An underground pipeline safety hazard prediction method based on privacy computing

Abstract

The application discloses a kind of underground pipeline safety hidden danger prediction methods based on privacy computing, it is related to underground pipeline safety management technical field, including, pipeline structure data, pipeline operation data and environmental data are collected and are standardized desensitization processing and fusion by differential privacy mechanism, generate pipeline privacy data;The graph structure construction operation is executed to pipeline privacy data, and generates privacy spatiotemporal graph set;Privacy spatiotemporal graph set is input into privacy protection computing environment formed based on multi-party secure computation protocol and homomorphic encryption, and graph feature extraction and timing feature coding processing are executed to privacy spatiotemporal graph set, generate encrypted graph structure feature and encrypted timing feature, through spatiotemporal feature fusion method, encrypted graph structure feature and encrypted timing feature are fused, and generate spatiotemporal risk data.The application is input into privacy protection computing environment by privacy spatiotemporal graph set and executes graph convolution operation and long short-term memory network coding, realizes the deep feature fusion of multi-source heterogeneous data.

Description

Technical Field

[0001] This invention relates to the field of underground pipeline safety management technology, and in particular to a method for predicting underground pipeline safety hazards based on privacy computing. Background Technology

[0002] As a crucial infrastructure for urban energy transmission and municipal security, underground pipelines are becoming increasingly complex in scale and structure with the advancement of urbanization and integrated development of underground space. They operate under conditions of high pressure, high corrosion, and high interference. To avoid leaking underlying business details while conducting joint analysis among multiple stakeholders and business units, the industry has begun to introduce privacy-preserving computation frameworks. These frameworks combine data security computation techniques such as differential privacy, secure multi-party computation, and homomorphic encryption with underground pipeline safety assessments, exploring technical pathways for hazard identification and risk classification without directly exposing raw data.

[0003] Existing methods still have some shortcomings. In the data fusion and feature modeling stages, it is difficult to perform detailed spatiotemporal correlation modeling of multi-source heterogeneous data under strict privacy constraints. Although pipeline structure data, operation monitoring data, and environmental data can participate in the analysis separately under encrypted or anonymized conditions, there is a lack of a unified graph structure and time series representation to characterize the coupling characteristics between pipeline topological connections, operating condition evolution trajectories, and environmental exposure intensity. This makes it difficult to reveal the risk patterns of hazard transmission along structural paths and the evolution driven by complex operating conditions in a timely manner. In the sample construction and risk learning stages, existing methods mostly rely on limited historical hazard events and accident records. Under privacy constraints, they lack the ability to systematically simulate and expand high-risk scenarios, making it difficult to generate a sample set of hazard scenarios covering multiple operating conditions and multiple disturbance combinations. This affects the ability of underground pipeline safety hazard prediction models to identify extreme risk situations and their generalization performance. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for predicting underground pipeline safety hazards based on privacy computing to solve the problems of data fusion difficulties under privacy conditions and insufficient samples in high-risk scenarios.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for predicting safety hazards in underground pipelines based on privacy computing. The method includes: collecting pipeline structure data, pipeline operation data, and environmental data; standardizing and desensitizing these data using a differential privacy mechanism and then fusing them to generate pipeline privacy data; performing graph structure construction on the pipeline privacy data to generate a privacy spatiotemporal graph atlas; inputting the privacy spatiotemporal graph atlas into a privacy-protected computing environment based on a multi-party secure computation protocol and homomorphic encryption; performing graph feature extraction and temporal feature encoding on the privacy spatiotemporal graph atlas to generate encrypted graph structure features and encrypted temporal features; fusing the encrypted graph structure features and encrypted temporal features using a spatiotemporal feature fusion method to generate spatiotemporal risk data; and constructing a hazard scenario generation model and inputting the spatiotemporal risk data into the model. The hazard scenario generation model performs feature combination, Gaussian noise perturbation, and transposed convolution expansion operations on spatiotemporal risk data to generate a hazard scenario sample set. In a privacy-preserving computing environment, a pipeline safety hazard prediction model is constructed using the spatiotemporal risk data and the hazard scenario sample set. Real-time pipeline operation data is collected and generated using the same processing flow as the spatiotemporal risk data generation model. This real-time node-level spatiotemporal risk data is input into the pipeline safety hazard prediction model to generate safety risk level identifiers for corresponding graph structure nodes. The safety risk level identifiers are spatially correlated with the pipeline structure data and combined with environmental data to form a safety hazard risk distribution map. Based on different levels of safety risk level identifiers, corresponding early warning information and operation and maintenance decisions are generated.

[0007] As a preferred embodiment of the privacy-preserving computation-based method for predicting safety hazards in underground pipelines according to the present invention, the following steps are taken: The collection of pipeline structure data, pipeline operation data, and environmental data is standardized, de-identified, and fused using a differential privacy mechanism to generate pipeline privacy data. Collect pipeline structure data, pipeline operation data, and environmental data, and perform format standardization and missing value processing to obtain preprocessed pipeline structure data, pipeline operation data, and environmental data; The preprocessed pipeline structure data, pipeline operation data, and environmental data are standardized and anonymized using a differential privacy mechanism to generate privacy-preserving pipeline structure data, privacy-preserving pipeline operation data, and privacy-preserving environmental data. Based on pipeline identification and corresponding time information, privacy pipeline structure data, privacy pipeline operation data, and privacy environment data are integrated to generate pipeline privacy data.

[0008] As a preferred embodiment of the privacy-based computation-based method for predicting underground pipeline safety hazards according to the present invention, the specific steps for performing graph structure construction operations on pipeline privacy data to generate a privacy spatiotemporal atlas are as follows: Each pipeline in the pipeline privacy data is mapped to a graph structure node, and the connection relationships and spatial proximity relationships in the pipeline privacy data are mapped to graph structure edges to generate a privacy graph structure. By integrating the privacy graph structures corresponding to information from different times, a privacy spatiotemporal atlas is formed.

[0009] As a preferred embodiment of the privacy-preserving computation-based method for predicting safety hazards in underground pipelines according to the present invention, the steps of inputting a privacy-preserving spatiotemporal atlas into a privacy-protected computing environment and performing graph feature extraction and temporal feature encoding on the privacy-preserving spatiotemporal atlas to generate encrypted graph structural features and encrypted temporal features are as follows. A privacy-preserving computing environment is formed based on a multi-party secure computation protocol and homomorphic encryption. The privacy-preserving spatiotemporal atlas is input into the privacy-preserving computing environment and homomorphically encrypted to generate an encrypted privacy-preserving spatiotemporal atlas. Graph convolution is performed on the encrypted privacy spatiotemporal graph, node-level encrypted graph structure features are preserved according to the graph structure node identifiers, and long short-term memory network operations are performed on the node-level encrypted graph structure features arranged by time information to generate the encrypted temporal features of the corresponding graph structure nodes.

[0010] As a preferred embodiment of the privacy-preserving computation-based method for predicting safety hazards in underground pipelines according to the present invention, the step of fusing encrypted graph structural features and encrypted temporal features through a spatiotemporal feature fusion method to generate spatiotemporal risk data is as follows: By splicing and fusing the node-level encrypted graph structure features and encryption time sequence features of the same graph structure node under the same time information, a preliminary encryption risk feature vector of the corresponding graph structure node is generated. The initial encrypted risk feature vector is input into the risk scoring function to obtain the encrypted risk value of the graph structure node under the corresponding time information. The encrypted risk values ​​are then integrated according to the time information to generate spatiotemporal risk data.

[0011] As a preferred embodiment of the privacy-preserving computation-based method for predicting safety hazards in underground pipelines according to the present invention, the specific steps for constructing the hazard scenario generation model are as follows: A potential hazard scenario generation model is constructed based on a generative adversarial network architecture; The hidden danger scenario generation model includes a generation network component and a discrimination network component.

[0012] As a preferred embodiment of the privacy-preserving computation-based method for predicting safety hazards in underground pipelines according to the present invention, the steps of inputting spatiotemporal risk data into a hazard scenario generation model, performing feature combination, Gaussian noise perturbation, and transposed convolution expansion operations on the spatiotemporal risk data to generate a hazard scenario sample set are as follows. The spatiotemporal risk data is input into the hazard scenario generation model. The model performs feature combination and Gaussian noise perturbation on the spatiotemporal risk data through the generation network component to generate basic hazard scenarios. The authenticity of basic hazard scenarios is scored by a discriminative network component. The scoring threshold is determined based on the authenticity score distribution of real basic hazard scenarios. Basic hazard scenarios with authenticity scores not lower than the scoring threshold are selected, and upsampling and transposed convolution are used to perform expansion operations to generate a hazard scenario sample set.

[0013] As a preferred embodiment of the privacy-preserving computation-based method for predicting underground pipeline safety hazards according to the present invention, the specific steps for constructing a pipeline safety hazard prediction model in a privacy-preserving computation environment using spatiotemporal risk data and a sample set of hazard scenarios are as follows. A pipeline safety hazard prediction model is constructed based on a gradient boosting decision tree architecture. Training samples are formed using spatiotemporal risk data, a sample set of potential hazard scenarios, and corresponding target output values. Based on the training samples, the partitioning features, splitting values, and output values ​​of each internal node in the gradient boosting decision tree are determined, and a configured pipeline safety hazard prediction model is generated.

[0014] As a preferred embodiment of the privacy-preserving computation-based underground pipeline safety hazard prediction method of the present invention, the specific steps for inputting real-time node-level spatiotemporal risk data into the pipeline safety hazard prediction model to generate safety risk level identifiers for corresponding graph structure nodes are as follows. Collect real-time pipeline operation data and process the real-time pipeline operation data according to the same processing flow used to generate spatiotemporal risk data to generate real-time node-level spatiotemporal risk data. Real-time node-level spatiotemporal risk data is input into the configured pipeline safety hazard prediction model. The risk probability distribution is calculated through gradient boosting decision tree, and the safety risk level identifier of the corresponding graph structure node is output.

[0015] As a preferred embodiment of the privacy-preserving computation-based method for predicting underground pipeline safety hazards according to the present invention, the specific steps for generating corresponding early warning information and operation and maintenance decisions based on different levels of safety risk identification are as follows: The safety risk level identifier is associated and matched with the spatial attributes in the pipeline structure data to generate pipeline spatial data with risk labels; By fusing pipeline space data with risk labels with environmental data, a risk dataset is generated. The risk dataset is then color-coded and labeled with symbols to generate a safety hazard risk distribution map. Based on the preset early warning rules, early warning information corresponding to different safety risk levels is generated, and operation and maintenance decisions are generated by combining the operation and maintenance rule table formed by pipeline operation and maintenance procedures, inspection system, maintenance methods and emergency plans.

[0016] The beneficial effects of this invention are as follows: By inputting a privacy-preserving spatiotemporal atlas into a privacy-protected computing environment to perform graph convolution operations and long short-term memory network encoding, deep feature fusion of multi-source heterogeneous data is achieved, transforming scattered pipeline operation data into unified encrypted risk indicators, improving the accuracy and real-time performance of the prediction model in identifying hidden risks, while ensuring that the original data is encrypted throughout the process; by constructing a risk scenario generation model, performing feature combination, Gaussian noise perturbation, and transposed convolution expansion operations on spatiotemporal risk data, diverse high-risk scenario samples are dynamically generated, solving the problem of weak model generalization ability caused by the scarcity of real high-risk samples, improving the sensitivity and reliability of early warning for sudden or low-probability risks, and avoiding early warning lag. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0018] Figure 1 This is a flowchart of a privacy-based computation method for predicting safety hazards in underground pipelines.

[0019] Figure 2 A flowchart for generating encryption features.

[0020] Figure 3 A flowchart for generating a sample set of potential hazard scenarios.

[0021] Figure 4 This is a flowchart for the distribution of safety hazards and early warning decisions. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for predicting safety hazards in underground pipelines based on privacy computing, including the following steps: S1. Collect pipeline structure data, pipeline operation data, and environmental data, and perform standardized desensitization and fusion through differential privacy mechanism to generate pipeline privacy data; apply graph structure construction operation to pipeline privacy data to generate privacy spatiotemporal atlas.

[0026] S1.1 Collect pipeline structure data, pipeline operation data, and environmental data, and perform format standardization and missing value processing to obtain preprocessed pipeline structure data, pipeline operation data, and environmental data.

[0027] It should be noted that pipeline structure data includes pipeline identification, pipeline location, depth, material, diameter, and burial method, obtained through underground pipeline surveying; pipeline operation data includes data during pipeline operation, such as flow rate, pressure, temperature, gas or liquid flow velocity, and pipeline aging level, collected in real time by sensor equipment installed on the pipeline; environmental data includes meteorological data, geological conditions, soil properties, groundwater level, terrain type, building type, population density classification, identification of important facilities, and spatial distance between important facilities and pipelines in the area where underground pipelines are located, obtained through meteorological stations, environmental monitoring equipment, geographic information data, planning facility data, and population statistics.

[0028] The formats of pipeline structure data, pipeline operation data, and environmental data collected from different sources are standardized. Specifically: spatial data in pipeline structure data is standardized using CGCS2000 coordinates to avoid data deviations caused by inconsistencies in coordinate systems. Spatial data includes the location and depth of the pipeline; the material in pipeline structure data is uniformly labeled as PE or steel, and the pipe diameter adopts the standard unit format of DN300; the units of meteorological data in environmental data need to be standardized to degrees Celsius, and precipitation to millimeters, etc.; for various sensor data in pipeline operation data, such as pressure values, flow rates, and temperatures, the units should be standardized to standard units such as Pa, L / s, degrees Celsius, etc.

[0029] When handling missing values ​​in pipeline structure data, if spatial data such as pipeline location and depth are missing, the pipeline identifier and segment sequence to which the missing record belongs are located. Within the same pipeline, the nearest known spatial point (containing at least planar coordinates and burial depth) is searched forward and backward along the topological sequence. The distance along the line from the missing point to the preceding and following known spatial points is calculated (obtained by summing the mileage markers corresponding to the segment sequence) and the ratio is calculated. Linear interpolation of the planar coordinates and burial depth is then performed using this ratio. If attribute data such as material and pipe diameter are missing, the missing values ​​are filled in based on the attributes of adjacent pipe segments: if the upstream phase of the missing pipe segment... If the attribute values ​​of adjacent and downstream pipe segments are consistent, the missing value will be assigned the same value as the upstream and downstream pipe segments. If the attributes of upstream and downstream pipe segments are inconsistent, the attribute of the pipe segment with the closest spatial distance will be assigned to the missing value (the spatial distance is the Euclidean distance from the geometric center of the missing pipe segment to the geometric center of the adjacent pipe segment). Since pipeline operation data and environmental data usually change over time and the changes are relatively stable, linear interpolation can be used to fill in the missing data in pipeline operation data and environmental data, generating preprocessed pipeline structure data, pipeline operation data, and environmental data. The expression for linear interpolation is as follows: ; in, This represents the interpolated value; This represents the first known data value; This represents the second known data value; Indicates the first known time point; This indicates the second known time point; This indicates the time point where interpolation is required.

[0030] S1.2 Apply differential privacy mechanism to the preprocessed pipeline structure data, pipeline operation data and environmental data for standardized desensitization processing to generate privacy pipeline structure data, privacy pipeline operation data and privacy environment data.

[0031] It should be noted that the preprocessed pipeline structure data, pipeline operation data, and environmental data are normalized to the [0,1] interval to form normalized pipeline structure data, pipeline operation data, and environmental data, in order to avoid the impact of dimensional differences on privacy protection. Each data record is defined as "a pipeline multi-source field record of a certain pipeline identifier within a certain time window". All pipeline multi-source field records are integrated to generate a pipeline multi-source field set. Adjacent pipeline multi-source field sets are defined as dataset pairs that differ only in one data record. That is, only one pipeline multi-source field record is allowed to be added or deleted between two pipeline multi-source field sets, thereby limiting the adjacency relationship of differential privacy.

[0032] For each numerical field (such as pipeline pressure value or coordinate information) in the normalized pipeline structure data, pipeline operation data, and environmental data, noise is generated based on the Laplace mechanism and added to the normalized pipeline structure data, pipeline operation data, and environmental data to protect individual privacy and generate privacy-preserving pipeline structure data, privacy-preserving pipeline operation data, and privacy-preserving environmental data.

[0033] The expression for generating noise using the Laplace distribution is: ; ; in, Indicates noise; Indicates sensitivity; Indicates privacy budget parameters; This indicates that the noise randomly generated from the Laplace distribution has a mean of 0 and a scale parameter of 1; This represents a set of multi-source field sets for adjacent pipelines; express and For two adjacent pipeline multi-source field sets, only one pipeline multi-source field record is allowed to be added or deleted between the two pipeline multi-source field sets; and This represents the query result after executing a query function on two pipeline multi-source field sets; This indicates the difference between the query results corresponding to two adjacent pipeline multi-source field sets.

[0034] Among them, privacy budget parameters It is a positive number greater than 0; in this embodiment, the data involving the precise location of the pipeline, its burial depth, and the distance to important facilities are taken as... Data involving pressure, flow rate, temperature, and valve status are collected. Data acquisition involving publicly available meteorological data This allows data fields with different levels of sensitivity to be protected against noise of varying strengths.

[0035] Differential privacy is a method used to protect data privacy, ensuring that even if an attacker obtains partial data or query results, they cannot infer whether a single data point exists in the dataset or the specific value of the data point. Privacy protection is achieved by adding noise to the query results, and the magnitude of the noise is determined by the sensitivity of the query results and the privacy budget parameter.

[0036] Privacy budget parameters refer to the level of information disclosure allowed during the query process. Privacy budget parameters control the intensity of added noise and determine the level of privacy protection in the query results. Privacy budget parameters should be set according to the privacy needs, data sensitivity, and query frequency in actual applications. The smaller the privacy budget parameter, the stronger the privacy protection.

[0037] S1.3. Based on pipeline identification and corresponding time information, integrate privacy pipeline structure data, privacy pipeline operation data and privacy environment data to generate pipeline privacy data.

[0038] It should be noted that by using pipeline identification and time information as the association fields for data fusion, privacy pipeline structure data and privacy pipeline operation data are merged to generate privacy pipeline structure operation data. This ensures that the pipeline's various attributes and status data are accurately matched through pipeline identification and time information, guaranteeing the consistency of the time sequence between data. Furthermore, privacy pipeline structure operation data is fused with privacy environment data, using pipeline identification and time information as association fields. This ensures that the corresponding time points of the privacy environment data are consistent with the privacy pipeline structure operation data, forming pipeline privacy data. This ensures that all data are accurately associated according to pipeline identification and time information, while preserving the privacy protection characteristics of the data.

[0039] S1.4. Assign each pipeline in the pipeline privacy data to a graph structure node, and assign the connection relationships and spatial proximity relationships in the pipeline privacy data to graph structure edges to generate a privacy graph structure.

[0040] It should be noted that each pipeline is identified by its pipeline identifier, and a unique graph structure node is assigned to each pipeline. The graph structure node represents the pipeline or pipe segment entity in the underground pipeline network, ensuring that the pipeline identifier of each pipeline corresponds to the graph structure node. Attribute data related to the pipeline, such as pipe diameter, material type, and spatial coordinates, are extracted from the pipeline privacy data and attached to each corresponding graph structure node.

[0041] Connections are determined by the physical topology of pipelines, and these connections are represented by pipeline identifiers, indicating which pipelines are physically connected through specific interfaces, valves, or other facilities. Graph edges are generated based on the connections between these pipeline identifiers, representing the physical connections between pipelines. Spatial proximity is determined by calculating the Euclidean distance between graph nodes. If the Euclidean distance between two graph nodes is less than a distance threshold, these two nodes are considered spatially adjacent, and a graph edge is created to represent this spatial proximity. Combining graph nodes and graph edges into a graph data structure results in a privacy graph structure.

[0042] It should also be noted that the distance threshold is defined based on the actual application needs and data characteristics in underground pipeline safety management. According to the accuracy requirements of pipeline network data mapping and the spatial distribution characteristics of pipelines, the distribution pattern of Euclidean distances between pipelines is calculated by analyzing historical pipeline data to understand the spatial clustering characteristics of pipeline clusters. The distance threshold is then defined in conjunction with the physical properties of the pipelines themselves and safety regulations. The range of the distance threshold is 5 to 30 meters, preferably 10 meters. If the distance threshold is greater than 30 meters, it will expand the identification range of spatial proximity relationships, leading to an increase in the number of graph structure edges and increasing the computational burden and storage overhead of graph convolution operations. If the distance threshold is less than 5 meters, pipelines with actual spatial influence relationships may not be included in the proximity relationships, resulting in incomplete early warning coverage.

[0043] S1.5 Integrate the privacy graph structures corresponding to different time information to form a privacy spatiotemporal graph atlas.

[0044] It should be noted that the privacy graph structure represents pipeline data at different points in time, and each privacy graph structure corresponds to the pipeline state and relationship data at a specific point in time. Based on the time information of the pipeline, all privacy graph structures are arranged in ascending order. By adding temporal edges representing state transmission between privacy graph structures with the same pipeline node at different times, privacy graph structures with different times are connected to form a privacy spatiotemporal graph atlas.

[0045] S2. Input the privacy spatiotemporal graph atlas into the privacy-preserving computing environment, and perform graph feature extraction and temporal feature encoding on the privacy spatiotemporal graph atlas to generate encrypted graph structure features and encrypted temporal features; fuse the encrypted graph structure features and encrypted temporal features through a spatiotemporal feature fusion method to generate spatiotemporal risk data.

[0046] S2.1. A privacy-preserving computing environment is formed based on a multi-party secure computation protocol and homomorphic encryption. The privacy-preserving spatiotemporal atlas is input into the privacy-preserving computing environment and homomorphically encrypted to generate an encrypted privacy-preserving spatiotemporal atlas.

[0047] It should be noted that the privacy-preserving computing environment consists of computing nodes of at least two data participants, and the homomorphic encryption parameters and multi-party secure computing parameters are collaboratively configured during the environment initialization phase: a homomorphic encryption public key is generated and distributed among the data participants, while decryption capabilities are distributed to each data participant in a threshold manner to prevent any single data participant from decrypting alone; multi-party secure computing collaborative rules for multiplication and nonlinear evaluation are established, enabling data participants to perform multiplication and polynomial evaluation on the "ciphertext-represented numerical value" without recovering the plaintext; two types of data exchange interfaces are defined between the homomorphic domain and the multi-party secure computing domain, including a ciphertext-to-share interface and a share-to-ciphertext interface. The ciphertext-to-share interface is used to convert the input of the homomorphic encrypted state into an addition share held by each data participant under threshold decryption collaboration, so that each data participant only obtains its own share and does not obtain the plaintext. The share-to-ciphertext interface is used to collaboratively re-encrypt the shares held by each data participant into homomorphic ciphertext after the multi-party secure computing is completed, so as to continue subsequent homomorphic domain operations.

[0048] The privacy-preserving spatiotemporal atlas is input into a privacy-preserving computing environment. Within this environment, the atlas is serialized, and the node attributes and adjacency relationships of the graph structure are split into data fragments. These fragments are then distributed to each participating node via a multi-party secure computation protocol. Each participating node uses a public key of a homomorphic encryption algorithm to encrypt the data fragments, generating ciphertext fragments. The homomorphic encryption algorithm ensures that the data fragments remain ciphertext during computation. Finally, the multi-party secure computation protocol aggregates all ciphertext fragments, reconstructing an encrypted privacy-preserving spatiotemporal atlas.

[0049] S2.2 Perform graph convolution operation on the encrypted privacy spatiotemporal graph to generate encrypted graph structure features; perform long short-term memory network operation on the encrypted graph structure features to generate encrypted temporal features.

[0050] It should be noted that graph convolution operations are performed on the encrypted privacy-preserving spatiotemporal graph using a graph convolutional network algorithm. Specifically, during the graph convolution operation, the adjacency matrix and node feature matrix of the encrypted privacy-preserving spatiotemporal graph are required. The adjacency matrix represents the connection relationship between nodes in the graph structure, and the node feature matrix represents the attributes of each node in the graph structure. The adjacency matrix and node feature matrix are transmitted and stored in encrypted form to ensure that no private data is leaked. The weight matrix of the graph convolutional layer is trained using historical pipeline hazard samples. The historical pipeline hazard samples include historical privacy-preserving spatiotemporal graphs and corresponding risk labels. The corresponding risk labels are determined by accident files, maintenance records, hazard investigation records, and manual review results. After training, the weight matrix of the graph convolutional layer is fixed and loaded into the privacy-preserving computing environment.

[0051] During graph convolution operations, linear transformations of ciphertext-plaintext multiplication and ciphertext addition are performed on the node feature matrix and the weight matrix of the graph convolution layer in the homomorphic domain to generate an encrypted linear intermediate feature matrix. The adjacency matrix and the encrypted linear intermediate feature matrix are then converted into additive shares held by multiple parties through a ciphertext-to-share interface. In the multi-party secure computation domain, the adjacency weights and feature components are aggregated by row-neighbor weighted summation to obtain the aggregated node share features. After a nonlinear transformation using a polynomial approximation, the aggregated node share features are re-encrypted back into the homomorphic domain through a share-to-ciphertext interface to form an encrypted node feature matrix. This process is iteratively performed on multiple graph convolution layers to generate an encrypted node feature matrix containing the representations of all nodes. The feature rows corresponding to each graph structure node in the encrypted node feature matrix are retained according to the graph structure node identifier. The encrypted node features of each graph structure node under the corresponding time information are used as node-level encrypted graph structure features. The expression for updating node features in graph convolution operations is as follows: ; ; ; in, The graph structure node is in the first position. The node features output by each graph convolutional layer, i.e., the new node features; Represents the ReLU nonlinear activation function; Represents the adjacency matrix; The graph structure node is in the first position. The node features output by the graph convolutional layer; This represents the weight matrix of the graph convolutional layer; This represents the hierarchical index of the graph volume layer; Represents the normalized adjacency matrix; This represents the adjacency matrix with self-loops after adding self-connections; express a diagonal matrix; express The diagonal element square root reciprocal matrix; This represents the identity matrix, with the same dimensions as the adjacency matrix, and is used to add self-connections to each adjacency matrix.

[0052] The node-level encrypted graph structure features are sorted according to the graph structure node identifier and time information to form a time series of encrypted graph features corresponding to each graph structure node. Each element in the time series represents the node-level encrypted graph structure feature of the same graph structure node under the corresponding time information.

[0053] The weight matrices and bias vectors corresponding to the forget gate, input gate, output gate, and candidate cell states in the Long Short-Term Memory (LSTM) network are obtained by training with historical pipeline time-series samples and corresponding risk labels. The historical pipeline time-series samples consist of node-level encrypted graph structure features of the same graph structure nodes under continuous time information. The corresponding risk labels are determined by accident files, maintenance records, hidden danger investigation records, and manual review results. After training, the weight matrices and bias vectors are fixed and loaded into a privacy-preserving computing environment.

[0054] For each time point, linear intermediate quantities of the forget gate, input gate, candidate cell state, and output gate are calculated in the homomorphic domain using ciphertext-plaintext multiplication and ciphertext addition. These linear intermediate quantities are then converted into addition shares held by each party through the ciphertext-to-shares interface. In the multi-party secure computation domain, polynomial approximations of the sigmoid function are performed on the forget gate, input gate, and output gate, and polynomial approximations of the tanh function required for updating the candidate cell state and cell state are also performed. Simultaneously, element-wise dot product and element-wise addition between the gate value and the cell state are performed to update the cell state and hidden state. After the update is completed, the hidden state is re-encrypted back into the homomorphic domain through the share-to-ciphertext interface as the input for the next time point.

[0055] After completing all time point iterations, the hidden state of the last time point is used as the encrypted temporal feature of the corresponding graph structure node.

[0056] S2.3. The node-level encrypted graph structure features and encrypted temporal features of the same graph structure node under the same time information are spliced ​​and fused to generate the preliminary encryption risk feature vector of the corresponding graph structure node; the preliminary encryption risk feature vector is input into the risk scoring function to obtain the encryption risk value of the graph structure node under the corresponding time information, and the encryption risk value is integrated according to the time information to generate spatiotemporal risk data.

[0057] It should be noted that the node-level encrypted graph structure features and the corresponding encrypted temporal features of the graph structure nodes are fused by concatenation: for the same graph structure node under the same time information, the node-level encrypted graph structure features are arranged first, and then the encrypted temporal features are arranged; within the same type of features, they are arranged in ascending order according to the dimension number when the features were generated and then concatenated with ciphertext to form the preliminary encryption risk feature vector of the graph structure node under the corresponding time information; the concatenation process is performed by homomorphic encryption algorithm to ensure that all encrypted features are still in an encrypted state.

[0058] The initial encryption risk feature vector is input into the risk scoring function, which calculates the encryption risk value for each graph node. The encryption risk values ​​of each graph node at each time point are integrated to form a time-series dataset. The time-series datasets from all time points are combined to construct spatiotemporal risk data reflecting the nodes' risks across different time dimensions. The expression for generating the encryption risk value using the risk scoring function is as follows: ; in, Indicates the encryption risk value; This represents the transpose of the weight vector of the risk scoring function; This represents the matrix transpose operation; This represents the initial encryption risk feature vector; This represents the bias term of the risk scoring function; This represents the Sigmoid activation function.

[0059] and The parameters are obtained from training with historical hidden danger samples and can also be used as joint training parameters for graph convolutional layers and long short-term memory network output layers. After training, they are fixed and loaded into a privacy-preserving computing environment.

[0060] S3. Construct a hidden danger scenario generation model, input spatiotemporal risk data into the hidden danger scenario generation model, perform feature combination, Gaussian noise perturbation and transposed convolution expansion operations on the spatiotemporal risk data to generate a hidden danger scenario sample set.

[0061] S3.1 Construct a hidden danger scenario generation model based on generative adversarial network architecture. The hidden danger scenario generation model includes a generation network component and a discrimination network component.

[0062] It should be noted that the generation network component and the discrimination network component are initialized within a privacy-preserving computing environment. The generation network component adopts a multilayer perceptron structure, with an input layer for receiving spatiotemporal risk data, and multiple fully connected layers and nonlinear activation layers are connected in series to map low-dimensional encrypted features layer by layer into synthetic hazard scene data consistent with the dimension of the hazard scene. The discrimination network component adopts a convolutional neural network structure, with an input layer for receiving mixed inputs of real hazard scene samples and generated hazard scene samples, and multiple levels of convolutional layers, pooling layers and fully connected layers are connected in sequence. The output layer provides a probability score representing the authenticity of the scene.

[0063] After the generation and discrimination network components have been initialized and their hierarchical connections completed, adversarial training is performed on them using anonymized historical hazard scenario samples as training samples. The historical hazard scenario samples are determined by accident files, maintenance records, hazard investigation records, and manual review results, and the anonymization method is consistent with the differential privacy processing method. After training, the network parameters of the generation and discrimination network components are fixed, and the fixed generation and discrimination network components are loaded into the privacy-preserving computing environment.

[0064] In a privacy-preserving computing environment, the generation network component is used to perform feature combination and Gaussian noise perturbation on spatiotemporal risk data under encryption, and the discrimination network component is used to score the authenticity of basic hidden danger scenarios under encryption; the network parameters of the generation network component and the discrimination network component are not updated at this stage.

[0065] It should also be noted that in the privacy-preserving computing environment, the generator network component and the discriminator network component only perform encrypted inference operations with fixed parameters, and do not perform loss function calculation, backpropagation, or network parameter updates. The fully connected layers, convolutional layers, and transposed convolutional layers in the generator network component and the discriminator network component are all converted into multiply-accumulate operations. Among them, the linear multiply-accumulate operations related to the fixed model parameters are completed in the homomorphic domain through ciphertext-plaintext multiplication and ciphertext addition. The activation function and pooling function are converted to the multi-party secure computing domain through the ciphertext-to-shares interface and executed. The corresponding outputs are obtained by using polynomial approximate activation and secure pooling. The output results are then re-encrypted back to the homomorphic domain through the share-to-ciphertext interface.

[0066] S3.2 Input the spatiotemporal risk data into the hidden danger scenario generation model, and perform feature combination and Gaussian noise perturbation operation on the spatiotemporal risk data through the generation network component to generate basic hidden danger scenarios.

[0067] It should be noted that the spatiotemporal risk data is used as the input layer of the generator network component of the hidden danger scenario generation model, which then divides the encrypted feature vector into multiple sub-vectors representing spatial risk distribution, temporal evolution trend, and pipeline attribute correlation. These sub-vectors are then fed into the first fully connected layer. In the first fully connected layer, a weighted combination of each sub-vector is achieved through weighted operations with weight vectors and bias vectors, resulting in an intermediate combined feature vector that integrates spatial, temporal, and attribute information. This intermediate combined feature vector is then used as the input to the second fully connected layer. In the second fully connected layer, a secure pseudo-random number generator is invoked to generate a random perturbation vector of the same dimension as the intermediate combined feature vector according to a Gaussian distribution. Homomorphic addition is then performed on the random perturbation vector. The encrypted Gaussian perturbation vector is obtained. The mean of the Gaussian distribution is 0, and the variance is 10% of the standard deviation of each dimension of the spatiotemporal risk feature vector in the training samples, so that the amplitude of the Gaussian noise perturbation matches the data fluctuation range of the training samples. With the support of homomorphic encryption, the intermediate combined feature vector and the encrypted Gaussian perturbation vector are subjected to element-wise addition perturbation operation to generate the hazard scene feature vector after feature combination and Gaussian noise perturbation. The hazard scene feature vector is used as input, and the one-dimensional hazard scene feature vector is converted into a multi-dimensional scene representation through a fully connected layer inside the generation network component. The basic hazard scene representing the risk status of pipelines and surrounding environment in the target time period is obtained in the output layer of the generation network component, and the basic hazard scene is kept encrypted throughout the operation link.

[0068] The basic hazard scenario includes at least pipeline identification, target time period, spatial risk distribution characteristics, temporal evolution characteristics, pipeline attribute association characteristics, and environmental association characteristics. Among them, spatial risk distribution characteristics are used to represent the distribution of risks in the spatial location of pipelines, temporal evolution characteristics are used to represent the changing trend of risks within a continuous time window, pipeline attribute association characteristics are used to represent the association between risks and pipe diameter, material, burial depth, and operating status, and environmental association characteristics are used to represent the association between risks and surrounding geology, groundwater level, population density, and important facilities.

[0069] S3.3. The authenticity of basic hidden danger scenarios is scored by the discriminant network component. The scoring threshold is determined according to the authenticity score distribution of the real basic hidden danger scenarios. Basic hidden danger scenarios with authenticity scores not lower than the scoring threshold are selected, and upsampling and transposed convolution are used to perform expansion operations to generate a sample set of hidden danger scenarios.

[0070] It should be noted that the basic hidden danger scene is input into the input layer of the discriminant network component in the form of an encrypted feature map. The input layer of the discriminant network component feeds the basic hidden danger scene feature map into the first convolutional layer. In the first convolutional layer and subsequent multi-level convolutional layers, convolution and non-linear activation operations are performed on the basic hidden danger scene feature map in sequence. Downsampling or pooling is performed in the intermediate layers to extract multi-scale spatial features. The feature map output by the last convolutional layer is flattened into a one-dimensional feature vector and fed into a fully connected layer group. In the fully connected layer group, the intermediate discriminant features of the discriminant network component are obtained by weighted operation with the weight vector and the bias vector. In the output layer of the discriminant network component, the authenticity score of the basic hidden danger scene with a value between 0 and 1 is calculated by the activation function.

[0071] The authenticity scores of basic hazard scenarios are converted into score shares held by each data participant through a ciphertext-to-share interface, and the score threshold is converted into a threshold share. Within the multi-party secure computation domain, the score shares and threshold shares are compared to obtain a comparison result indicating whether the authenticity score is not lower than the score threshold. Based on the comparison result, only basic hazard scenarios with authenticity scores not lower than the score threshold are retained and organized into a set of basic hazard scenarios to be expanded in batches. Each basic hazard scenario in the set of basic hazard scenarios to be expanded is used as the input for the expansion operation. In the expansion operation, a spatial upsampling operator is applied to the basic hazard scenarios, and interpolation upsampling is performed in the row and column dimensions to improve spatial resolution. The magnified basic hazard scene is obtained, and then the magnified basic hazard scene is used as input to perform transposed convolution operation to generate the first extended hazard scene. By changing the upsampling ratio, transposed convolution kernel parameters, and repeatedly performing the upsampling and transposed convolution combination operation, multiple extended hazard scene feature maps with differences in hazard area range, boundary shape, and local details are generated for the same basic hazard scene. All extended hazard scenes obtained from the same basic hazard scene are collected and formed into the hazard scene sample corresponding to the current basic hazard scene. The extension step is performed sequentially on all basic hazard scenes in the set of basic hazard scenes to be extended, and the hazard scene samples generated for each basic hazard scene are uniformly summarized into a hazard scene sample set.

[0072] It should also be noted that a verification sample set of basic hazard scenarios with completed labeling is selected. The verification sample set contains basic hazard scenarios labeled as real. The verification sample set is input into the discriminant network component one by one, and the authenticity score corresponding to each sample is recorded. A set of real sample score sequences is obtained. The real sample score sequences are sorted in ascending order. The authenticity scores corresponding to the real basic hazard scenarios are arranged in ascending order. The score value corresponding to the 20th position after sorting is used as the score threshold. Basic hazard scenarios with scores lower than the lower range of real sample scores are eliminated, while most basic hazard scenarios with real scene characteristics are retained. An example value of 0.7 is used for the score threshold. If the score threshold is greater than 0.7, the screening conditions will be too strict, resulting in a decrease in diversity and fewer covered scenario types. If the score threshold is less than 0.7, the screening conditions will be too lenient, resulting in an increase in the number of samples.

[0073] Samples were drawn from the existing accident files, maintenance records, and hazard investigation records of pipeline operating units. For scenarios that were clearly identified as having safety hazards such as corrosion, leakage, support instability, and external damage risks by on-site inspections, third-party testing, or accident investigation reports within a certain period of time, and for which there were subsequent disposal measures or accident records, the corresponding spatiotemporal risk data and pipeline spatial layout were compiled into "real basic hazard scenarios" and labeled "real".

[0074] This solution scores the authenticity of basic hazard scenarios and filters them using a scoring threshold. Then, it performs upsampling and transposed convolution expansion on the filtered basic hazard scenarios. This controls the quality of scenarios entering the expansion stage and constructs multiple hazard manifestations with different details for each basic scenario. This results in more credible, diverse, and numerous hazard scenario samples, thereby improving the accuracy of subsequent underground pipeline safety hazard prediction while ensuring data privacy.

[0075] S4. In a privacy-preserving computing environment, a pipeline safety hazard prediction model is constructed using spatiotemporal risk data and a sample set of potential hazard scenarios. Real-time pipeline operation data is collected and real-time node-level spatiotemporal risk data is generated according to the same processing flow as the spatiotemporal risk data. The real-time node-level spatiotemporal risk data is then input into the pipeline safety hazard prediction model to generate the safety risk level identifier of the corresponding graph structure node.

[0076] S4.1 Construct a pipeline safety hazard prediction model based on gradient boosting decision tree architecture. Use spatiotemporal risk data, hazard scenario sample set and corresponding target output values ​​to form training samples. Based on the training samples, determine the partitioning features, split values ​​and output values ​​of each internal node in the gradient boosting decision tree and generate the configured pipeline safety hazard prediction model.

[0077] It should be noted that, based on the types and quantities of spatial risk features, temporal evolution features, and pipeline attribute features contained in the spatiotemporal risk data, the complete feature vector dimension is statistically determined, and this complete feature vector dimension is used as the feature dimension that each decision tree needs to receive as input. The arithmetic mean of the target output values ​​of the training samples is used as the initial constant output term. For each newly added decision tree, the difference between the previous round of comprehensive prediction and the target output value is used as the residual signal, and the newly added decision tree learns the residual signal. The overall prediction result is obtained by summing the initial constant output term and the output results of each decision tree.

[0078] Using spatiotemporal risk data and a sample set of potential hazard scenarios as training samples, each record in the training samples includes a spatiotemporal risk feature vector and a corresponding target output value. The target output value is used to characterize the reference magnitude of the potential hazard risk corresponding to the current record. The target output value is determined by the pipeline operation unit's existing accident files, maintenance records, potential hazard investigation records, and manual review results. For records where no potential hazards are found and no action is required, the target output value is set to 0. For records where a general potential hazard is confirmed by inspection and routine action is taken, the target output value is set to 0.5. For records where an accident, leakage, support instability, external force damage, or emergency action is required, the target output value is set to 1. Different potential hazard levels are mapped to continuous risk reference values ​​in the range of 0 to 1 according to the enterprise's potential hazard classification standards. The squared error is selected as the loss function, and the arithmetic mean of the target output values ​​of all training samples is calculated. The arithmetic mean is written into the initial constant output term, so that the initial prediction given for all training samples before any decision tree is used is equal to the initial constant output term.

[0079] For hazard scenario samples obtained by expanding the hazard scenario generation model, if the expanded sample comes from the same basic hazard scenario, the expanded sample inherits the target output value of the corresponding basic hazard scenario; when the expanded sample is identified as having a higher risk intensity than the corresponding basic hazard scenario after manual review, the target output value is increased to no more than 1; when the expanded sample is identified as having a lower risk intensity than the corresponding basic hazard scenario after manual review, the target output value is decreased to no less than 0.

[0080] Using the initial constant output term as a baseline for prediction, the difference between the current prediction and the target output value of each training sample is calculated. This difference is considered as the residual signal that the first decision tree needs to learn. On the node skeleton of the first decision tree, starting from the root node, each splitting feature and its corresponding splitting value are tried sequentially. Specifically, the values ​​of the training samples in the current node on the splitting feature are sorted in ascending order, and the midpoint between two adjacent different values ​​is taken as the corresponding splitting value. For each combination of feature and splitting point tried, the training samples contained in the current root node are divided into left and right subsets according to the combination. The sum of squares of the residuals in the left and right subsets is calculated, and then the sum of the squares of the left and right subsets is compared with the sum of squares of the residuals of the whole set without splitting. The feature number and splitting value with the largest decrease in the sum of squares of the residuals are selected and written into the reserved field of the current root node, and left and right child nodes are generated.

[0081] The comparison of feature values, comparison of split values, partitioning of left and right subsets, calculation of residual sum of squares, and selection of optimal split conditions corresponding to the above-mentioned splitting features are all completed within the multi-party secure computation domain. Among them, the feature values ​​to be compared and the corresponding split values ​​in the training samples are converted into the shares held by each data participant through the encrypted-to-share interface and then compared. The residuals and residual sum of squares are calculated in the form of shares. After obtaining the optimal split conditions, the corresponding splitting feature number and split value are written into the current decision tree node.

[0082] Repeat the steps of "traversing the partitioning features and splitting values, calculating the sum of squared residuals before and after the split and comparing the reduction, and writing the feature and splitting value that brings the maximum reduction into the current node" for the newly generated left and right child nodes; when a node to be split cannot reduce the sum of squared residuals of the training samples in the node after further splitting, mark the current node to be split as a leaf node and stop splitting downwards; for the parts that have been determined as leaf nodes, collect the training samples contained in each leaf node, calculate the arithmetic mean of the residuals corresponding to the training samples, and write the arithmetic mean into the output value position of the leaf node as the correction amount of the current decision tree on the current leaf node.

[0083] The initial constant output term is summed with the output of the first decision tree to form a new comprehensive prediction. The difference between the new comprehensive prediction and the target output value is calculated again for all training samples. The new difference is used as the residual signal of the next decision tree. The node splitting conditions and leaf node output values ​​of the next decision tree are determined in the same way.

[0084] After each new decision tree is added, the squared error before and after the addition is calculated. When the squared error after adding a new decision tree is no longer less than the squared error before the addition, the addition of new trees is stopped. The initial constant output term, the partition feature number, split value and output value of all internal nodes in each decision tree and the leaf node are uniformly fixed as the internal parameters and decision logic of the gradient boosting decision tree combination structure.

[0085] S4.2 Collect real-time pipeline operation data and process the real-time pipeline operation data according to the same processing flow for generating spatiotemporal risk data to generate real-time node-level spatiotemporal risk data.

[0086] It should be noted that real-time pipeline operation data is collected by pressure sensors, flow sensors, temperature sensors, valve opening acquisition devices, pump operation status acquisition devices, and surrounding third-party construction monitoring devices deployed along the pipeline and in the station. Each original record is accompanied by a collection timestamp, measuring point identifier, and measurement type identifier. Original records from different measuring points within the same collection period are aggregated into a time-stamped real-time data stream in chronological order.

[0087] Using a unified reference clock as a benchmark, a fixed-length time window is defined for the current prediction time. The start and end times of the time window serve as alignment boundaries. Each raw record in the real-time data stream is assigned to its corresponding time window based on its timestamp. Aligned monitoring sequences, such as pressure, flow rate, temperature, valve status, pump status, and alarm event sequences for a specific pipe segment within the current time window, are then generated using pipeline identifiers as keys. After time window alignment, each monitoring sequence is used as input for feature extraction. For continuous quantities such as pressure, flow rate, and temperature, the current... Within a time window, the mean, maximum, minimum, fluctuation amplitude, and slope of change are calculated. For discrete states such as valve opening and pump start / stop, the cumulative duration and percentage of each state are statistically analyzed within the current time window. For discrete events such as leakage alarms and disturbance alarms, the number of events and the minimum interval between adjacent events are statistically analyzed within the current time window. All statistical results are arranged into feature values. Preprocessing such as missing value imputation, dimension unification, and interval normalization is performed on the feature values ​​to compress the statistical results from different physical quantities into a comparable numerical range, forming a real-time pipeline operation feature vector.

[0088] The real-time pipeline operation data, after time window alignment, is correlated with the corresponding pipeline structure data and environmental data to form real-time pipeline multi-source data. Following the same method used to generate pipeline privacy data, the real-time pipeline multi-source data undergoes format unification, missing value handling, normalization, and differential privacy processing to generate real-time pipeline privacy data. Following the same method used to generate privacy spatiotemporal atlases, the real-time pipeline privacy data is constructed into a real-time privacy spatiotemporal atlas. Following the same method used to generate spatiotemporal risk data, the real-time privacy spatiotemporal atlas undergoes node-level graph feature extraction, node-level temporal feature encoding, and node-level risk scoring to generate real-time node-level spatiotemporal risk data, ensuring that the real-time prediction input and the training input of the pipeline safety hazard prediction model are of the same data type.

[0089] S4.3 Input real-time node-level spatiotemporal risk data into the configured pipeline safety hazard prediction model, calculate the risk probability distribution through gradient boosting decision tree, and output the safety risk level identifier of the corresponding graph structure node.

[0090] It should be noted that real-time node-level spatiotemporal risk data is sent to the pipeline safety hazard prediction model in encrypted form. Each feature value in the real-time node-level spatiotemporal risk data is mapped one by one to the partition feature number recorded by the root node and internal nodes of each decision tree. During the gradient boosting decision tree inference process, for the feature value to be compared corresponding to the current internal node, the feature value to be compared in the real-time node-level spatiotemporal risk data is converted into the feature share held by each data participant through the encrypted-to-share interface, and the segmentation value recorded by the current internal node is converted into the segmentation threshold share. In the multi-party secure computation domain, the feature share and segmentation are... The threshold share is compared to obtain the comparison result of the current internal node. Based on the comparison result, it is determined whether to enter the left or right child node, and the process proceeds down the tree structure layer by layer until the leaf node is reached. After reaching the leaf node, the output value of the leaf node is used as the risk contribution of the current decision tree to the real-time node-level spatiotemporal risk data, and the risk contribution is re-encrypted back to the homomorphic domain through the share-to-ciphertext interface. The risk contributions obtained from all decision trees are summed with the initial constant output term to form the comprehensive risk score corresponding to the current time window. The comprehensive risk score is mapped to a risk probability distribution in the interval of 0 to 1 through the Sigmoid function.

[0091] The risk probability distribution is transformed into a corresponding safety risk level identifier based on the upper and lower risk level thresholds. The lower and upper risk level thresholds are adaptively determined using a data-driven approach: a set of verification samples with real risk labels is constructed, and each sample is input into the pipeline safety hazard prediction model to obtain the verification risk probability distribution for each sample. The distribution is incremented from 0 to 1 in a fixed step size (e.g., 0.01) within the range of 0 to 1, resulting in a sequence of multiple candidate thresholds. A candidate threshold sequence is generated. The current candidate threshold is extracted sequentially, and samples with a verification risk probability not lower than the current candidate threshold are classified as high-risk candidates, while those with a verification risk probability lower than the current candidate threshold are classified as non-high-risk candidates. The F1 score corresponding to the current candidate threshold is calculated, and the current candidate threshold with the optimal F1 score is taken as the upper risk level threshold. A second candidate threshold sequence is generated within the range of samples with a verification risk probability distribution lower than the upper risk level threshold, and the F1 score is calculated similarly to determine the lower risk level threshold.

[0092] For each risk probability distribution, interval judgment is performed. When the risk probability is less than the lower risk level threshold, a low risk level label is output. When the risk probability is greater than or equal to the lower risk level threshold but less than the upper risk level threshold, a medium risk level label is output. When the risk probability is greater than or equal to the upper risk level threshold, a high risk level label is output. This allows the three risk levels to be automatically determined by the recognition effect of the verification data, ensuring that the risk level label matches the actual risk recognition capability.

[0093] S5. Spatially associate the safety risk level identifier with the pipeline structure data and combine it with environmental data to form a safety hazard risk distribution map; generate corresponding early warning information and operation and maintenance decisions based on different levels of safety risk level identifiers.

[0094] S5.1. Associate and match the safety risk level identifier with the spatial attributes in the pipeline structure data to generate pipeline spatial data with risk labels.

[0095] It should be noted that the pipeline identifier, time window identifier, and safety risk level identifier are compiled into a risk labeling list. Spatial attribute information is read from the pipeline structure data. The spatial attribute information includes at least the pipeline identifier, the coordinates of the pipeline segment's starting point, the coordinates of the pipeline segment's ending point, the coordinate sequence of inflection points along the pipeline, and the identifier of the station or region to which it belongs. The pipeline identifier is used as the association key to perform a one-to-one match between the risk labeling list and the pipeline structure data. During the matching process, for each record of "pipeline identifier, time window identifier, and safety risk level identifier" in the risk labeling list, the spatial attribute information with the same pipeline identifier in the pipeline structure data is retrieved. The coordinates of the pipeline segment's starting point, the coordinates of the pipeline segment's ending point, and the coordinate sequence of inflection points along the pipeline are combined and stored with the current safety risk level identifier in the same data record. A new field for recording the time window identifier and the safety risk level identifier is added to the attribute table of the spatial attribute information, and the time window identifier and the safety risk level identifier are written into the corresponding field, forming pipeline spatial data with risk labels that contain both geometric coordinate information and safety risk level identifiers at different times.

[0096] S5.2. Integrate the pipeline space data with risk labels with environmental data to generate a risk dataset. Color-code and symbolize the risk dataset to generate a safety hazard risk distribution map.

[0097] It should be noted that the pipeline spatial data with risk labels is used as the spatial basis data. The spatial basis data and environmental data are imported into the same geographic information software. The spatial location overlap relationship is used as the matching condition. Spatial overlay operation is performed on each pipeline spatial data with risk labels to determine which environmental elements the line segments corresponding to the starting point, ending point and inflection point coordinates along the pipeline intersect or fall into. An environmental attribute field is added to the attribute table of spatial attribute information for each pipeline spatial record. The terrain category, building type, population density classification and important facility identification information corresponding to the intersection or inclusion relationship are written into the corresponding fields to obtain a risk dataset that includes pipeline geometric coordinates, time window identifiers, safety risk level identifiers and environmental attribute information.

[0098] In the geographic information software, the color coding rules (green for low-risk level, yellow for medium-risk level, and red for high-risk level) are called according to the safety risk level identifier field in the risk dataset. Each pipeline geometric object is assigned a corresponding color. At the same time, according to the environmental attribute field, different shapes of legend symbols are superimposed on important facilities and densely populated areas for auxiliary labeling. All pipeline objects and environmental elements within the current time window are rendered on the same base map according to color coding and legend symbol labeling, and a safety hazard risk distribution map is output.

[0099] S5.3 Generate early warning information corresponding to different safety risk levels based on preset early warning rules, and generate operation and maintenance decisions by combining the operation and maintenance rule table formed by pipeline operation and maintenance procedures, inspection system, maintenance methods and emergency plans.

[0100] It should be noted that before generating early warning information, the environmental sensitivity is determined based on the environmental attributes in the risk dataset, and the importance of the pipeline is determined based on the pipeline structure data. The environmental sensitivity is determined by population density classification, spatial distance between important facilities and pipelines, building type, groundwater level, and geological conditions. Among these, the environmental sensitivity is determined to be high when the population density is high, the spatial distance between important facilities and pipelines is less than 50 meters, the building type is a school, hospital, transportation hub, or gas station, the groundwater level is higher than the pipeline burial depth, or the geological conditions are soft soil or easily subsidence strata. The importance of the pipeline is determined by the pipe diameter, the transported medium, the service area, and the impact range of historical faults. Among these, the importance of the pipeline is determined to be high when the pipe diameter is not less than DN300, the transported medium is gas or heating medium, the service area includes residential areas or important public facilities, or the impact range of historical faults exceeds adjacent pipe sections within the same station.

[0101] Using the risk dataset as input, the records that need to be processed are filtered by time window identifier and pipeline identifier. For each record, the safety risk level identifier, time window identifier, spatial attribute field and environmental attribute field are read and used as input conditions for early warning determination.

[0102] In the early warning rules, different combinations of safety risk level identifiers, environmental sensitivity, and pipeline importance are configured with corresponding early warning levels and response requirements. For each record in the risk dataset, a matching entry is searched in the early warning rules based on the safety risk level identifier and environmental attributes to obtain the early warning level corresponding to the current record, such as general early warning, key early warning, or emergency early warning, as well as response requirements such as whether a notification needs to be sent, whether a control room prompt is needed, or whether on-site verification needs to be arranged. Information such as pipeline identifier, time window identifier, spatial location, safety risk level identifier, and early warning level are combined and written into the early warning information record list.

[0103] By centralizing the safety risk level indicators of the same pipeline within multiple adjacent time windows in the risk dataset, and comparing the level changes between consecutive time windows, pipeline records that maintain a medium-risk or high-risk level indicator for three or more consecutive time windows, as well as those that rise directly from a low-risk to a high-risk level indicator within a single time window, are identified. Explanatory fields such as "risk remains high" or "risk suddenly increases" are added to the warning information to highlight the risk evolution trend. An operation and maintenance rule table is formed by combining pipeline operation and maintenance procedures, inspection systems, maintenance methods, and emergency plans. This rule table pre-determines the combination of each warning level and environmental sensitivity. For each of the multiple maintenance recommendations, including the recommended range for adjusting operating pressure or flow, the recommended cycle for increasing inspection frequency, whether on-site leak testing or partial excavation verification is required, whether a shutdown and maintenance plan is required, and whether an emergency plan needs to be activated in case of an emergency warning, the corresponding maintenance recommendation is searched in the maintenance rule table using the warning level and environmental attributes as indexes. The found maintenance recommendations are then appended to each record in the warning information record list in the form of a "list of recommended operation items," forming a set of maintenance decision records. The maintenance decision records are then sorted according to the severity of the warning level and the time order.

[0104] According to the company's existing safety management system, hazard classification standards and emergency plans, a unified pre-defined early warning rule is defined. Specifically, several information items that need to be referenced are first selected, such as safety risk level identification, environmental sensitivity and pipeline importance. Then, according to different combinations of information items, each item is discussed to determine "which level of early warning should be considered when this situation occurs, and what actions should be taken at least". The discussion conclusions are then compiled into an early warning rule table. Each row in the early warning rule table clearly states the corresponding safety risk level identification, environmental attribute combination, early warning level and recommended actions.

[0105] In summary, this invention achieves deep feature fusion of multi-source heterogeneous data by inputting a privacy-preserving spatiotemporal atlas into a privacy-protected computing environment to perform graph convolution operations and long short-term memory network encoding. This transforms scattered pipeline operation data into a unified encrypted risk indicator, improving the accuracy and real-time performance of the prediction model in identifying hidden risks, while ensuring end-to-end encryption of the original data. Furthermore, by constructing a risk scenario generation model, this invention performs feature combination, Gaussian noise perturbation, and transposed convolution expansion operations on spatiotemporal risk data to dynamically generate diverse high-risk scenario samples. This solves the problem of weak model generalization ability caused by the scarcity of real high-risk samples, improves the sensitivity and reliability of early warning for sudden or low-probability risks, and avoids early warning lag.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting safety hazards in underground pipelines based on privacy computing, characterized in that: include, Pipeline structure data, pipeline operation data, and environmental data are collected and then standardized, de-identified, and fused using a differential privacy mechanism to generate pipeline privacy data. Perform graph structure construction operations on pipeline privacy data to generate a privacy spatiotemporal atlas; The privacy spatiotemporal atlas is input into a privacy-preserving computing environment based on a multi-party secure computation protocol and homomorphic encryption. Graph feature extraction and temporal feature encoding are performed on the privacy spatiotemporal atlas to generate encrypted graph structural features and encrypted temporal features. The encrypted graph structural features and encrypted temporal features are then fused using a spatiotemporal feature fusion method to generate spatiotemporal risk data. A hazard scenario generation model is constructed, and spatiotemporal risk data is input into the hazard scenario generation model. Feature combination, Gaussian noise perturbation and transposed convolution expansion operations are performed on the spatiotemporal risk data to generate a hazard scenario sample set. In a privacy-preserving computing environment, a pipeline safety hazard prediction model is constructed using spatiotemporal risk data and a sample set of potential hazard scenarios. Real-time pipeline operation data is collected and real-time node-level spatiotemporal risk data is generated according to the same processing flow as the spatiotemporal risk data. The real-time node-level spatiotemporal risk data is then input into the pipeline safety hazard prediction model to generate the safety risk level identifier of the corresponding graph structure node. The safety risk level identifier is spatially correlated with pipeline structure data, and combined with environmental data to form a safety hazard risk distribution map; corresponding early warning information and operation and maintenance decisions are generated based on different levels of safety risk identifiers.

2. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 1, characterized in that: The collected pipeline structure data, pipeline operation data, and environmental data are standardized, de-identified, and fused using a differential privacy mechanism to generate pipeline privacy data. The specific steps are as follows: Collect pipeline structure data, pipeline operation data, and environmental data, and perform format standardization and missing value processing to obtain preprocessed pipeline structure data, pipeline operation data, and environmental data; The preprocessed pipeline structure data, pipeline operation data, and environmental data are standardized and anonymized using a differential privacy mechanism to generate privacy-preserving pipeline structure data, privacy-preserving pipeline operation data, and privacy-preserving environmental data. Based on pipeline identification and corresponding time information, privacy pipeline structure data, privacy pipeline operation data, and privacy environment data are integrated to generate pipeline privacy data.

3. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 2, characterized in that: The steps for performing graph structure construction on pipeline privacy data to generate a privacy spatiotemporal graph atlas are as follows. Each pipeline in the pipeline privacy data is mapped to a graph structure node, and the connection relationships and spatial proximity relationships in the pipeline privacy data are mapped to graph structure edges to generate a privacy graph structure. By integrating the privacy graph structures corresponding to information from different times, a privacy spatiotemporal atlas is formed.

4. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 3, characterized in that: The process involves inputting the privacy-preserving spatiotemporal graph atlas into the privacy-preserving computing environment, and performing graph feature extraction and temporal feature encoding on the atlas to generate encrypted graph structure features and encrypted temporal features. The specific steps are as follows: A privacy-preserving computing environment is formed based on a multi-party secure computation protocol and homomorphic encryption. The privacy-preserving spatiotemporal atlas is input into the privacy-preserving computing environment and homomorphically encrypted to generate an encrypted privacy-preserving spatiotemporal atlas. Graph convolution is performed on the encrypted privacy spatiotemporal graph, node-level encrypted graph structure features are preserved according to the graph structure node identifiers, and long short-term memory network operations are performed on the node-level encrypted graph structure features arranged by time information to generate the encrypted temporal features of the corresponding graph structure nodes.

5. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 4, characterized in that: The method of fusing encrypted graph structural features and encrypted temporal features through spatiotemporal feature fusion to generate spatiotemporal risk data is described in the following steps. By splicing and fusing the node-level encrypted graph structure features and encryption time sequence features of the same graph structure node under the same time information, a preliminary encryption risk feature vector of the corresponding graph structure node is generated. The initial encrypted risk feature vector is input into the risk scoring function to obtain the encrypted risk value of the graph structure node under the corresponding time information. The encrypted risk values ​​are then integrated according to the time information to generate spatiotemporal risk data.

6. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 5, characterized in that: The specific steps for constructing the potential hazard scenario generation model are as follows: A potential hazard scenario generation model is constructed based on a generative adversarial network architecture; The hidden danger scenario generation model includes a generation network component and a discrimination network component.

7. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 6, characterized in that: The process involves inputting spatiotemporal risk data into the hazard scenario generation model, performing feature combination, Gaussian noise perturbation, and transposed convolution expansion operations on the spatiotemporal risk data to generate a hazard scenario sample set. The specific steps are as follows: The spatiotemporal risk data is input into the hazard scenario generation model. The model performs feature combination and Gaussian noise perturbation on the spatiotemporal risk data through the generation network component to generate basic hazard scenarios. The authenticity of basic hazard scenarios is scored by a discriminative network component. The scoring threshold is determined based on the authenticity score distribution of real basic hazard scenarios. Basic hazard scenarios with authenticity scores not lower than the scoring threshold are selected, and upsampling and transposed convolution are used to perform expansion operations to generate a hazard scenario sample set.

8. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 7, characterized in that: The specific steps for constructing a pipeline safety hazard prediction model in a privacy-preserving computing environment using spatiotemporal risk data and a sample set of potential hazard scenarios are as follows. A pipeline safety hazard prediction model is constructed based on a gradient boosting decision tree architecture. Training samples are formed using spatiotemporal risk data, a sample set of potential hazard scenarios, and corresponding target output values. Based on the training samples, the partitioning features, splitting values, and output values ​​of each internal node in the gradient boosting decision tree are determined, and a configured pipeline safety hazard prediction model is generated.

9. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 8, characterized in that: The specific steps for inputting real-time node-level spatiotemporal risk data into the pipeline safety hazard prediction model to generate safety risk level identifiers for corresponding graph structure nodes are as follows. Collect real-time pipeline operation data and process the real-time pipeline operation data according to the same processing flow used to generate spatiotemporal risk data to generate real-time node-level spatiotemporal risk data. Real-time node-level spatiotemporal risk data is input into the configured pipeline safety hazard prediction model. The risk probability distribution is calculated through gradient boosting decision tree, and the safety risk level identifier of the corresponding graph structure node is output.

10. The method for predicting underground pipeline safety hazards based on privacy computing as described in claim 9, characterized in that: The specific steps for generating corresponding early warning information and operation and maintenance decisions based on different levels of security risk identification are as follows. The safety risk level identifier is associated and matched with the spatial attributes in the pipeline structure data to generate pipeline spatial data with risk labels; By fusing pipeline space data with risk labels with environmental data, a risk dataset is generated. The risk dataset is then color-coded and labeled with symbols to generate a safety hazard risk distribution map. Based on the preset early warning rules, early warning information corresponding to different safety risk levels is generated, and operation and maintenance decisions are generated by combining the operation and maintenance rule table formed by pipeline operation and maintenance procedures, inspection system, maintenance methods and emergency plans.

Images