.

[0048] The specific formula of the stress intensity factor compensation equation is as follows:

[0049] wherein, ΔK is the stress intensity factor range, used to represent the stress field intensity at the crack tip; m is the Paris index, used to describe the constant of the material's fatigue crack propagation resistance.

[0050] The specific formula of the ion concentration compensation equation is as follows:

[0051] wherein, and are the hydrogen ion and potassium ion concentrations, respectively, and a, b are preset weight values.

[0052] In one possible implementation manner, the running state detection model is obtained by jointly constructing and training a plurality of historical state data based on an ensemble learning model and a time series deep learning model, including: a plurality of historical state data sets of the steam pipeline in different dominant failure modes at different periods are collected; each of the historical state data sets is added with a corresponding real label in an artificial labeling manner, and the real label includes a real running state and a residual life of the steam pipeline at the corresponding period; a training data set is constructed according to each of the historical state data sets and the corresponding real label; an initial running state detection model is jointly constructed based on an ensemble learning model, a time series deep learning model and a plurality of physical constraint equations, and the initial running state detection model includes a state detection branch and a life prediction branch; the initial running state detection model is jointly trained by using the training data set in a supervised learning manner, to obtain the running state detection model, wherein the state detection branch is trained by using a cross-entropy loss, and the life prediction branch is trained by using a Huber loss.

[0053] The model training method of the running state detection model provided in the embodiments of the present application jointly constructs an initial model based on an ensemble learning model and a time series deep learning model, the ensemble learning model efficiently mines the nonlinear relationship between multiple source features, and the time series model accurately captures the evolution trend, so as to realize the collaborative optimization of state detection and life prediction. In the training process, a physical constraint equation is introduced, the physical law of material failure is directly embedded into the model, the prediction result that violates the scientific principle generated by pure data driving is avoided, and the generalization ability under extreme working conditions is significantly improved. At the same time, the real label is artificially labeled, the state detection branch is trained by using a cross-entropy loss, and the life prediction branch is trained by using a Huber loss, so as to guarantee the convergence stability and prediction robustness of the model training, and improve the accuracy of subsequent running state detection of the high-temperature steam pipeline.

[0054] Embodiment Two: As shown in Figure 5 Embodiment Two provides a running state detection system of a high-temperature steam pipeline based on failure mode recognition, including an acquisition module 10, a first feature extraction module 20, a mode recognition module 30, a second feature extraction module 40 and a running state detection module 50. The acquisition module 10 is configured to acquire real-time state data of the steam pipeline. The first feature extraction module 20 is configured to extract features from the real-time state data to obtain real-time state features of the steam pipeline. The mode recognition module 30 is configured to input the real-time state features into a preset failure mode recognition model, so that the failure mode recognition model recognizes a dominant failure mode of the steam pipeline according to the real-time state features. The second feature extraction module 40 is configured to extract corresponding target features from a plurality of historical state data of the steam pipeline according to a type of the dominant failure mode, wherein if the dominant failure mode is creep-dominant, temperature gradient evolution features and thermal fluctuation features are extracted from a plurality of historical temperature distribution data; if the dominant failure mode is fatigue crack-dominant, geometric morphology evolution features and texture evolution features are extracted from a plurality of historical image data; and if the dominant failure mode is corrosion-dominant, corrosion kinetics evolution features and ion migration features are extracted from a plurality of historical chemical composition data. The operating state detection module 50 is configured to input stress distribution features in the real-time state features and the target features into a preset operating state detection model, so that the operating state detection model generates a current operating state and a remaining life of the steam pipeline according to the input data. The failure mode recognition model is constructed based on a convolutional neural network and a Transformer network and is trained using a plurality of historical state data, and the operating state detection model is constructed based on an ensemble learning model and a time series deep learning model and is trained using a plurality of historical state data.

[0055] In a possible implementation, the failure mode recognition model recognizes the dominant failure mode of the steam pipeline according to the real-time state features, including: The temperature distribution features in the real-time state features are sequentially subjected to a plurality of convolution operations and MaxPooling dimension reduction operations by a first convolutional neural network branch to generate a temperature feature vector of a preset dimension; The stress distribution features in the real-time state features are subjected to spatial position coding by a second convolutional neural network branch to obtain corresponding position embedding data, and then the embedding data is subjected to full connection mapping to generate a stress distribution feature vector of a preset dimension; The crack morphology features in the real-time state features are sequentially subjected to a plurality of residual connections, global average pooling, and feature normalization operations by a third convolutional neural network branch to generate a crack morphology feature vector of a preset dimension; The chemical corrosion features in the real-time state features are sequentially subjected to Min-Max normalization and feature coding operations by a fourth convolutional neural network branch to generate a chemical corrosion feature vector of a preset dimension; inputting the temperature feature vector, the stress distribution feature vector, the crack morphology feature vector and the chemical corrosion feature vector into a Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on an attention fusion mechanism, to obtain a context perception feature of a preset dimension, and then generating a dominant failure mode of the steam pipeline according to the context perception feature through a Softmax classifier.

[0056] In a possible implementation manner, the running state detection model generates the current running state and the remaining life of the steam pipeline according to input data, including: performing feature fusion on the stress distribution feature and the target feature based on a cross-modal attention mechanism, to generate a fusion feature vector of a preset dimension; predicting a probability distribution corresponding to each running state of the steam pipeline according to the fusion feature vector through a gradient boosting decision tree in the running state detection model, and then outputting the current running state of the steam pipeline through probability interval mapping; generating a corresponding fusion prediction value through a bidirectional LSTM network in the running state detection model according to the fusion feature vector, and performing Bayesian prediction on the fusion prediction value to generate the remaining life of the steam pipeline.

[0057] The embodiment of the present application provides a running state detection system for a high-temperature steam pipeline based on failure mode recognition. The system extracts real-time state features of the steam pipeline, determines the current dominant failure mode by using a failure mode recognition model, then constructs target features according to the dominant failure mode, and finally predicts the current running state and the remaining life of the steam pipeline by using stress distribution features and target features, thereby realizing deep fusion of multi-source heterogeneous data and progressive state detection. In the embodiment of the present application, a CNN-Transformer fusion model is used to autonomously identify the dominant failure mode based on input features, thereby significantly improving the reliability of failure mechanism identification under complex working conditions. Then, the dominant failure mode is dynamically associated with historical data features, and the core time sequence features in the historical data are adaptively extracted according to the dominant failure mode. When creep is dominant, the temperature gradient evolution law is focused on; when fatigue is dominant, the crack geometric morphology expansion trend is analyzed; and when corrosion is dominant, the ion migration dynamics process is tracked. This strategy avoids redundant calculation of irrelevant historical data, improves feature extraction efficiency, ensures that the extracted features strictly match the current failure mechanism, provides high-correlation evolution law input for the life prediction model, reduces evaluation bias from the source, provides a mechanism basis for remaining life assessment, and forms a closed-loop framework of "data acquisition-failure mode recognition-state detection", thereby essentially improving the accuracy of running state detection of the high-temperature steam pipeline and reducing the risk of unplanned shutdown.

[0058] The working principle and step flow of the embodiment can be but not limited to referring to the related description of embodiment one.

[0059] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition, characterized in that: include: Get real-time status data of steam pipelines; Performing feature extraction on the real-time status data to obtain real-time status features of the steam pipeline; Inputting the real-time status feature into a preset failure mode recognition model, so that the failure mode recognition model recognizes the dominant failure mode of the steam pipeline according to the real-time status feature; Extracting corresponding target features from a number of historical state data of the steam pipeline according to the type of the dominant failure mode, wherein, if the dominant failure mode is creep-dominated, extracting temperature gradient evolution features and thermal fluctuation features from a number of historical temperature distribution data; if the dominant failure mode is fatigue crack-dominated, extracting geometric morphology evolution features and texture evolution features from a number of historical image data; if the dominant failure mode is corrosion-dominated, extracting corrosion kinetics evolution features and ion migration features from a number of historical chemical composition data; Inputting the stress distribution feature in the real-time state feature and the target feature into a preset operation state detection model, so that the operation state detection model generates the current operation state and the remaining life of the steam pipeline according to the input data; Among them, the failure mode recognition model is jointly constructed based on a convolutional neural network and a Transformer network and is obtained by training using a number of historical state data, and the operation status detection model is jointly constructed based on an integrated learning model and a time series deep learning model and is obtained by training using a number of historical state data.

2. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 1, characterized in that: The obtaining of real-time status data of the steam pipeline includes: Acquiring real-time status data of the steam pipeline through a plurality of sensors at various preset locations; wherein, the temperature distribution data of the steam pipeline surface is obtained by an optical fiber temperature sensor array; collecting stress distribution data of the steam pipeline through a strain gauge array; Acquiring image data of the steam pipeline surface by using an infrared thermal imager and a visible light camera; The chemical composition data of the steam pipeline surface and the surrounding environment are obtained through chemical sensors.

3. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 2, characterized in that: The extracting features of the real-time status data to obtain the real-time status features of the steam pipeline includes: Calculating temperature gradient data of the steam pipeline according to the temperature distribution data, and then constructing temperature distribution characteristics; Performing strain data analysis on the stress distribution data to determine the maximum stress point and stress concentration area of ​​the steam pipeline, and then constructing stress distribution characteristics; Combining the infrared image and the visible light image in the image data, using a computer vision algorithm to identify a number of existing surface cracks on the steam pipeline, and determining the crack length and crack direction of each surface crack, thereby constructing crack morphological characteristics; Determining the real-time concentration of corrosive ions and the environmental pH value at each preset location of the steam pipeline based on the chemical composition data, calculating the corrosion intensity index and the environmental erosion degree at each preset location of the steam pipeline based on the real-time concentration of each corrosive ion and each environmental pH value, and then constructing a chemical corrosion signature based on each corrosion intensity index and each environmental erosion degree; The real-time state characteristics are obtained by combining the temperature distribution characteristics, stress distribution characteristics, crack morphology characteristics and chemical corrosion characteristics.

4. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 1, characterized in that: The failure mode identification model identifies the dominant failure mode of the steam pipeline according to the real-time state characteristics, including: Performing multiple convolution operations and MaxPooling dimensionality reduction operations on the temperature distribution features in the real-time state features in sequence through the first convolutional neural network branch to generate a temperature feature vector of a preset dimension; The stress distribution features in the real-time state features are spatially encoded by a second convolutional neural network branch to obtain corresponding position embedded data, and then the embedded data is fully connected and mapped to generate a stress distribution feature vector of a preset dimension; Performing multi-layer residual connection, global average pooling, and feature normalization operations on the crack morphology features in the real-time state features in sequence through a third convolutional neural network branch to generate a crack morphology feature vector of a preset dimension; performing Min-Max normalization and feature encoding operations on the chemical corrosion features in the real-time state features in sequence through a fourth convolutional neural network branch to generate a chemical corrosion feature vector of a preset dimension; The temperature feature vector, stress distribution feature vector, crack morphology feature vector and chemical corrosion feature vector are input into the Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on the attention fusion mechanism, obtains context-aware features of preset dimensions, and then generates the dominant failure mode of the steam pipeline through the Softmax classifier based on the context-aware features.

5. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 1, characterized in that: The operation status detection model generates the current operation status and the remaining life of the steam pipeline according to the input data, including: Performing feature fusion on the stress distribution feature and the target feature based on a cross-modal attention mechanism to generate a fused feature vector of a preset dimension; According to the fused feature vector, the probability distribution corresponding to each operating state of the steam pipeline is predicted by the gradient boosting decision tree in the operating state detection model, and the current operating state of the steam pipeline is output through probability interval mapping; According to the fused feature vector, a corresponding fused prediction value is generated through the bidirectional LSTM network in the operating status detection model, and the fused prediction value is subjected to Bayesian prediction to generate the remaining life of the steam pipeline.

6. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 5, characterized in that: The method of generating a corresponding fusion prediction value based on the fusion feature vector by using a bidirectional LSTM network in the operation status detection model, and performing Bayesian prediction on the fusion prediction value to generate the remaining life of the steam pipeline includes: Determine the corresponding physical constraint equation according to the dominant failure mode, wherein the creep-dominated mode corresponds to the Arrhenius temperature compensation equation, the fatigue crack-dominated mode corresponds to the stress intensity factor compensation equation, and the corrosion-dominated mode corresponds to the ion concentration compensation equation; Numerically correcting the fused eigenvector according to the physical constraint equation to obtain a physical constraint eigenvector; Performing time series formatting on the physical constraint feature vector based on a preset time sliding window to obtain a formatted time series matrix; Inputting the formatted time series matrix into the bidirectional LSTM network so that the bidirectional LSTM network generates a corresponding fusion prediction value; Randomly generate a number of prediction values ​​according to the fused prediction value based on the Monte Carlo sampling method; Bayesian prediction is performed based on each of the predicted values ​​to generate the remaining life of the steam pipeline.

7. The method for detecting the operating status of a high-temperature steam pipeline based on failure mode recognition according to claim 1, characterized in that: The operation status detection model is obtained by jointly building an integrated learning model and a time series deep learning model and training with a number of historical status data, including: Collect several historical status data sets of steam pipelines with different dominant failure modes at different periods; Adding corresponding real labels to each of the historical status data sets by manual labeling, wherein the real labels include the real operating status and remaining life of the steam pipeline in the corresponding period; Constructing a training dataset based on each of the historical state datasets and the corresponding true labels; An initial operating state detection model is jointly constructed based on an integrated learning model, a time series deep learning model, and several physical constraint equations. The initial operating state detection model includes a state detection branch and a life prediction branch. Based on a supervised learning approach, the initial operating status detection model is jointly trained using the training data set to obtain the operating status detection model, wherein the status detection branch is trained using cross entropy loss, and the life prediction branch is trained using Huber loss.

8. A high-temperature steam pipeline operation status detection system based on failure mode recognition, characterized in that: It includes an acquisition module, a first feature extraction module, a pattern recognition module, a second feature extraction module and an operation status detection module; Wherein, the acquisition module is used to obtain real-time status data of the steam pipeline; The first feature extraction module is used to extract features from the real-time status data to obtain real-time status features of the steam pipeline; The pattern recognition module is used to input the real-time state characteristics into a preset failure mode recognition model, so that the failure mode recognition model recognizes the dominant failure mode of the steam pipeline according to the real-time state characteristics; The second feature extraction module is used to extract corresponding target features from a number of historical state data of the steam pipeline according to the type of the dominant failure mode, wherein, if the dominant failure mode is creep-dominated, temperature gradient evolution features and thermal fluctuation features are extracted from a number of historical temperature distribution data; if the dominant failure mode is fatigue crack-dominated, geometric morphology evolution features and texture evolution features are extracted from a number of historical image data; if the dominant failure mode is corrosion-dominated, corrosion kinetics evolution features and ion migration features are extracted from a number of historical chemical composition data; The operation status detection module is used to input the stress distribution feature and the target feature in the real-time state feature into a preset operation status detection model, so that the operation status detection model generates the current operation status and the remaining life of the steam pipeline according to the input data; Among them, the failure mode recognition model is jointly constructed based on a convolutional neural network and a Transformer network and is obtained by training using a number of historical state data, and the operation status detection model is jointly constructed based on an integrated learning model and a time series deep learning model and is obtained by training using a number of historical state data.

9. The high-temperature steam pipeline operation status detection system based on failure mode recognition according to claim 8, characterized in that: The failure mode identification model identifies the dominant failure mode of the steam pipeline according to the real-time state characteristics, including: Performing multiple convolution operations and MaxPooling dimensionality reduction operations on the temperature distribution features in the real-time state features in sequence through the first convolutional neural network branch to generate a temperature feature vector of a preset dimension; The stress distribution features in the real-time state features are spatially encoded by a second convolutional neural network branch to obtain corresponding position embedded data, and then the embedded data is fully connected and mapped to generate a stress distribution feature vector of a preset dimension; Performing multi-layer residual connection, global average pooling, and feature normalization operations on the crack morphology features in the real-time state features in sequence through a third convolutional neural network branch to generate a crack morphology feature vector of a preset dimension; performing Min-Max normalization and feature encoding operations on the chemical corrosion features in the real-time state features in sequence through a fourth convolutional neural network branch to generate a chemical corrosion feature vector of a preset dimension; The temperature feature vector, stress distribution feature vector, crack morphology feature vector and chemical corrosion feature vector are input into the Transformer network, so that the Transformer network performs multi-head attention calculation on each input data based on the attention fusion mechanism, obtains context-aware features of preset dimensions, and then generates the dominant failure mode of the steam pipeline through the Softmax classifier based on the context-aware features.

10. The high-temperature steam pipeline operation status detection system based on failure mode recognition according to claim 8, characterized in that: The operation status detection model generates the current operation status and the remaining life of the steam pipeline according to the input data, including: Performing feature fusion on the stress distribution feature and the target feature based on a cross-modal attention mechanism to generate a fused feature vector of a preset dimension; According to the fused feature vector, the probability distribution corresponding to each operating state of the steam pipeline is predicted by the gradient boosting decision tree in the operating state detection model, and the current operating state of the steam pipeline is output through probability interval mapping; According to the fused feature vector, a corresponding fused prediction value is generated through the bidirectional LSTM network in the operating status detection model, and the fused prediction value is subjected to Bayesian prediction to generate the remaining life of the steam pipeline.

Images