A gcn-tcn fusion network ship pipeline leakage fault diagnosis system and method

Abstract

The application belongs to the technical field of industrial equipment and discloses a GCN-TCN fusion network ship pipeline leakage fault diagnosis system and method, which comprises a data acquisition module, a data interpolation module, a mixed graph construction module, a GCN-TCN diagnosis module and a fault classification output module.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of industrial equipment technology, and particularly relates to a GCN-TCN fusion network ship pipeline leakage fault diagnosis system and method. Background Technology

[0002] Marine piping systems are widely used in fuel transportation, cooling circulation, hydraulic transmission, ballast water transportation, and other fluid media transportation processes, serving as a crucial component for ensuring the stable operation of the ship's power system and auxiliary equipment. Due to the long-term exposure of ships to vibration, swaying, temperature and humidity fluctuations, and corrosive environments, pipes, valves, joints, and welded areas are prone to leakage failures due to fatigue, corrosion, and decreased sealing performance. Pipeline leaks not only cause media loss but can also lead to abnormal equipment operation and, in severe cases, compromise ship safety. Therefore, timely and accurate diagnosis of marine pipeline leaks is of paramount importance.

[0003] Currently, pipeline leak monitoring on ships primarily involves deploying sensors at critical locations along the pipeline to continuously collect state parameters during operation and determining whether a leak has occurred based on changes in the monitored data. With the development of machine learning and deep learning technologies, some existing techniques are beginning to utilize models such as convolutional neural networks and recurrent neural networks to analyze the collected time-domain signals, frequency-domain signals, or time-frequency features. By learning the characteristic differences between normal and different leak states, pipeline fault classification can be achieved. Some solutions also employ multiple sensors to jointly collect data, thereby improving the identifiability of leak characteristics under complex operating conditions.

[0004] However, the actual operating environment of ships is quite complex. When sensors are affected by factors such as mechanical vibration, electromagnetic interference, communication anomalies, and short-term equipment failures, continuously collected data is prone to missing, abnormal, or asynchronous phenomena. When processing data by deleting missing samples, filling in the mean, or interpolation, it is difficult to fully maintain the probability distribution and temporal variation patterns of the original monitoring data, which may cause the loss of transient characteristics of leaks, thereby affecting the subsequent fault diagnosis results.

[0005] Meanwhile, ship pipeline systems exhibit a distinct network structure. Different sensor nodes not only have spatial adjacency relationships determined by pipeline connection locations, but may also develop cross-node correlations due to the propagation of pressure waves, flow disturbances, and acoustic signals within the pipeline. Existing methods typically process sensor data as relatively independent time series or determine node relationships solely based on a fixed pipeline topology, making it difficult to simultaneously reflect both the local spatial connections between sensors and the nonlinear global correlations formed during operation. Furthermore, leakage faults evolve over time and propagate along the pipeline space. Simply performing spatial feature analysis or temporal feature extraction is insufficient to fully characterize the spatiotemporal coupling features corresponding to the leakage state. Further improvements are needed in identifying similar fault states such as minor and severe leaks, as well as in determining the leakage area. Summary of the Invention

[0006] To address the problems existing in current technologies, this invention provides a GCN-TCN fusion network-based system and method for diagnosing leaks in ship pipelines. It utilizes a hybrid architecture that integrates graph convolutional networks and temporal convolutional networks, combined with advanced data preprocessing and graph structure construction techniques, to perform real-time, high-precision online monitoring and diagnosis of leaks in ship pipeline systems. This invention spans multiple interdisciplinary fields, including marine engineering, predictive maintenance, artificial intelligence, and sensor data analysis.

[0007] This invention is implemented as follows: a GCN-TCN fused network ship pipeline leakage fault diagnosis system, comprising:

[0008] Data acquisition module: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, multi-source heterogeneous time series data are collected in real time, including pressure, acoustic emission, flow rate, and temperature. These raw data constitute a multivariate time series tensor.

[0009] Data imputation module: The collected raw data, which may contain missing values, is input into the Generative Adversarial Imputation Network (GAIN) module; the missing data points are repaired with high fidelity, and a repaired data tensor with complete data and consistent distribution is output.

[0010] Hybrid Graph Construction Module: This module inputs the repaired, complete data into the hybrid adjacency matrix construction module. It generates a hybrid adjacency matrix that comprehensively characterizes the complex relationships between sensors by parallel computation of global nonlinear correlations and local spatial correlations.

[0011] The GCN-TCN diagnostic module takes the repaired data tensor and the hybrid adjacency matrix as inputs and feeds them into the GCN-TCN fusion diagnostic module; through stacked spatiotemporal processing units, it deeply extracts the spatiotemporal features from the data.

[0012] Fault classification output module: The high-dimensional feature vector output by the diagnostic module is processed by the classifier and finally outputs the diagnostic results of the current state of the pipeline system, including normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

[0013] Furthermore, the adversarial interpolation network (GAIN) module consists of a generator (G) and a discriminator (D). Through adversarial training, it processes ship pipeline sensor data containing missing information (missing data is tagged as...). This allows for adaptive repair of missing values.

[0014] (1) Generator G: Input contains missing data Missing markers and random noise Output the repaired data The purpose is to fit the distribution of the original data. The process can be represented as:

[0015]

[0016] (2) Discriminator D: Input real data Or repair data and missing markers Its task is to determine whether the input data is a true value and output a probability between (0,1).

[0017] Furthermore, the GAIN (Adversarial Interpolation Network) module also includes:

[0018] Discriminator loss The goal of the discriminator is to distinguish between real data and repaired data. Its loss function is binary cross-entropy, designed to maximize the probability of correctly classifying real and generated data.

[0019]

[0020] in: Expressing expectations; It is the discriminator's prediction of the "true probability" of real data; It is the discriminator's prediction of the true probability of repairing the data.

[0021] Generator loss The goal of the generator is to make the repaired data as close as possible to the real data, and to fool the discriminator. Its loss function consists of two parts: adversarial loss and reconstruction loss.

[0022] Adversarial loss: The aim is to make the discriminator unable to distinguish between repaired data and real data, i.e., to minimize... ;

[0023] Reconstruction loss aims to minimize the difference between the repaired data and the original data at the non-missing locations. Here, the L1 norm (representing the sum of the absolute values ​​of element-wise differences) is used as a metric.

[0024]

[0025] The generator's total loss function is:

[0026]

[0027] in It is a hyperparameter greater than 0, used to balance the ability to resist and the accuracy of repair.

[0028] Furthermore, the network training iteration process includes:

[0029] Step 1, Training the discriminator: Fix the generator parameters and input real data. (mask at this time) (a matrix of all 1s) and the repair data output by the generator. (At this point, M represents the original missing label), minimized using gradient descent. And update the discriminator parameters.

[0030] Step 2, Train the generator: Fix the discriminator parameters and input data containing missing values. , Minimize the random noise Z using gradient descent. And update the generator parameters.

[0031] Step 3, Alternating Iteration: Repeat steps 1 and 2 until the generator's repaired data reaches a balance between the discriminator's deception and the reconstruction error with the real data. At this point, training is complete, and the output is... This refers to the sensor data from the ship's pipeline after the final repair.

[0032] Furthermore, the hybrid adjacency matrix construction module specifically includes:

[0033] (1) Global correlation matrix A global (Based on the maximum information coefficient MIC):

[0034] Suppose the ship's pipeline system has N sensors, and each sensor collects time-series data as follows: ( , (Time step).

[0035] Mutual Information Quantizing data from two sensors and The degree of information correlation is calculated using the following formula:

[0036]

[0037] in, yes marginal probability, yes marginal probability, yes and The joint probability.

[0038] To eliminate the impact of binning method on mutual information, MIC iterates through all reasonable binning numbers and takes the normalized maximum mutual information, as shown in the formula:

[0039]

[0040] : right point box, point Mutual information calculated after the box;

[0041] The upper limit for the number of boxes is usually taken as... Ensure computational efficiency;

[0042] Normalize the mutual information to make .

[0043] Each pair of sensors Fill in the matrix to obtain the global correlation matrix:

[0044]

[0045] (2) Local correlation matrix A local (Based on a one-dimensional convolutional neural network (CNN)):

[0046] Sensor data repaired using GAIN , One sensor, At each time step, it is reshaped into a tensor. (Batch size is 1, number of channels is 1).

[0047] Define convolution kernel , For time step, For the number of output channels, Perform 1D convolution:

[0048]

[0049] The specific calculation for the convolution operation is as follows:

[0050]

[0051] Local feature map;

[0052] : Activation function;

[0053] : No. The bias of each channel.

[0054] right Perform global average pooling (GAP) to compress the time dimension:

[0055]

[0056] get Then, it is mapped to a similarity matrix through a fully connected layer:

[0057]

[0058] in, ,element Indicates sensor and The degree of local correlation.

[0059] (3) Matrix fusion:

[0060] To obtain a final adjacency matrix that includes both global information and local details, this invention will use A global and A local Perform fusion. The fusion method can be a learnable weighted sum:

[0061]

[0062] Here, α is a learnable hyperparameter that is automatically adjusted during model training to balance the importance of global and local information.

[0063] Furthermore, the GCN-TCN fusion diagnostic module specifically includes: a spatio-temporal network architecture that deeply integrates GCN and TCN, composed of multiple stacked spatio-temporal blocks, designed to collaboratively and deeply extract fault features. The spatial topology and temporal evolution of the pipeline are relatively independent feature dimensions; concatenation would force temporal features into already processed spatial features, destroying the integrity of the original temporal pattern. The contribution of the spatial propagation and temporal development of leakage faults to diagnosis varies with the fault type, and the concatenated architecture cannot dynamically adapt to this difference. Therefore, parallel processing is chosen, followed by dynamic attention fusion.

[0064] GCN branch: Taking the adjacency matrix and the repaired data as input, it extracts pure spatial topological features through graph convolution.

[0065]

[0066] TCN branch: Taking the repaired time series as input, it extracts pure temporal evolution features through dilated convolution.

[0067]

[0068] Dynamic attention fusion is employed, which learns the dependency weights of each sensor node on "spatial" and "temporal" features through an attention mechanism, and then performs weighted fusion.

[0069]

[0070]

[0071] For the first Spatial feature weights of individual sensors are dynamically adapted to different leakage scenarios.

[0072] Another objective of this invention is to provide a method for diagnosing leaks in ship pipelines using the GCN-TCN fusion network ship pipeline leak diagnosis system, comprising:

[0073] S1, Data Acquisition: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, multi-source heterogeneous time-series data are collected in real time, including pressure, acoustic emission, flow rate, and temperature. These raw data constitute a multivariate time-series tensor.

[0074] S2, Data Imputation: The collected raw data, which may contain missing values, is input into the Generative Adversarial Imputation Network (GAIN) module; the missing data points are repaired with high fidelity, and a repaired data tensor with complete data and consistent distribution is output.

[0075] S3, Hybrid Graph Construction: The repaired complete data is input into the hybrid adjacency matrix construction module. This module generates a hybrid adjacency matrix that can comprehensively characterize the complex relationships between sensors by parallel computing of global nonlinear correlations and local spatial correlations.

[0076] S4, GCN-TCN Diagnosis: The repaired data tensor and the hybrid adjacency matrix are simultaneously input into the GCN-TCN fusion diagnosis module; through stacked spatiotemporal processing units, the spatiotemporal features in the data are extracted in depth.

[0077] S5, Fault Classification Output: The high-dimensional feature vector output by the diagnostic module is processed by the classifier, and finally outputs the diagnostic results of the current state of the pipeline system, including normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

[0078] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, causing the processor to perform the steps of the GCN-TCN fused network ship pipeline leakage fault diagnosis method.

[0079] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the GCN-TCN fused network ship pipeline leakage fault diagnosis method.

[0080] Another objective of this invention is to provide an information data processing terminal, which includes the aforementioned GCN-TCN fusion network ship pipeline leakage fault diagnosis system.

[0081] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0082] This invention aims to overcome the aforementioned deficiencies of existing technologies and provide a robust, accurate, and adaptive method and system for diagnosing ship pipeline leaks. This invention mainly addresses the following technical problems:

[0083] Enhancing the robustness of diagnostic systems under real-world conditions: Addressing the technical challenge of frequent data loss due to sensor malfunctions or communication problems on ships, which severely impacts the performance of diagnostic models, this invention introduces a generative adversarial interpolation network to perform high-fidelity adaptive repair of missing data. This ensures the integrity and accuracy of the data input to the diagnostic model, thereby guaranteeing the stable operation and reliable diagnosis of the entire system even with incomplete data.

[0084] Achieving high-precision leak fault diagnosis: Addressing the technical challenge of low diagnostic accuracy and high false alarm rate in existing methods due to their failure to fully capture the complex spatiotemporal dynamic relationships in pipeline networks, this invention proposes an innovative spatiotemporal feature learning scheme that integrates global and local information. This scheme is achieved through the following two collaborative technical means:

[0085] Constructing a hybrid adjacency matrix: A novel, data-driven hybrid graph construction method is proposed, which can simultaneously capture the nonlinear global functional associations and complex local proximity associations between sensors, thereby providing a richer and more accurate representation of the system spatial structure for diagnostic models than existing techniques.

[0086] Deep spatiotemporal feature fusion: A network architecture that deeply integrates GCN and TCN is designed. This architecture can efficiently extract deep spatial features from the above hybrid graph and use temporal convolutional networks to capture the long-term evolution of these features in the time dimension, thereby accurately identifying weak spatiotemporal patterns caused by leakage and significantly improving the accuracy and sensitivity of diagnosis.

[0087] This invention possesses extremely high robustness and reliability: by introducing a GAIN module, it effectively addresses the common problem of missing sensor data in actual ship operations. Unlike existing technologies where performance severely degrades or even fails when data is missing, this invention can reconstruct missing information with high precision by learning the true distribution of the data, thereby ensuring the stability of the entire diagnostic system and the reliability of diagnostic results under harsh operating conditions.

[0088] This invention offers superior diagnostic accuracy: The hybrid adjacency matrix construction method designed in this invention, by fusing the global nonlinear correlations from MIC computation and the local complex patterns from CNN learning, provides GCN with far more accurate and richer graph structure information than traditional physical topology graphs. This enables the model to capture weak but crucial cross-sensor correlation patterns caused by leaks that are undetectable by traditional methods, thereby significantly improving the accuracy of fault identification and greatly reducing the false alarm rate.

[0089] This invention possesses comprehensive spatiotemporal dynamic modeling capabilities: the GCN-TCN fusion architecture of this invention is tailor-made to solve such spatiotemporal dependency problems. The GCN module effectively models the instantaneous spatial state of the pipeline network at any given moment, while the TCN module efficiently captures the long-range dynamic patterns of these spatial states evolving over time. This deep synergy between spatial and temporal modeling enables this invention to effectively detect slowly developing early leaks or complex transient anomalies, achieving a comprehensive understanding of where and when the fault occurs.

[0090] This invention possesses excellent adaptability: because the hybrid adjacency matrix is ​​dynamically constructed based on data-driven principles, this system exhibits strong adaptability. When changes occur in the ship's operating conditions (such as load and speed), leading to alterations in the fluid dynamics characteristics within the pipeline system, this system can adapt to these changes by updating the adjacency matrix online, consistently maintaining optimal diagnostic performance—something difficult to achieve with existing technologies based on fixed physical models.

[0091] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0092] This invention addresses the need for online monitoring and diagnosis of leaks and faults in critical pipeline systems such as ship fuel lines, cooling water lines, ballast water lines, and hydraulic lines. It enables real-time acquisition, missing data repair, spatiotemporal correlation modeling, and fault classification output of multi-source sensor data during ship operation. After technical transformation, this solution can be integrated into ship intelligent operation and maintenance systems, engine room condition monitoring systems, ship fault early warning platforms, or port remote operation and maintenance platforms to achieve early leak identification, fault level judgment, and auxiliary fault area location in ship pipeline systems.

[0093] Compared to methods relying on manual inspections, single-sensor threshold alarms, or traditional physical model analysis, this invention maintains good diagnostic stability and identification accuracy even when sensor data is missing, there is noise interference, and operating conditions change. This reduces equipment downtime, energy loss, increased maintenance costs, and safety risks caused by leaks. For ship operators, this invention improves the timeliness of fault detection and reduces unplanned downtime and maintenance costs. For ship intelligent manufacturing and intelligent operation and maintenance companies, this invention can serve as a core algorithm module in ship pipeline condition monitoring products, ship digital twin systems, or predictive maintenance systems, demonstrating significant engineering promotion value and commercial application prospects.

[0094] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0095] Existing ship pipeline leak diagnosis technologies mostly employ acoustic detection, pressure wave analysis, flow balance analysis, or alarm methods based on fixed thresholds. While these methods are effective in laboratory or stable operating conditions, they are easily affected by factors such as engine room vibration, pump and valve impact, fluid pulsation, sea state changes, and sensor communication anomalies in the actual operating environment of ships, leading to high false alarm and false negative rates. In recent years, although deep learning methods have been increasingly applied to industrial fault diagnosis, most methods directly use sensor data as ordinary time series input models, lacking joint modeling of the complex spatial correlations and temporal evolution patterns among sensors in ship pipeline systems.

[0096] Especially in the scenario of diagnosing leaks in ship pipelines, missing sensor data is a common phenomenon. Existing methods typically use simple methods such as mean imputation, linear interpolation, and forward imputation to handle missing data. These methods struggle to maintain the original spatiotemporal correlation structure between multi-sensor data, potentially further weakening the performance of subsequent fault diagnosis models. Furthermore, existing graph neural network methods often rely on a fixed physical topology graph to construct an adjacency matrix. However, in ship pipeline systems, the functional relationships between sensors dynamically change under different loads, speeds, pump and valve states, and fluid conditions. A fixed topology cannot accurately reflect the true fault propagation relationships.

[0097] To address the aforementioned shortcomings, this invention proposes a ship pipeline leakage fault diagnosis scheme that combines generative adversarial interpolation networks, hybrid adjacency matrix construction, and GCN-TCN parallel spatiotemporal feature fusion. This scheme not only considers the high-fidelity repair problem in the case of missing ship sensor data, but also extracts global nonlinear correlations through maximum information coefficients, extracts local correlation features through convolutional neural networks, and further utilizes GCN and TCN to model spatial topological features and temporal evolution features, respectively. Therefore, this invention provides a novel technical approach in the field of ship pipeline leakage fault diagnosis that balances data integrity repair, dynamic correlation modeling, and spatiotemporal feature fusion, overcoming the shortcomings of existing technologies in robust diagnosis under complex ship operating conditions.

[0098] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0099] Ship piping systems are constantly exposed to vibration, shock, temperature and humidity variations, strong noise, and complex fluid disturbances. Early signs of leaks are often weak and easily masked by fluctuations in normal operating conditions. How to accurately identify early leaks and leaks of different levels, even with missing sensor data, noise, and changing operating conditions, has long been a challenging technical problem in the field of intelligent ship maintenance.

[0100] Traditional leak diagnosis methods typically rely on changes in a single physical quantity or empirical thresholds. When the leak is small, the leak location is concealed, or the pipeline system has complex branches, changes in a single sensor signal are not significant, making it difficult to reliably determine the leak status. Even when using multiple sensors for monitoring, if the spatial correlation and temporal evolution patterns between different sensors cannot be effectively characterized, it is difficult to extract stable leak characteristics from complex data. Furthermore, when some sensor data is missing, traditional diagnostic methods often directly discard abnormal data or use simple interpolation to complete it, which can easily disrupt the original data distribution and the correlation structure between multiple sensors, leading to unstable diagnostic results.

[0101] This invention adaptively repairs missing sensor data using a generative adversarial interpolation network, ensuring that the repaired data maintains distribution characteristics as consistent as possible with the real data. It constructs global and local correlation matrices using MIC and CNN respectively, enabling the adjacency matrix to simultaneously reflect nonlinear functional correlations and local proximity correlations between sensors. A parallel GCN-TCN structure extracts spatial and temporal features, and a dynamic attention mechanism is used to achieve adaptive feature fusion under different leakage scenarios. Therefore, this invention solves the technical challenge of traditional methods simultaneously addressing missing data repair, dynamic spatial correlation modeling, and long-range temporal feature extraction, improving the accuracy, robustness, and engineering applicability of ship pipeline leakage fault diagnosis.

[0102] (4) The technical solution of the present invention overcomes technical bias: The present invention overcomes the technical bias of relying on fixed physical topology for leak diagnosis, and proposes to construct a hybrid adjacency matrix using a data-driven approach. This adjacency matrix not only considers the global nonlinear correlation between sensors, but also the proximity correlation reflected by local temporal characteristics, thereby enabling a more realistic characterization of the dynamic correlation of the ship pipeline system under actual operating conditions. Attached Figure Description

[0103] Figure 1 This is a schematic diagram of the GCN-TCN fused network ship pipeline leakage fault diagnosis system provided in an embodiment of the present invention;

[0104] Figure 2 This is a flowchart of the GAIN process provided in an embodiment of the present invention.

[0105] Figure 3 This is the construction of the adjacency matrix provided in the embodiments of the present invention.

[0106] Figure 4 This is the GCN-TCN fusion diagnostic module provided in this embodiment of the invention.

[0107] Figure 5 This is a flowchart of the GCN-TCN fusion network ship pipeline leakage fault diagnosis method provided in the embodiments of the present invention.

[0108] Figure 6 is a schematic diagram of the experimental platform and sensor arrangement of the ship pipeline at Wuhan University of Technology provided in an embodiment of the present invention.

[0109] Figure 7 This is a confusion matrix of classification results for different leakage states under three-pump operating conditions provided in the embodiments of the present invention.

[0110] Figure 8 This is the t-SNE visualization result of the characteristic distribution of different leakage states under three-pump operating conditions provided in the embodiment of the present invention. Detailed Implementation

[0111] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0112] like Figure 1 As shown, the overall architecture of the fault diagnosis system provided in this embodiment of the invention includes five core parts: data acquisition, data interpolation module, hybrid graph construction module, GCN-TCN diagnosis module, and fault classification output. Its detailed workflow is as follows:

[0113] (1) Data acquisition module: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, multi-source heterogeneous time series data, such as pressure, acoustic emission, flow rate, and temperature, are collected in real time. These raw data constitute a multivariate time series tensor.

[0114] (2) Data imputation module: The collected raw data, which may contain missing values, is input into the Generative Adversarial Imputation Network (GAIN) module. This module performs high-fidelity repair on the missing data points and outputs a repaired data tensor with complete data and consistent distribution.

[0115] (3) Hybrid Graph Construction Module: The repaired complete data is input into the hybrid adjacency matrix construction module. This module generates a hybrid adjacency matrix that can comprehensively characterize the complex relationships between sensors by parallel computing of global nonlinear correlation and local spatial correlation.

[0116] (4) GCN-TCN diagnostic module: The repaired data tensor and the hybrid adjacency matrix are simultaneously input into the GCN-TCN fusion diagnostic module. This module deeply extracts the spatiotemporal features in the data through stacked spatiotemporal processing units.

[0117] (5) Fault classification output module: The high-dimensional feature vector output by the diagnostic module is processed by the classifier and finally outputs the diagnostic results of the current state of the pipeline system, such as normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

[0118] Data interpolation module: Generate adversarial interpolation network (GAIN).

[0119] The present invention provides a GCN-TCN fusion network ship pipeline leakage fault diagnosis system, which realizes the diagnosis of pipeline leakage status and fault area by continuously collecting multi-source sensor data during the operation of ship pipelines, repairing missing data, modeling sensor correlation relationships, and jointly extracting spatiotemporal features.

[0120] During system operation, a distributed sensor network deployed at key nodes of the ship's pipeline continuously acquires pressure, acoustic emission, flow, and temperature data. Based on sampling time and sensor nodes, this data is uniformly organized to form a multivariate time-series tensor characterizing the pipeline's operating status. Due to factors such as vibration, electromagnetic interference, communication anomalies, and short-term sensor failures in the ship's operating environment, some data may be missing. Therefore, the original time series containing missing values ​​is input into a generator adversarial interpolation network. This network reconstructs the missing locations based on the temporal variation patterns of the observed data and the data distribution relationships between different variables, obtaining a complete, repaired data tensor, providing a continuous data foundation for subsequent fault characteristic analysis.

[0121] Subsequently, the repaired data is input into the hybrid adjacency matrix construction module. On the one hand, the correlation at the data level is determined based on the global nonlinear correlation between the time series of different sensors. On the other hand, the physical connection relationship is determined by combining the local spatial correlation of sensors in the ship's pipeline. The two types of relationships are then fused to form a hybrid adjacency matrix that describes the degree of correlation of each sensor node, so that the actual spatial structure of the pipeline network and the data correlation generated under the operating state can be characterized simultaneously.

[0122] During the fault diagnosis phase, the repaired data tensor and the hybrid adjacency matrix are input into the GCN-TCN fusion diagnostic module. GCN uses the hybrid adjacency matrix to propagate and aggregate features among interconnected sensor nodes, extracting spatial correlation features formed by the propagation of leakage disturbances along the pipeline. TCN, on the other hand, performs convolution processing on continuously sampled data along the time dimension, extracting temporal evolution features exhibited by pressure fluctuations, acoustic emission changes, flow anomalies, and temperature variations. By stacking spatiotemporal processing units, spatial correlation information and temporal dynamic information are fused layer by layer, forming high-dimensional spatiotemporal features with leakage state identification capabilities.

[0123] Finally, the high-dimensional spatiotemporal features are input into the fault classifier, and diagnostic results such as normal, minor leakage, and severe leakage are output according to the feature distribution corresponding to different leakage states. Combined with the sensor nodes that produce significant abnormal responses and their spatial correlation, the pipeline areas that may leak are correlated and judged, so as to realize the state identification and area location of ship pipeline leakage faults.

[0124] like Figure 2 As shown, this invention employs a Generative Adversarial Interpolation (GAIN) module. This module consists of a generator (G) and a discriminator (D), and through adversarial training, it processes ship pipeline sensor data containing missing information (missing data is marked as...). This allows for adaptive repair of missing values.

[0125] (1) Generator G: Input contains missing data Missing markers and random noise Output the repaired data The purpose is to fit the distribution of the original data. The process can be represented as:

[0126]

[0127] (2) Discriminator D: Input real data Or repair data and missing markers Its task is to determine whether the input data is a true value and output a probability between (0,1).

[0128] 1. Discriminator loss

[0129] The discriminator aims to distinguish between real data and repaired data. Its loss function is binary cross-entropy, designed to maximize the probability of correctly classifying real and generated data.

[0130]

[0131] in:

[0132] Expressing expectations;

[0133] It is the discriminator's prediction of the "true probability" of real data;

[0134] It is the discriminator's prediction of the true probability of repairing the data.

[0135] 2. Generator Loss

[0136] The goal of the generator is to make the repaired data as close as possible to the real data, and to fool the discriminator. Its loss function consists of two parts: adversarial loss and reconstruction loss.

[0137] Adversarial loss: The aim is to make the discriminator unable to distinguish between repaired data and real data, i.e., to minimize... ;

[0138] Reconstruction loss aims to minimize the difference between the repaired data and the original data at the non-missing locations. Here, the L1 norm (representing the sum of the absolute values ​​of element-wise differences) is used as a metric.

[0139]

[0140] In summary, the total loss function of the generator is:

[0141]

[0142] in It is a hyperparameter greater than 0, used to balance the ability to resist and the accuracy of repair.

[0143] Network training iterative process

[0144] Step 1: Train the discriminator

[0145] Fixed generator parameters, input real data (mask at this time) (a matrix of all 1s) and the repair data output by the generator. (At this point, M represents the original missing label), minimized using gradient descent. And update the discriminator parameters.

[0146] Step 2: Training the generator

[0147] With fixed discriminator parameters, input data containing missing values. , Minimize the random noise Z using gradient descent. And update the generator parameters.

[0148] Step 3: Alternating Iteration

[0149] Repeat steps 1 and 2 until the generator's repaired data reaches a balance between the discriminator's deception and the reconstruction error with the real data. At this point, training is complete, and the output is... This refers to the sensor data from the ship's pipeline after the final repair.

[0150] Hybrid adjacency matrix building module:

[0151] like Figure 3 As shown, the performance of graph convolutional networks largely depends on the quality of their adjacency matrix, which defines the information transmission paths between nodes. This invention innovatively proposes a data-driven hybrid adjacency matrix construction method to replace the static adjacency matrix based on physical topology, thereby more accurately characterizing the functional correlations between sensors in a ship pipeline system. This module constructs a global correlation matrix and a local correlation matrix in parallel and then fuses them.

[0152] (1) Global correlation matrix A global (Based on the maximum information coefficient MIC):

[0153] Suppose the ship's pipeline system has N sensors, and each sensor collects time-series data as follows: ( , (Time step).

[0154] Mutual Information Quantizing data from two sensors and The degree of information correlation is calculated using the following formula:

[0155]

[0156] in, yes marginal probability, yes marginal probability, yes and The joint probability.

[0157] To eliminate the impact of binning method on mutual information, MIC iterates through all reasonable binning numbers and takes the normalized maximum mutual information, as shown in the formula:

[0158]

[0159] : right point box, point Mutual information calculated after the bin;

[0160] : Upper limit of the number of boxes (usually taken as ) (Ensuring computational efficiency);

[0161] Normalize the mutual information to make .

[0162] Each pair of sensors Fill in the matrix to obtain the global correlation matrix:

[0163]

[0164] (2) Local correlation matrix A local (Based on a one-dimensional convolutional neural network (CNN)):

[0165] Sensor data repaired using GAIN ( One sensor, (at each time step), reshape it into a tensor (Batch size is 1, number of channels is 1).

[0166] Define convolution kernel ( For time step, (number of output channels), for Perform 1D convolution (temporal dimension convolution)

[0167]

[0168] The specific calculation for the convolution operation is as follows:

[0169]

[0170] Local feature map;

[0171] Activation function (e.g., ReLU);

[0172] : No. The bias of each channel.

[0173] right Perform global average pooling (GAP) to compress the time dimension:

[0174]

[0175] get Then, it is mapped to a similarity matrix through a fully connected layer:

[0176]

[0177] in, ,element Indicates sensor and The degree of local correlation.

[0178] (3) Matrix fusion:

[0179] To obtain a final adjacency matrix that includes both global information and local details, this invention will use A global and A local Perform fusion. The fusion method can be a learnable weighted sum:

[0180]

[0181] Here, α is a learnable hyperparameter that is automatically adjusted during model training to balance the importance of global and local information.

[0182] GCN-TCN fusion diagnostic module: such as Figure 4 As shown, the core diagnostic module of this invention adopts a spatio-temporal network architecture that deeply integrates GCN and TCN. This architecture consists of multiple stacked spatio-temporal blocks, designed to collaboratively and deeply extract fault features. The spatial topology of the pipeline (such as the connection relationship of sensor nodes) and temporal evolution (such as the decay of pressure over time) are relatively independent feature dimensions. Concatenation would force temporal features to be embedded in the processed spatial features, destroying the integrity of the original temporal pattern. The contribution of the spatial propagation and temporal development of leakage faults to diagnosis varies with the fault type (e.g., small leaks depend more on temporal trends, while multi-location leaks depend more on spatial correlations), and the concatenated architecture cannot dynamically adapt to this difference. Therefore, parallel processing is chosen, followed by dynamic attention fusion.

[0183] GCN branch: Taking the adjacency matrix and the repaired data as input, it extracts pure spatial topological features through graph convolution.

[0184]

[0185] TCN branch: Taking the repaired time series as input, it extracts pure temporal evolution features through dilated convolution.

[0186]

[0187] Dynamic attention fusion is employed, which learns the dependency weights of each sensor node on "spatial" and "temporal" features through an attention mechanism, and then performs weighted fusion.

[0188]

[0189]

[0190] For the first Spatial feature weights of individual sensors are dynamically adapted to different leakage scenarios.

[0191] like Figure 5 As shown, the GCN-TCN fusion network ship pipeline leakage fault diagnosis method provided in this embodiment of the invention includes:

[0192] S1, Data Acquisition: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, multi-source heterogeneous time-series data are collected in real time, including pressure, acoustic emission, flow rate, and temperature. These raw data constitute a multivariate time-series tensor.

[0193] S2, Data Imputation: The collected raw data, which may contain missing values, is input into the Generative Adversarial Imputation Network module; the missing data points are repaired with high fidelity, and a repaired data tensor with complete data and consistent distribution is output;

[0194] S3, Hybrid Graph Construction: The repaired complete data is input into the hybrid adjacency matrix construction module; this module generates a hybrid adjacency matrix that can comprehensively characterize the complex relationships between sensors by parallel computing of global nonlinear correlation and local spatial correlation.

[0195] S4, GCN-TCN Diagnosis: The repaired data tensor and the hybrid adjacency matrix are simultaneously input into the GCN-TCN fusion diagnosis module; the spatiotemporal features in the data are deeply extracted through stacked spatiotemporal processing units;

[0196] S5, Fault Classification Output: The high-dimensional feature vector output by the diagnostic module is processed by the classifier, and finally outputs the diagnostic results of the current state of the pipeline system, including normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

[0197] An application embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the GCN-TCN fusion network ship pipeline leakage fault diagnosis method.

[0198] An application embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the GCN-TCN fused network ship pipeline leakage fault diagnosis method.

[0199] An application embodiment of the present invention provides an information data processing terminal, which includes a GCN-TCN fusion network ship pipeline leakage fault diagnosis system.

[0200] The embodiments of the present invention can be applied to scenarios such as online status monitoring, leakage fault diagnosis, fault level identification and intelligent operation and maintenance early warning of ship pipeline systems, and are especially suitable for ship engine room fuel pipelines, cooling water pipelines, ballast water pipelines, hydraulic pipelines and other complex pipeline systems with multi-sensor monitoring requirements.

[0201] In practical applications, this invention can be deployed in ship engine room status monitoring systems, ship pipeline leak early warning systems, ship intelligent operation and maintenance platforms, ship digital twin systems, and shore-based remote monitoring platforms. By deploying sensors for pressure, flow, temperature, acoustic emission, and vibration at key pipeline nodes, this invention can continuously collect ship pipeline operating status data and automatically identify pipeline leaks using a data interpolation module, a hybrid graph construction module, and a GCN-TCN fusion diagnostic module.

[0202] Specifically, the present invention can form the following related products or system modules:

[0203] (1) Online diagnostic system for ship pipeline leakage faults: used for real-time acquisition of ship pipeline system operation data, repair of missing data, identification of leakage faults and output of fault level;

[0204] (2) Fault diagnosis algorithm module in the intelligent operation and maintenance platform of ship engine room: As the core algorithm unit in the integrated monitoring platform of ship engine room, it performs intelligent analysis of pipeline leakage, sensor abnormality and changes in operating status;

[0205] (3) Ship Pipeline Health Status Monitoring Terminal: Deployed in the ship's local edge computing equipment or information data processing terminal to realize on-site monitoring and early warning of the pipeline system's operating status;

[0206] (4) Shore-based remote fault diagnosis and maintenance decision-making platform: The data collected by the ship is transmitted to the shore-based platform, and the model described in this invention performs status identification, providing a basis for remote operation and maintenance, maintenance plan formulation and safety management;

[0207] (5) Pipeline status perception module in ship digital twin system: The normal status, leakage status of different degrees and their evolution trend output by the present invention are used as the status input of the digital twin model to realize the visualization of the operating status of the ship pipeline system and fault early warning.

[0208] Through the above applications, this invention can improve the automation level of ship pipeline leakage fault diagnosis, reduce reliance on manual inspection and experience judgment, and reduce equipment downtime, media loss, increased maintenance costs and safety accident risks caused by leakage faults. It has good engineering application value and industrialization prospects.

[0209] To verify the effectiveness of the GCN-TCN fused network ship pipeline leakage fault diagnosis system and method described in this invention, experimental verification was conducted on the ship pipeline experimental platform at Wuhan University of Technology. Figure 6As shown, the experimental platform includes a main body of the pipeline experimental platform, pump set, flow sensor, pressure sensor, vibration sensor and valve opening regulator, etc. It can simulate the fluid transport process of the ship pipeline system under different pump set operating states and different valve opening conditions, and collect multi-source sensor data during the operation of the pipeline system.

[0210] In this embodiment, the experimental platform operates with three pumps. Different levels of leakage are simulated by adjusting valve openings, and pipeline operation data are collected under both normal and leaking conditions. The experimental setup includes six operating states: normal, 10% leakage, 20% leakage, 30% leakage, 40% leakage, and 50% leakage. Sensor data collected under each state includes, but is not limited to, pressure, flow, and vibration signals, used to characterize the dynamic response of the ship's pipeline system under different levels of leakage.

[0211] To further verify the robustness of this invention in the face of incomplete data under real-world ship operating conditions, this embodiment introduces a 30% random missing component in the normally operating data before testing. This simulates potential sensor malfunctions, communication packet loss, sampling anomalies, or localized data loss that may occur during actual ship operation. Subsequently, the sensor data containing missing values ​​is input into an adversarial interpolation network module to repair the missing data. Then, a hybrid adjacency matrix construction module extracts the global nonlinear correlations and local dynamic correlations between sensors. Finally, the data is input into a GCN-TCN fusion diagnostic module to complete the leakage status classification.

[0212] like Figure 7 As shown in the confusion matrix results, the present invention can effectively identify normal conditions and different leakage levels under three-pump operating conditions. Except for a small number of normal samples that were misclassified as 10% leakage, the remaining 10%, 20%, 30%, 40%, and 50% leakage states were accurately classified. Specifically, 62 samples in the normal state were correctly identified, and 10 were misclassified as 10% leakage; 72 samples each in the 10%, 20%, 30%, 40%, and 50% leakage states were correctly identified. The total number of samples in the experiment was 432, with 422 correctly identified, achieving an overall classification accuracy of 97.69%.

[0213] The above results demonstrate that this invention maintains a high accuracy rate in identifying leakage faults even with 30% random data loss. On one hand, the generative adversarial interpolation network effectively repairs missing sensor data, reducing the impact of data loss on the input quality of subsequent diagnostic models. On the other hand, the hybrid adjacency matrix and GCN-TCN fusion diagnostic module fully exploit the spatial correlation and temporal evolution features in multi-source sensor data, thereby improving the ability to distinguish between different leakage levels.

[0214] like Figure 8 As shown, this invention further employs the t-SNE method to perform two-dimensional visualization analysis on the high-dimensional features extracted by the diagnostic model. The visualization results show that the normal state, 10% leakage state, 20% leakage state, 30% leakage state, 40% leakage state, and 50% leakage state form relatively clear cluster distributions in the two-dimensional feature space, with obvious separation boundaries between different categories. Furthermore, the feature clusters corresponding to different leakage levels are distributed in different regions, indicating that the model described in this invention can learn deep features that are discriminative between different leakage levels.

[0215] The confusion matrix and t-SNE visualization results together demonstrate that this invention not only achieves high classification accuracy but also effectively distinguishes different leakage levels in the feature space. These results prove that the proposed generative adversarial interpolation network, hybrid adjacency matrix construction method, and GCN-TCN parallel spatiotemporal feature fusion structure can effectively capture the spatial correlation changes and temporal evolution characteristics caused by leakage faults in ship pipeline systems.

[0216] In summary, this invention, through its technical process of generating an adversarial interpolation network for missing data repair, constructing a hybrid adjacency matrix, extracting spatiotemporal features using GCN-TCN, dynamic attention fusion, and outputting fault classification, achieved a classification accuracy of 97.69% on three-pump operating data from the ship pipeline experimental platform at Wuhan University of Technology. This experimental result demonstrates that this invention can still achieve high-precision identification of different leakage levels even when sensor data has random gaps, exhibiting good robustness, accuracy, and engineering applicability.

[0217] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0218] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A GCN-TCN fused network ship pipeline leakage fault diagnosis system, characterized in that, include: Data acquisition module: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, it collects multi-source heterogeneous time series data in real time, including pressure, acoustic emission, flow rate, and temperature. These raw data constitute a multivariate time series tensor. Data imputation module: Inputs the collected raw data, which may contain missing values, into the generative adversarial imputation network module; Perform high-fidelity repair on missing data points and output a repaired data tensor with complete data and consistent distribution; Hybrid Graph Construction Module: Input the repaired complete data into the hybrid adjacency matrix construction module; this module generates a hybrid adjacency matrix that can comprehensively characterize the complex relationships between sensors by parallel computing of global nonlinear correlation and local spatial correlation; GCN-TCN diagnostic module: The repaired data tensor and the hybrid adjacency matrix are simultaneously input into the GCN-TCN fusion diagnostic module; through stacked spatiotemporal processing units, the spatiotemporal features in the data are extracted in depth. Fault classification output module: The high-dimensional feature vector output by the diagnostic module is processed by the classifier and finally outputs the diagnostic results of the current state of the pipeline system, including normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

2. The GCN-TCN converged network ship pipeline leakage fault diagnosis system according to claim 1, characterized in that, The adversarial interpolation network module consists of a generator and a discriminator. Through adversarial training, it repairs missing ship pipeline sensor data, achieving adaptive repair of missing values. (1) Generator G: Input contains missing data Missing markers and random noise Output the repaired data The purpose is to fit the distribution of the original data; The process can be represented as: ； (2) Discriminator D: Input real data Or repair data and missing markers Its task is to determine whether the input data is a true value and output a probability between (0,1).

3. The GCN-TCN converged network ship pipeline leakage fault diagnosis system according to claim 1, characterized in that, The adversarial interpolation network module also includes: Discriminator loss The discriminator aims to distinguish between real data and repaired data; its loss function is binary cross-entropy, designed to maximize the probability of correctly identifying real and generated data. ； in: Expressing expectations; It is the discriminator's prediction of the "true probability" of real data; It is the discriminator's prediction of the true probability of the repaired data; Generator loss The goal of the generator is to make the repaired data as close as possible to the real data and to fool the discriminator; its loss function consists of two parts: adversarial loss and reconstruction loss. Adversarial loss: The aim is to make the discriminator unable to distinguish between repaired data and real data, i.e., to minimize... ; Reconstruction loss: The goal is to minimize the difference between the repaired data and the original data at the non-missing locations; here, the L1 norm (representing the sum of the absolute values ​​of element-wise differences) is used as the metric. ； The generator's total loss function is: ； in It is a hyperparameter greater than 0, used to balance the ability to resist and the accuracy of repair.

4. The GCN-TCN converged network ship pipeline leakage fault diagnosis system according to claim 1, characterized in that, The network training iteration process includes: Step 1, Training the discriminator: Fix the generator parameters and input real data. and the repair data output by the generator Minimize using gradient descent And update the discriminator parameters; Step 2, Train the generator: Fix the discriminator parameters and input data containing missing values. , Minimize the random noise Z using gradient descent. and update the generator parameters; Step 3, Alternating Iteration: Repeat steps 1 and 2 until the generator's repaired data reaches a balance between the discriminator's deception and the reconstruction error with the real data; at this point, training is complete, and the output is... This refers to the sensor data from the ship's pipeline after the final repair.

5. The GCN-TCN converged network ship pipeline leakage fault diagnosis system according to claim 1, characterized in that, The hybrid adjacency matrix construction module specifically includes: (1) Global correlation matrix A global : Suppose the ship's pipeline system has N sensors, and each sensor collects time-series data as follows: , , For time step; Mutual Information Quantizing data from two sensors and The degree of information correlation is calculated using the following formula: ； in, yes marginal probability, yes marginal probability, yes and The joint probability; To eliminate the impact of binning method on mutual information, MIC iterates through all reasonable binning numbers and takes the normalized maximum mutual information, as shown in the formula: ； : right point box, point Mutual information calculated after the box; The upper limit for the number of boxes is usually taken as... Ensure computational efficiency; Normalize the mutual information to make ; Each pair of sensors Fill in the matrix to obtain the global correlation matrix: ； (2) Local correlation matrix A local : Sensor data repaired using GAIN , One sensor, At each time step, it is reshaped into a tensor. ; Define convolution kernel , For time step, For the number of output channels, Perform 1D convolution: ； The specific calculation for the convolution operation is as follows: ； Local feature map; : Activation function; : No. The bias of each channel; right Perform global average pooling to compress the time dimension: ； get Then, it is mapped to a similarity matrix through a fully connected layer: ； in, ,element Indicates sensor and Local correlation degree; (3) Matrix fusion: To obtain a final adjacency matrix that includes both global information and local details, this invention will use A global and A local To perform fusion; the fusion method can be a learnable weighted sum: ； Here, α is a learnable hyperparameter that is automatically adjusted during model training to balance the importance of global and local information.

6. The GCN-TCN converged network ship pipeline leakage fault diagnosis system according to claim 1, characterized in that, The GCN-TCN fusion diagnostic module specifically includes: a spatiotemporal network architecture that deeply integrates GCN and TCN, composed of multiple stacked spatiotemporal blocks, designed to collaboratively and deeply extract fault features; the spatial topology and temporal evolution of the pipeline are relatively independent feature dimensions, and concatenation would force temporal features to be embedded in the processed spatial features, destroying the integrity of the original temporal pattern; the contribution of the spatial propagation and temporal development of leakage faults to diagnosis varies with the fault type, and the concatenated architecture cannot dynamically adapt to this difference; therefore, parallel processing is chosen, and finally dynamic attention fusion is performed; GCN branch: Taking the adjacency matrix and the repaired data as input, it extracts pure spatial topological features through graph convolution. ； TCN branch: Taking the repaired time series as input, it extracts pure temporal evolution features through dilated convolution. ； Dynamic attention fusion is employed, which learns the dependence weights of each sensor node on spatial and temporal features through an attention mechanism, and then performs weighted fusion: ； ； For the first Spatial feature weights of individual sensors are dynamically adapted to different leakage scenarios.

7. A method for diagnosing leaks in ship pipelines using the GCN-TCN fusion network ship pipeline leakage fault diagnosis system as described in any one of claims 1 to 6, characterized in that, include: S1, Data Acquisition: Through a distributed sensor network deployed at key nodes of the ship's pipeline system, multi-source heterogeneous time-series data are collected in real time, including pressure, acoustic emission, flow rate, and temperature. These raw data constitute a multivariate time-series tensor. S2, Data imputation: The collected raw data, which may contain missing values, is input into the generative adversarial imputation network module; Perform high-fidelity repair on missing data points and output a repaired data tensor with complete data and consistent distribution; S3, Hybrid Graph Construction: The repaired complete data is input into the hybrid adjacency matrix construction module; this module generates a hybrid adjacency matrix that can comprehensively characterize the complex relationships between sensors by parallel computing of global nonlinear correlation and local spatial correlation. S4, GCN-TCN Diagnosis: The repaired data tensor and the hybrid adjacency matrix are simultaneously input into the GCN-TCN fusion diagnosis module; the spatiotemporal features in the data are deeply extracted through stacked spatiotemporal processing units; S5, Fault Classification Output: The high-dimensional feature vector output by the diagnostic module is processed by the classifier, and finally outputs the diagnostic results of the current state of the pipeline system, including normal, minor leakage, serious leakage, etc., and can be further associated with possible fault areas.

8. A computer device comprising a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the GCN-TCN fusion network ship pipeline leakage fault diagnosis method as described in claim 7.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the GCN-TCN converged network ship pipeline leakage fault diagnosis method as described in claim 7.

10. An information data processing terminal, comprising the GCN-TCN converged network ship pipeline leakage fault diagnosis system as described in any one of claims 1 to 6.

Images