Extreme low temperature medium transportation pipeline state diagnosis and prediction method

Abstract

A state diagnosis and prediction method for an extreme low-temperature medium transportation pipeline relates to the technical field of pipeline monitoring, and comprises the following steps: acquiring temperature shock optical fiber data to obtain temperature field and structural dynamic information of the pipeline; a voiceprint sensor array arranged on the outer wall of the pipeline is used for collecting a voiceprint signal of pipeline vibration, and pipeline deformation data are obtained through calculation; external environment parameters of the pipeline are synchronously collected; obtaining a multi-source fusion feature set according to the temperature shock optical fiber data, the voiceprint signal and the external environment parameters; training a pre-established multi-modal deep learning diagnosis model by using the multi-source fusion feature set; according to the response of the trained multi-mode deep learning diagnosis model to the current multi-source fusion features, a pipeline fault diagnosis result is obtained, and the technical problems that under the extreme low temperature, the internal state of a pipeline is difficult to perceive, the micro-displacement monitoring precision is low, the diagnosis dimension is single, and prediction is inaccurate are solved; and the safety and the intelligent level of pipeline operation and maintenance are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of pipeline monitoring technology, and more specifically, to a method for diagnosing and predicting the condition of pipelines transporting extreme cryogenic media. Background Technology

[0002] Long-distance pipeline transportation of cryogenic media such as liquefied natural gas, liquid nitrogen, and cryogenic chemical raw materials is a critical infrastructure in the energy and chemical industry. These pipelines operate under ultra-low temperature conditions, and their structural integrity and sealing reliability directly affect transportation safety and efficiency. Real-time monitoring, accurate diagnosis, and fault prediction of pipeline conditions are of paramount importance for preventing major safety accidents, environmental pollution, and economic losses caused by leaks and ruptures.

[0003] Currently, the industry mainly relies on traditional sensor technology and data analysis methods for monitoring pipeline conditions. However, when dealing with the special scenario of pipelines transporting extreme low-temperature media, existing technologies have the following significant drawbacks: Traditional pipeline temperature detection equipment, such as thermocouples and infrared thermometers, cannot directly contact extreme low-temperature media below -40°C. They can only detect the temperature of the outer wall of the pipeline and cannot reflect the temperature distribution of the internal medium, resulting in lagging and large deviations in temperature data, which cannot support the judgment of internal conditions. Traditional displacement detection equipment, such as strain gauges and laser rangefinders, are prone to failure in extreme low-temperature, high-humidity, and high-salt-alkali environments, and their positioning accuracy is mostly above the centimeter level, which cannot capture the millimeter-level micro-displacements caused by changes in medium temperature and environmental loads, making it difficult to assess the stability of the pipeline structure. Existing systems rely on a single parameter to judge the pipeline condition without combining the characteristics of the transported medium and environmental factors, resulting in one-sided diagnostic results and a high rate of false positives and false negatives. Traditional prediction methods cannot effectively extract the correlation features between historical and current pipeline data. Especially in scenarios with large fluctuations in extreme environmental data, the prediction accuracy is low, the timeliness is poor, and it is impossible to warn of potential risks in advance.

[0004] Existing technologies are limited by the shortcomings of sensor hardware performance, data processing methods and analysis models, and have obvious shortcomings in direct sensing of the internal state of pipelines transporting extreme cryogenic media, accurate monitoring of millimeter-level displacement, multi-dimensional state fusion diagnosis and high-precision forward-looking prediction. Summary of the Invention

[0005] In response to the problems in related technologies, this invention proposes a method for diagnosing and predicting the condition of pipelines transporting extreme cryogenic media, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] A method for diagnosing and predicting the condition of pipelines transporting cryogenic media includes the following steps:

[0008] Collect temperature and vibration data from distributed temperature and vibration fiber optic sensors deployed along the pipeline axis, and obtain the temperature field and structural dynamics information of the pipeline based on the temperature and vibration data.

[0009] The acoustic signature signal of the pipe vibration is collected by an array of acoustic signature sensors deployed on the outer wall of the pipe, and the pipe deformation data is calculated. The pipe deformation data includes the pipe displacement and the position after displacement.

[0010] Simultaneously collect external environmental parameters of the pipeline, including ambient temperature, humidity, wind speed, soil salinity and precipitation;

[0011] Based on the temperature field, structural dynamics information, acoustic signature signal, and external environmental parameters, a multi-source fusion feature set is obtained.

[0012] A pre-built multimodal deep learning diagnostic model is trained using a multi-source fusion feature set;

[0013] Based on the response of the trained multimodal deep learning diagnostic model to the current multi-source fusion features, the diagnostic results of pipeline faults are obtained.

[0014] Based on the current diagnostic results and historical fault monitoring data from the past N months, a time-series sample is constructed.

[0015] The time series samples are input into a pre-established prediction model that combines bidirectional LSTM with an attention mechanism;

[0016] Based on the output of the prediction model, the prediction results of pipeline faults are obtained, and the diagnosis results include the probability of fault occurrence, the type of anomaly, the maximum degree of anomaly, and the earliest possible time.

[0017] Risk warning information is generated based on the prediction results and abnormal locations.

[0018] The aforementioned technical solution, by integrating temperature and seismic fiber optic data, acoustic signature displacement data, and conventional environmental data, constructs a multi-source fusion feature set, enabling comprehensive and multi-dimensional monitoring of pipeline temperature field, structural dynamics, displacement deformation, and external environmental factors. Through multi-source data fusion, deep learning modeling, and dynamic prediction updates, a complete perception-diagnosis-prediction-decision closed-loop system is constructed, overcoming the limitations of traditional single-parameter diagnosis. This significantly improves the accuracy and comprehensiveness of condition diagnosis, effectively reducing the risk of misjudgment and missed judgment. The prediction model captures the forward and reverse temporal dependencies of historical data through bidirectional LSTM and dynamically focuses on key features using an attention mechanism, effectively overcoming the shortcomings of traditional prediction methods in scenarios with large data fluctuations in extreme environments. This method forms a complete closed loop of monitoring-diagnosis-prediction-early warning-update. The system dynamically updates the prediction results hourly based on the latest data, ensuring that the early warning information always reflects the latest status of the pipeline.

[0019] Furthermore, methods for obtaining temperature field and structural dynamics information of pipelines based on thermo-vibration fiber optic data include:

[0020] The collected temperature and vibration fiber optic data are preprocessed to remove noise and outliers;

[0021] Fourier transform or wavelet transform are used to analyze the vibration frequency components and determine the dynamic information of the structure.

[0022] The temperature-vibration fiber optic data is processed by sliding window averaging and first-order difference to extract the temperature change rate and temperature gradient distribution characteristics, read the thermal expansion coefficient of the pipe material, and calculate the temperature change based on the structural dynamic information.

[0023] The temperature field of the pipeline is obtained by tracking continuous temperature changes and setting an initial temperature value.

[0024] Furthermore, methods for obtaining temperature field and structural dynamics information of pipelines based on thermo-vibration fiber optic data also include:

[0025] The pipeline is divided into sections, and the spatial temperature gradient and the rate of change of temperature difference in the time domain of the sections are calculated based on the temperature field of the pipeline.

[0026] Based on the spatial temperature gradient and the temporal rate of temperature difference change, regions of non-uniform temperature change are obtained, and regions of abnormal stress are obtained based on regions of non-uniform temperature change.

[0027] Furthermore, methods for calculating pipe deformation data by using an array of acoustic sensors deployed on the outer wall of the pipe to collect the acoustic signals of pipe vibration include:

[0028] The pipe displacement is calculated using triangulation based on data from multiple acoustic sensor readings.

[0029] The location of the pipe after displacement was determined by analyzing the time delay of the acoustic signature signal.

[0030] Furthermore, the method for obtaining a multi-source fusion feature set based on the temperature-vibration fiber optic data, acoustic signature signal, and external environmental parameters includes:

[0031] Extract temperature trend features and deep vibration frequency features from thermo-vibration fiber optic data;

[0032] Extracting time-frequency domain features from voiceprint signals and extracting deep time-frequency features from convolutional neural networks;

[0033] Extracting multi-parameter correlation features from environmental data;

[0034] By integrating temperature trend features, deep vibration frequency features, time-frequency domain features, deep time-frequency features, and multi-parameter correlation features through an attention mechanism, and then associating them with fault types, a multi-source fusion feature set is formed.

[0035] An electronic device, including a processor and a memory;

[0036] The processor is connected to the memory;

[0037] The memory is used to store executable program code;

[0038] The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, in order to execute a method for diagnosing and predicting the condition of a pipeline transporting cryogenic media.

[0039] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for diagnosing and predicting the condition of a pipeline transporting cryogenic media.

[0040] A computer program product includes a computer program that, when executed by a processor, implements a method for diagnosing and predicting the condition of a pipeline transporting cryogenic media.

[0041] The beneficial effects of this invention are as follows:

[0042] 1. In practical use, this invention improves the accuracy and timeliness of pipeline status prediction, reduces the risk of misjudgment and omission, and adopts a hybrid model combining Bi-LSTM and Transformer. It captures the positive and negative dependencies of historical data through bidirectional temporal feature extraction, and strengthens the correlation between the current state and historical features by combining it with a cross-modal attention fusion module. This greatly improves the accuracy of anomaly occurrence prediction and anomaly type identification compared with traditional time-series prediction methods. At the same time, the accuracy of anomaly degree prediction error and anomaly occurrence time prediction error is also greatly improved. It accurately quantifies the degree of risk and time nodes, and avoids maintenance delays caused by ambiguous predictions.

[0043] 2. In practical use, this invention optimizes the allocation of operation and maintenance resources, reduces pipeline maintenance costs, and uses a precise matching mechanism based on risk level and maintenance strategy. For high-risk anomalies, emergency inspections only need to be arranged for specific mileage markers within 2 hours before the anomaly occurs. For medium-risk anomalies, a plan needs to be formulated within 24 hours. For low-risk or no anomalies, the regular cycle is maintained. Compared with the traditional inspection mode, this reduces the amount of ineffective inspections and lowers the cost of manpower and equipment. By predicting anomalies in advance, intervention measures can be taken before the failure occurs to avoid serious failures such as pipeline leakage and structural damage, thereby reducing the cost of fault repair and the economic losses and environmental risks caused by media leakage.

[0044] 3. In practical use, this invention enhances pipeline risk prevention and control capabilities under extreme environments, thereby improving operational safety. The prediction model is trained based on multimodal historical data under extreme low temperatures, high salinity, and alkalinity conditions, effectively capturing the special changing patterns of pipeline states in extreme environments. Compared to models not optimized for extreme environments, the prediction accuracy in extreme scenarios is improved, avoiding misjudgments of risks due to poor environmental adaptability. This forms a risk control system of hierarchical management and key prevention and control, significantly improving the safety and stability of pipeline operation under extreme environments.

[0045] 4. In practical applications, this invention provides intelligent support for pipeline operation and maintenance by deeply mining the value of multimodal data. By integrating the temporal characteristics of acoustic, temperature and vibration optical fiber, and routine environmental data through a multimodal information fusion layer, it explores the potential correlations between different modal data, forming multimodal fusion that enhances the generalization ability of the prediction model and fully releases the application value of multi-source monitoring data. At the same time, the prediction model can be iteratively optimized by continuously accumulating historical data, and the prediction accuracy and adaptability of decision recommendations are continuously improved in long-term use. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a method for diagnosing and predicting the condition of a pipeline transporting extreme cryogenic media, according to an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown in the embodiment of the present invention, a method for diagnosing and predicting the condition of a pipeline for transporting extreme cryogenic media is provided, comprising the following steps:

[0050] S1. Static Parameter Acquisition

[0051] S2. Collaborative data acquisition from multiple sensor modules

[0052] Temperature and vibration data are collected using special integrated temperature and vibration optical fibers that are resistant to ultra-high and low temperatures (-196℃~550℃) and have a salt spray resistance level of 9. These fibers are embedded in the inner wall insulation layer of the LNG pipeline to be monitored. Monitoring points are set up along the pipeline axis at intervals of 0.5 meters. Using DTS / DAS fusion distributed optical fiber sensing technology, the temperature and vibration signals of the medium inside the pipeline are collected in real time at a frequency of 1 time / minute. The data is transmitted to the central database via optical cable.

[0053] Millimeter-level acoustic signature displacement data is collected. An array of acoustic signature sensors is deployed every 2 meters along the outer wall of the pipeline. Each array contains three acoustic signature sensors with a low-temperature resistant titanium alloy shell and a built-in heating and defrosting module. These sensors collect the acoustic signature signals of micro-vibrations generated by thermal expansion and contraction of the pipeline and foundation settlement. After amplification and processing at the acquisition terminal, the displacement and location information of the pipeline are calculated using a triangulation algorithm. The displacement accuracy reaches ±0.1 mm, and the data is stored in the database at a frequency of once per minute.

[0054] Routine environmental data is collected by deploying industrial-grade low-temperature resistant temperature and humidity sensors, ultrasonic anemometers, and soil salinity detectors along the pipeline route to collect parameters such as ambient temperature, humidity, wind speed, precipitation, and soil salinity. All environmental data is transmitted to the data center via a Modbus-RTU industrial bus at a frequency of 1 time per minute, and is stored in association with the data from the aforementioned two types of modules at the same frequency. Finally, all sensor data is centrally archived in the database.

[0055] S3. Targeted cleaning and spatiotemporal alignment of multi-source sensor data

[0056] After cleaning the temperature and vibration fiber optic data and removing common outliers, further data with temperature values ​​exceeding the range of -196℃ to 550℃ were removed. Additionally, data with obviously unreasonable values ​​were removed within the normal operating temperature range of LNG (-80℃ to -20℃). For vibration signals, disconnected data caused by poor local contact during the spiral winding of the fiber optic cable were removed. Furthermore, based on the spatial distribution logic of monitoring points every 0.5m, a linear interpolation algorithm using synchronous data from adjacent monitoring points was employed to supplement any missing data caused by temporary signal interruptions, ensuring spatial data continuity.

[0057] The acoustic signature displacement data was cleaned. Based on the removal of common outliers, the transient interference signals (duration ≤2s) generated at the moment the sensor heating defrost module was activated under extreme low temperature were checked in the micro-vibration acoustic signature signal. The high-frequency noise introduced by the sensor heating defrost activation and the high salinity environment was filtered out using wavelet denoising algorithm. Outliers with displacement values ​​exceeding the ±0.1mm accuracy range were removed. The calculation logic was based on the spatial layout of 1 group of 2m and 3 acoustic signature probes per group. The matching of displacement position with sensor placement position was verified. Data with positioning deviation greater than 0.5m were deleted to ensure the accuracy of displacement positioning.

[0058] After removing common outliers from the routine environmental data, values ​​that jump due to equipment failure are also removed. Based on a sampling frequency of 1 time per minute, a moving average algorithm is used to supplement missing data caused by industrial bus (Modbus-RTU) transmission delay. At the same time, synchronous data from temperature and vibration fiber optic and acoustic positioning modules are correlated to remove data that contradicts the actual scenario, ensuring consistency between the environmental data and the pipeline operation scenario.

[0059] To improve data processing efficiency and consistency, after each of the three types of sensor data undergoes its own specialized cleaning process, the following processing steps must be uniformly performed to ensure consistency in data format, dimensions, and analysis standards:

[0060] Data standardization was implemented. For the three types of data with different physical dimensions, the Min-Max standardization method was used to convert all parameters, such as temperature (°C, range -196~550), displacement (mm, range ±0.1), humidity (%, range 0~1000), and salinity (mg / kg, according to the actual monitoring range), into standardized data in the range [0,1]. This eliminated the impact of dimensional differences on subsequent multi-dimensional analysis and ensured the fairness of the weights of different parameters in the model calculation.

[0061] Spatiotemporal alignment is achieved based on a unified spatiotemporal reference, completing the association and matching of three types of data. The specific coordinate rules and synchronization logic are as follows:

[0062] A three-dimensional spatial coding system is adopted, consisting of pipeline mileage markers, circumferential angles, and module identifiers. The pipeline mileage markers are based on the pipeline's starting point and are accurate to 0.1m; for example, K100+050.3 represents a distance of 100050.3m from the starting point. The circumferential angles are based on the center of the top of the pipeline as 0° and increase clockwise to 1°; for example, 90° corresponds to a horizontal position on the right side of the pipeline, adapting to the deployment characteristics of temperature-vibration fiber optic spiral winding and acoustic sensor arrays. Module identifiers are distinguished by letters; for example, F represents temperature-vibration fiber optic, A represents acoustic positioning, and E represents normal environmental conditions, ensuring accurate correspondence between data from different modules at the same spatial location. The code K25+123.4-90°-F represents a temperature-vibration fiber optic monitoring point at a horizontal position (90°) on the right side of the pipeline, 25123.4 meters from the starting point.

[0063] Using a system-level high-precision clock with an error of ≤1ms as a benchmark, millisecond-level timestamps are uniformly added to the three types of data, in the format YYYY-MM-DD HH:MM:SS.sss. The temperature-vibration fiber optic and acoustic fingerprint positioning modules generate timestamps at a sampling frequency of 1 time / second, while the environmental routine module generates timestamps at a sampling frequency of 1 time / minute. For environmental data, an interpolation completion + time anchoring method is used to associate the environmental parameter values ​​within 1 minute with the timestamps of the temperature-vibration fiber optic and acoustic fingerprint positioning data of the same period. For example, the environmental data at 10:00:00.000 is synchronously matched with the temperature-vibration and acoustic fingerprint data from 10:00:00.000 to 10:00:59.999 to avoid matching deviations caused by time differences.

[0064] Finally, using spatial coordinate encoding and timestamps as indexes, a three-dimensional data matrix of space-time-parameters is constructed to achieve complete alignment of all sensor data in the spatiotemporal dimensions. For example, index K100+050.3-90°-10:00:00.000 corresponds to the spatial location, the temperature / vibration data of the optical fiber, the acoustic signature displacement data, and the environmental temperature, humidity, and salinity data at that time point, achieving complete alignment of the three types of data in the spatiotemporal dimensions and providing a unified data foundation for subsequent pipeline condition diagnosis.

[0065] S4. Multimodal Feature Extraction and Fusion

[0066] Based on the micro-vibration acoustic signature signal from the acoustic signature localization displacement detection module after prior spatiotemporal alignment, common data processing, and specialized cleaning, the following steps are performed:

[0067] Based on the previous standardization results of the [0,1] interval, the continuous signal is divided into short-time signal frames of 1s / segment according to the millisecond-level timestamp. Each signal segment corresponds to the spatial coordinates of "pipeline mileage station number + circumferential angle" to establish the "spatiotemporal-signal frame" correspondence. At the same time, the signal frame is converted into a 2D feature map, with the horizontal axis representing the time series and the vertical axis representing the signal amplitude, to adapt to the deep learning input format.

[0068] For each short-time signal frame, four time-domain features—short-time average amplitude, short-time root mean square value, peak factor, and zero-crossing rate—are calculated, along with the dominant frequency feature extracted via FFT. Subsequently, a simplified version of the LeNet-5 convolutional neural network is introduced. Using the generated 2D signal feature map as input, and through two convolutional layers and one pooling layer, deep time-frequency correlation features of the signal are automatically extracted. These traditional features are then concatenated with the deep features from the CNN to form an acoustic fusion feature pool, enhancing the feature's ability to represent displacement signals.

[0069] Deep learning feature selection employs a multilayer perceptron (MLP), with an input layer consisting of a fusion feature pool dimension, a single hidden layer (64 neurons), and an output layer for binary classification (effective features / redundant features). It is trained on a pre-annotated feature-displacement association dataset and automatically selects features strongly correlated with pipeline displacement, such as deep time-frequency features extracted by CNN, dominant frequency, and short-time root mean square value, while eliminating redundant features, thus replacing traditional empirical selection methods.

[0070] Feature fusion and output: The filtered features are standardized in the [0,1] interval. Then, through the attention mechanism and the lightweight self-attention module, the weight of each feature is calculated to optimize the feature contribution. Finally, the optimized acoustic features are associated with the synchronous temperature and humidity features of the optical fiber and the ambient temperature and humidity features using the timestamp and pipeline mileage station as indexes. The results are then input into the deep learning regression model, and the spatiotemporal optimized feature dataset is output to support subsequent displacement positioning and status diagnosis.

[0071] Temperature and vibration fiber optic data preprocessing: Based on the cleaned temperature (accuracy ±0.1℃) and vibration signal 1s / s segment, the data is segmented synchronously with the acoustic signal frame, and each segment corresponds to the spatial coordinates of "pipeline mileage station number + circumferential angle"; the vibration signal is converted into a 2D time-frequency graph by using short-time Fourier transform (STFT, window length 50ms), with time on the horizontal axis, frequency on the vertical axis, and color representing energy; the temperature data is constructed into a 1D feature sequence according to the time series, and adapted to different deep learning input formats.

[0072] Environmental routine data preprocessing: For the cleaned environmental parameters, such as temperature, humidity, wind speed, and soil salinity, the data is completed at 1 second intervals. Based on the original acquisition frequency of once per minute, linear interpolation is used to complete the data to 1 second intervals, synchronized with other modules. Data segments with more than three consecutive missing values ​​due to temporary sensor offline are removed. The completed multi-dimensional environmental parameters are normalized, using the [0,1] interval standard, to construct a time-parameter two-dimensional data matrix.

[0073] Thermo-optical fiber feature extraction:

[0074] Traditional feature calculation: For each data segment, calculate temperature features (mean, variance, and maximum temperature difference within the segment) and vibration features (peak vibration amplitude, mean energy, and frequency bandwidth within the segment) to form a set of 6 basic features;

[0075] Deep learning feature enhancement: A lightweight CNN is introduced, taking the 2D time-frequency map of the vibration signal as input, and automatically extracting the deep frequency change features of the vibration signal, such as the special frequency patterns corresponding to the deformation of the pipeline structure, through 3 layers of deep separable convolution (3×3 kernel, stride 1) and 2 layers of max pooling (2×2 pooling kernel). At the same time, 1D convolution (kernel size 5, stride 1) is used to process the 1D temperature sequence to extract temperature trend features, such as the time node features of sudden drop / rise. The two types of deep features output by the CNN are concatenated with the traditional 6 basic features to form a temperature-vibration-fiber fusion feature pool.

[0076] Environmental routine feature extraction:

[0077] Traditional feature calculation: For each data segment, calculate the mean and rate of change of each environmental parameter within the segment, such as the change in humidity per minute and the maximum wind speed, to form a set of 8 basic features;

[0078] Deep learning feature enhancement: A lightweight MLP is adopted, with 8 basic feature dimensions in the input layer and 2 hidden layers with 32 neurons each. The activation function is ReLU. Feature mapping is performed on multi-dimensional environmental parameters, and the correlation features between parameters are automatically mined, such as the synergistic effect of low temperature + high humidity. The correlation features output by MLP are concatenated with the traditional 8 basic features to form a conventional fusion feature pool of environment.

[0079] Feature selection was performed using a lightweight classifier based on SVM and CNN, inputting it into a fusion feature pool of temperature- and vibration-induced fiber optic and environmental features, retaining features strongly correlated with pipeline state. The selected temperature- and vibration-induced fiber optic and environmental features were standardized within the [0,1] interval, and a lightweight cross-modal attention module was introduced to calculate the weights of different features. Finally, using timestamps and pipeline mileage markers as indices, all optimized temperature- and vibration-induced fiber optic features, environmental features, and acoustic optimization features were associated to form a spatiotemporal-multi-source fusion feature dataset.

[0080] S5. Pipeline condition diagnosis based on multimodal data

[0081] 1. Multimodal diagnostic data preprocessing and input adaptation:

[0082] The spatiotemporal-multi-source fusion feature dataset is grouped by pipeline mileage marker + timestamp, and each group contains three types of modal features:

[0083] Acoustic modal features: encompassing basic time-frequency domain features and deep time-frequency features extracted by CNN, with a dimension of 24;

[0084] Thermo-vibration fiber modal characteristics: including temperature trend characteristics, deep vibration frequency abrupt change characteristics and traditional statistical characteristics, with a dimension of 32;

[0085] Environmental constant-scale characteristics: including environmental parameter correlation characteristics and traditional mean / rate of change characteristics, with 16 dimensions.

[0086] Standard deviation standardization was performed on the three types of modal features to eliminate the magnitude differences between different modal features. The samples were then reorganized according to the single spatiotemporal unit-multimodal feature matrix format to form the input sample set of the multimodal model. At the same time, the corresponding pipeline status labels of the samples were labeled, such as normal operation, low temperature stress anomaly, displacement deformation anomaly, and environmental corrosion anomaly.

[0087] 2. Construction of a multimodal feature fusion model:

[0088] The fused feature dataset is input into the constructed multimodal fusion model, which employs a three-level architecture of intramodal feature enhancement, cross-modal feature interaction, and global feature fusion.

[0089] Intramodal feature enhancement: The acoustic mode uses a 1D CNN with 2 convolutional layers, a kernel size of 3, and a stride of 1 to enhance the local correlation of acoustic features, highlighting the differences in acoustic features corresponding to displacement anomalies; the temperature-vibration fiber optic mode uses a 2D CNN with 3 convolutional layers, a kernel size of 3×3, and a stride of 1 to process the vibration time-frequency map derived features and enhance the frequency change features caused by low-temperature stress; the environmental constant mode uses an MLP with 2 hidden layers and 32 and 16 neurons respectively to mine the nonlinear correlation between environmental parameters and highlight the temperature-salinity synergy features corresponding to corrosion risk.

[0090] A modal attention interaction module is introduced to calculate the association weights between each modal feature and the pipeline state label, and to achieve information interaction between different modal features through a cross-attention mechanism, eliminating information redundancy between modalities. The interacted modal features are then input into a multimodal fusion layer, which uses a combination of concatenation fusion and element-weighted fusion to generate a 64-dimensional global fusion feature. At the same time, a BatchNorm layer is used to normalize the feature distribution, improving the model's generalization ability.

[0091] 3. Implementation of Pipeline Status Classification and Anomaly Diagnosis

[0092] Based on global fusion features, a classification-regression dual-task multimodal diagnostic model is constructed. A Softmax classifier is used to output the probabilities of four states: normal operation, low-temperature stress anomaly, displacement deformation anomaly, and environmental corrosion anomaly. If the probability is ≥0.85, it is determined to be the state; otherwise, it is marked as an anomaly state to be confirmed, triggering an anomaly early warning mechanism.

[0093] A 3-layer MLP regressor is used to output anomaly quantification values. With global fusion features as input, the anomaly quantification values ​​are predicted. When the quantification value is ≥0.6, it is judged as a high-risk anomaly and needs to be dealt with first.

[0094] The model uses a weighted sum of cross-entropy loss and MSE loss as the loss function, and is trained using the Adam optimizer. The final test set classification accuracy is ≥92%, and the anomaly prediction error is ≤0.05.

[0095] 4. Diagnostic Result Output and Spatiotemporal Localization

[0096] The model diagnostic results are correlated with the spatiotemporal information of pipeline mileage station number + timestamp to generate a three-dimensional diagnostic report of spatiotemporal-state-anomaly degree:

[0097] For normal state samples, output pipeline mileage station number + time + basic information of normal operation; for abnormal state samples, in addition to outputting basic spatiotemporal and state information, add the abnormality degree quantification value, risk level (high / medium / low) and associated modal characteristics (such as abnormal displacement deformation, the abnormal indicators of acoustic displacement characteristics and temperature and vibration fiber optic vibration characteristics need to be marked).

[0098] Based on the diagnostic report, a spatiotemporal distribution map of the pipeline status is constructed, and abnormal areas are visualized in the form of a heat map (high-risk abnormal areas are marked in red, and medium-risk areas are marked in yellow), providing accurate spatiotemporal location data for subsequent pipeline maintenance.

[0099] S6. Predictive analysis based on current status and historical data

[0100] Using the three-dimensional diagnostic results of spatiotemporal-state-abnormality as the current state data, and combining the historical multimodal monitoring data of the pipeline stored in the database for the past 12 months, such as acoustic, temperature and vibration optical fiber, environmental conventional characteristics and corresponding state labels, a time-series deep learning algorithm is introduced to construct a predictive model to realize predictive analysis of the future operating status and abnormal risks of the pipeline.

[0101] 1. Predictive data preprocessing and time series dataset construction

[0102] Historical data screening and cleaning: Extract nearly 12 months of pipeline multimodal historical data from the database, group them by pipeline mileage station, remove invalid samples (≤3%) caused by sensor failure in each group, and retain valid samples containing complete acoustic, temperature and vibration fiber optic, environmental routine characteristics and corresponding status labels (normal / abnormal type / abnormal degree); perform standard deviation standardization on the feature values ​​in the historical data in the same way as the current data to ensure uniform data distribution.

[0103] Predictive input samples are constructed using time series windows as units, and time series samples are constructed using 24-hour windows with a sliding step of 1 hour. Each sample contains the multimodal feature sequence of the previous 24 hours and the output current state features. The current state features are the output pipeline state label at the current moment and the quantified value of the degree of abnormality, and are marked with whether the state label of the next 12 hours is abnormal, the type of abnormality, and the maximum degree of abnormality.

[0104] 2. Construction of Temporal Deep Learning Prediction Model

[0105] A three-level architecture of temporal feature extraction, multimodal information fusion, and prediction output is adopted to construct a hybrid prediction model based on LSTM and Transformer, which adapts to the temporal correlation of pipeline states and the complexity of multimodal features.

[0106] A bidirectional LSTM (Bi-LSTM) network is introduced as the core for temporal feature extraction. The multimodal temporal sequence of historical window features is input into the Bi-LSTM layer (128 hidden units and 2 layers). The forward and backward LSTM units capture the forward and backward temporal dependencies of historical data, such as the influence of temperature change trends on subsequent stress anomalies and the correlation between temporal abrupt changes in displacement signals and subsequent deformation. The output is a historical temporal global feature with a dimension of 256.

[0107] The historical time-series features output by Bi-LSTM and the current state features (encoded as 32-dimensional vectors) are input into the cross-modal attention fusion module to calculate the association weights between historical and current features. For example, if slight low-temperature stress has occurred, the weight of historical temperature time-series features is increased. A fusion feature with a dimension of 256 is generated by weighted summation. At the same time, a BatchNorm layer and a Dropout layer (dropout probability of 0.2) are introduced to reduce overfitting and improve the model's generalization ability.

[0108] The output layer adopts a multi-task design:

[0109] Anomaly prediction branch: The Softmax classifier predicts the anomaly type for the next 12 hours, and maps the fused features to the probability distribution of four results for the next 12 hours: no anomaly, low temperature stress anomaly, displacement deformation anomaly, and environmental corrosion anomaly. The category with the highest probability is determined as the anomaly prediction result.

[0110] Anomaly severity prediction branch: A linear regression layer is used to map the fused features to the maximum anomaly severity quantification value (range 0-1) within the next 12 hours, and at the same time outputs the predicted value of the earliest time node when the anomaly occurs, accurate to the hour.

[0111] 3. Model training and prediction performance optimization

[0112] A multi-task joint loss function is adopted. The anomaly prediction branch uses cross-entropy loss to calculate the probability difference between the predicted category and the true category, while the anomaly severity prediction branch uses MSE loss to calculate the error between the predicted quantified value and the true value. The weight ratio of the two is set to 1:1.5. The total loss function is the weighted sum of the two types of losses, ensuring that the model optimizes both classification accuracy and regression precision at the same time.

[0113] The AdamW optimizer was used with an initial learning rate of 0.001, which decayed by 0.8 every 5 epochs for training, with 50 training epochs. An early stopping strategy (patience=5) was employed: training was stopped and the optimal model parameters were saved when the total loss on the validation set did not decrease for 5 consecutive epochs. Hyperparameters such as the number of hidden layer units and attention weight coefficients of the Bi-LSTM were tuned using a grid search method to ensure optimal model performance. On the test set, multi-dimensional metrics were used to evaluate model performance, requiring anomaly occurrence prediction accuracy ≥90%, anomaly type identification accuracy ≥88%, anomaly severity prediction error ≤0.06, and anomaly occurrence time prediction error ≤1 hour, meeting the requirements of pipeline maintenance decision-making for prediction accuracy and timeliness.

[0114] 4. Prediction result output and maintenance decision support

[0115] Input the current pipeline time series sample to be predicted into the trained prediction model, and output a pipeline status prediction report for the next 12 hours. The report includes the probability of anomaly occurrence in the next 12 hours, the most likely anomaly type, the maximum anomaly severity quantification value, the earliest time node of anomaly occurrence and the corresponding risk level. Among them, high risk: anomaly severity ≥ 0.6, medium risk: 0.3 ≤ anomaly severity < 0.6, low risk: anomaly severity < 0.3.

[0116] For high-risk predictions, it is recommended to arrange an emergency inspection within 2 hours before the anomaly occurs, focusing on the pipeline section at the corresponding mileage marker; for medium-risk predictions, it is recommended to develop an inspection plan within 24 hours and increase the real-time monitoring frequency of the corresponding area; for low-risk or no-anomaly predictions, maintain the regular maintenance cycle.

[0117] Every hour, based on the latest monitoring data and current status diagnosis results, the time series sample is updated and the prediction is re-executed. The prediction results and maintenance recommendations for the next 12 hours are dynamically adjusted to ensure the timeliness and accuracy of the prediction, providing a dynamic decision-making basis for advance pipeline maintenance.

[0118] This embodiment successfully achieved accurate perception, diagnosis, and advanced prediction of the status of LNG pipelines in extreme cryogenic conditions through the above-described systematic process. After actual deployment and testing, compared with traditional methods, the amount of ineffective inspections and the potential failure rate avoided by early warning were reduced, significantly improving the operational safety and maintenance economy of the pipeline.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing and predicting the condition of pipelines transporting extreme cryogenic media, characterized in that, Includes the following steps: Collect temperature and vibration data from distributed temperature and vibration fiber optic sensors deployed along the pipeline axis, and obtain the temperature field and structural dynamic information of the pipeline based on the temperature and vibration fiber optic data. The acoustic signature signal of the pipe vibration is collected by an array of acoustic signature sensors deployed on the outer wall of the pipe, and the pipe deformation data is calculated. The pipe deformation data includes the pipe displacement and the position after displacement. Simultaneously collect external environmental parameters of the pipeline; Based on the temperature field and structural dynamics information, acoustic signature signal, and external environmental parameters, a multi-source fusion feature set is obtained; A pre-built multimodal deep learning diagnostic model is trained using a multi-source fusion feature set; Based on the response of the trained multimodal deep learning diagnostic model to the current multi-source fusion features, the diagnostic results of pipeline faults are obtained. Based on the current diagnostic results and historical fault monitoring data from the past N months, a time-series sample is constructed. The time series samples are input into a pre-established prediction model; Based on the output of the prediction model, the prediction results of pipeline faults are obtained; Risk warning information is generated based on the diagnostic results and abnormal locations.

2. The method for diagnosing and predicting the condition of a pipeline transporting extreme cryogenic media according to claim 1, characterized in that, Methods for obtaining temperature field and structural dynamics information of pipelines based on thermo-seismic fiber optic data include: The collected temperature and seismic fiber optic data are preprocessed to remove noise and outliers; Fourier transform or wavelet transform are used to analyze the vibration frequency components to determine the structural dynamic information; The temperature-vibration fiber optic data is processed by sliding window averaging and first-order difference to extract the temperature change rate and temperature gradient distribution characteristics, read the thermal expansion coefficient of the pipe material, and calculate the temperature change based on the structural dynamic information. The temperature field of the pipeline is obtained by tracking continuous temperature changes and setting an initial temperature value.

3. The method for diagnosing and predicting the condition of a pipeline transporting extreme cryogenic media according to claim 2, characterized in that, Other methods for obtaining temperature field and structural dynamics information of pipelines based on thermo-seismic fiber optic data include: The pipeline is divided into sections, and the spatial temperature gradient and the rate of change of temperature difference in the time domain of the sections are calculated based on the temperature field of the pipeline. Based on the spatial temperature gradient and the temporal rate of temperature difference change, regions of non-uniform temperature change are obtained, and regions of abnormal stress are obtained based on regions of non-uniform temperature change.

4. The method for diagnosing and predicting the condition of a pipeline transporting extreme cryogenic media according to claim 1, characterized in that, Methods for calculating pipe deformation data by collecting acoustic signature signals of pipe vibration using an array of acoustic signature sensors deployed on the outer wall of the pipe include: The pipe displacement is calculated using triangulation based on data from multiple acoustic sensor readings. The location of the pipe after displacement was determined by analyzing the time delay of the acoustic signature signal.

5. The method for diagnosing and predicting the condition of a pipeline transporting extreme cryogenic media according to claim 1, characterized in that, The method for obtaining a multi-source fusion feature set based on the temperature-vibration fiber optic data, acoustic signature signal, and external environmental parameters includes: Extract temperature trend features and deep vibration frequency features from thermo-vibration fiber optic data; Extracting time-frequency domain features from voiceprint signals and extracting deep time-frequency features from convolutional neural networks; Extracting multi-parameter correlation features from environmental data; By integrating temperature trend features, deep vibration frequency features, time-frequency domain features, deep time-frequency features, and multi-parameter correlation features through an attention mechanism, and then associating them with fault types, a multi-source fusion feature set is formed.

6. An electronic device, characterized in that, Including the processor and memory; The processor is connected to the memory; The memory is used to store executable program code; The processor runs a program corresponding to the executable program code by reading the executable program code stored in the memory, in order to perform the method as described in any one of claims 1-5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-5.

Images