A risk prediction method for pipeline corridor construction process based on multi-task learning

Abstract

The present invention discloses a risk prediction method for a pipeline corridor construction process based on multi-task learning, which belongs to the technical field of pipeline corridor safety status assessment. The present invention adopts a multi-task learning model, and by combining multiple information sources such as multi-perspective image data and structural health monitoring data, a deep learning model capable of processing spatiotemporal correlations in complex scenarios is constructed, thereby realizing real-time prediction and alarm of potential safety hazards at the construction site. The method overcomes the shortcomings of the prior art such as single data, strong manual dependence, and delayed risk prediction, and effectively improves the accuracy and response speed of construction risk prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of pipe gallery safety status assessment, and in particular to a pipe gallery construction process risk prediction method based on multi-task learning. Background Art

[0002] With the acceleration of global urbanization, the development and utilization of underground space has become an important part of urban infrastructure construction. As an important part of urban underground space, the construction demand for underground pipe galleries and tunnels is increasing. Shield construction has been widely used in this field due to its high efficiency and precision. However, the shield construction process is complex, involving multiple environmental factors such as geological structure, groundwater hydrology, construction machinery, personnel operation and changes in construction conditions. These factors can easily cause safety hazards and pose a serious threat to construction safety.

[0003] At present, although there are a variety of construction safety monitoring and management measures, most of them rely on manual experience and traditional technical means, and it is difficult to achieve a comprehensive and accurate prediction of construction risks. Therefore, developing a risk prediction method for the pipeline corridor construction process based on multi-task learning is of great significance for ensuring construction safety and reducing accidents. Summary of the invention

[0004] The purpose of the present invention is to provide a risk prediction method for the construction process of a pipeline corridor based on multi-task learning. By combining multiple information sources such as multi-perspective image data and structural health monitoring data, a deep learning model is constructed that can handle spatiotemporal correlations in complex scenarios, thereby achieving real-time prediction and alarm of potential safety hazards at the construction site. This method overcomes the shortcomings of the prior art such as single data, strong manual dependence, and delayed risk prediction, and effectively improves the accuracy and response speed of construction risk prediction.

[0005] To achieve the above object, the present invention provides a risk prediction method for the pipeline corridor construction process based on multi-task learning, comprising the following steps:

[0006] S1. Pipe gallery construction and on-site structural health monitoring to obtain original structural health monitoring data;

[0007] S2, preprocessing the original structural health monitoring data in S1;

[0008] S3, converting the preprocessed structural health monitoring data into images through time-frequency transformation and feature extraction technology;

[0009] S4, convert the image of S3 into a standard format that can be processed by the multi-task learning algorithm;

[0010] S5. Establish a risk prediction model for the pipeline corridor construction process based on the structural health monitoring data in S2;

[0011] S6, based on the standard format image converted by S4, a multi-task learning model for predicting the mechanical behavior process during the construction of the pipeline corridor is established on the basis of S5;

[0012] S7: Train and test the multi-task learning model established in S6, and make risk predictions for the construction site to identify potential safety hazards and sound an alarm if there are any.

[0013] Preferably, the specific steps of constructing the pipe gallery and performing on-site structural health monitoring to obtain original structural health monitoring data in S1 are as follows:

[0014] S11. During the deep foundation pit excavation stage, high-precision displacement sensors and strain gauges are deployed to monitor and analyze data in real time, providing timely warning of collapse risks;

[0015] S12. After the deep foundation pit construction is completed, during the main structure construction phase of the tunnel, strain gauges and crack gauge sensors are installed to monitor the stress, strain and crack development of the structure in real time.

[0016] Preferably, the process of preprocessing data in S2 is as follows:

[0017] S21. Clean the data to remove redundant, erroneous or missing information. For missing values, use mean filling or median filling according to the data distribution law and business logic.

[0018] S22, after the S21 process is completed, identifying and processing abnormal values ​​of the data;

[0019] S23. After processing the outliers in the data, perform data conversion and standardization.

[0020] Preferably, the image in S3 is a two-dimensional image or a three-dimensional image generated by transforming and visualizing the original sensor data, and serves as the input of multi-task deep learning.

[0021] Preferably, the specific steps of converting the image into a standard format that can be processed by the multi-task deep learning algorithm in S4 are as follows:

[0022] S41. Adjust the size so that all images have the same width and height;

[0023] S42, performing normalization processing after resizing, scaling the pixel values ​​of the image from the range of [0, 255] to the range of [0, 1] or [-1, 1];

[0024] S43, after the normalization process, adjusting the color channel order of the image according to the requirements of the deep learning framework;

[0025] S44, after adjusting the color channel order, enhancing the data;

[0026] S45. After data enhancement is completed, batch processing is performed to form batches;

[0027] After S46 and S45 are completed, format conversion is performed: according to the requirements of the deep learning framework, the processed image data is converted into a file format that can be directly read and processed by the deep learning framework.

[0028] Preferably, the process of establishing the risk prediction model for the pipeline corridor construction process in S5 is as follows:

[0029] S51, define a data matrix: regard the monitoring data collected from one sensor as a time series, and obtain a data matrix X composed of the monitoring data collected by multiple sensors.

[0030] X∈R n×L；

[0031] Where n represents the number of sensors, L represents the length of each time series, and R represents the pipeline corridor monitoring data matrix;

[0032] S52. After obtaining the data matrix, define the indicator type:

[0033] , ;

[0034] Where m is the type of monitoring indicator, and is the number of sensors corresponding to monitoring indicator i. The prediction results of x and X are assigned to and , Represents the data matrix of the i-th category monitoring indicator, Represents scalar data stored by the sensor;

[0035] S53. After the indicator type is defined, the time scale is defined. In the empirical process, the time series will be gradually input into the model for training. The input length is called the observation window, denoted as p. On this basis, the information of the next time q is predicted;

[0036] S54. After the above process is completed, the prediction implementation begins, and the data is imported into the recurrent neural network algorithm to predict the monitoring data of the pipeline corridor construction process.

[0037] S55. Conduct risk prediction; set thresholds and define risks at the tunnel construction site based on the prediction results of monitoring data.

[0038] Preferably, the specific steps of establishing a multi-task learning model for predicting the mechanical behavior process during the pipe gallery construction in S6 are as follows:

[0039] S61. Input the monitoring data of different indicators into the model and perform parallel prediction of multiple time series. The process is as follows:

[0040] ;

[0041] ;

[0042] Where argmin() is the function that calculates the minimum error. Represents the time series monitoring data from time t-p+1 to time t, Represents the time series monitoring data from time t+1 to time t+p, represents a function for calculating the minimum error between sensor data and corresponding predicted data;

[0043] S62. After the parallel prediction is completed, a multi-task learning model is established. Specifically, the working process is as follows: the original monitoring data is input into the encoder module for spatiotemporal correlation learning. After the encoding process of multiple spatiotemporal attention blocks, the output of the last step is used as input and transferred to the decoder module for final prediction.

[0044] Preferably, the multi-task learning model adopts and based on The algorithms are used to perform spatial correlation learning and temporal dependence learning respectively, and The framework is based on the two symmetrical encoder and decoder modules, each of which consists of several layers called spatiotemporal attention blocks. The network is the core algorithm for extracting spatiotemporal correlations. is a and The coupled network composed of algorithms, a spatiotemporal attention block is Multiple calculations are superimposed, and then their encoded results are input into the subsequent spatiotemporal attention block for high-jump information learning.

[0045] Preferably, GCN is a topological network for capturing node features in space from structural information. GCN updates the state of each sequence at a certain time by fusing the information of all adjacent sequences in the adjacent matrix A as follows:

[0046] ;

[0047] in , , , D represents the number of features of each time series, F is the dimension of the target feature, and the output , calculated on this basis , the process is as follows:

[0048] ;

[0049] ;

[0050] in and yes The features captured by the layers A and L, A and W are shared by all nodes in all layers, and E1 and E2 are two nodes with learnable parameters;

[0051] yes A variant of the model uses two gate elements, the reset gate and the update gate, to capture the long-term dependency of the sequence. The reset gate is used to limit the amount of information in the update part, where the historical hidden state is compressed and the current information is mixed. The update gate is used to generate new information using the information of the reset gate and the activation function. The specific process is:

[0052] ;

[0053] ;

[0054] ;

[0055] ;

[0056] in and represents the reset gate and update gate, and represents weight parameters and bias, , and Respectively represent the historical state, current state and candidate state at the current moment, represents the input data of the current time step, σ and is the activation function; thus determined The model generation process is as follows:

[0057] ;

[0058] ;

[0059] ;

[0060] ;

[0061] in Represents the weight parameters of graph convolution.

[0062] Preferably, the steps of determining the potential safety hazard area and issuing an alarm are as follows:

[0063] S71. Determine risk indicators and indicators that can reflect the health status of the structure;

[0064] S72. After the risk indicators are determined, a weight is assigned to each risk indicator according to its impact on the structural safety, and a basic score is set for each risk indicator;

[0065] S73. The comprehensive score is obtained by multiplying each indicator by its weight and adding the basic score;

[0066] S74. Set thresholds based on the structural safety requirements and historical data. When the comprehensive score exceeds the threshold, the system issues an early warning, defining the risk as the trigger for the corresponding early warning measures.

[0067] S75. Continue to issue risk warnings to the system and adjust weights, basic scores and thresholds based on new data.

[0068] Therefore, the risk prediction method for the pipeline corridor construction process based on multi-task learning using the above structure has the following advantages:

[0069] 1. For the first time, structural health monitoring and environmental monitoring are systematically integrated during the construction of underground pipe corridors to comprehensively evaluate construction safety;

[0070] 2. Standardize and enhance the original data to ensure data consistency and high efficiency of model training, and improve the accuracy of model prediction;

[0071] 3. A GCN-GRU coupling network is proposed, which can deeply learn spatiotemporal features, capture dynamic changes in the construction process, achieve efficient prediction of complex construction scenes, and improve the generalization ability of the model;

[0072] 4. A risk prediction model based on graph network was built, which can analyze monitoring data in real time and issue safety alarms, greatly improving the intelligent level of construction safety management;

[0073] 5. Combined with risk assessment at each stage of construction, full-process safety monitoring and management from the beginning to the end of construction can be achieved to form a closed-loop control.

[0074] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 A multi-task learning model structure diagram of a pipeline corridor construction process risk prediction method based on multi-task learning according to the present invention;

[0076] Figure 2The invention provides a risk prediction method for the pipeline corridor construction process based on multi-task learning. Network structure diagram;

[0077] Figure 3 The present invention is a flow chart of a method for predicting risks in a pipeline corridor construction process based on multi-task learning. DETAILED DESCRIPTION

[0078] Example

[0079] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0080] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0081] like Figures 1 to 3 As shown, the present invention provides a risk prediction method for the pipeline corridor construction process based on multi-task learning, comprising the following steps:

[0082] S1. Pipe gallery construction and on-site structural health monitoring to obtain original structural health monitoring data;

[0083] S11. During the deep foundation pit excavation stage, high-precision displacement sensors and strain gauges are deployed to monitor and analyze data in real time, providing timely warnings of possible collapse risks;

[0084] S12. After the deep foundation pit construction is completed, during the main structure construction phase of the tunnel, strain gauges, crack meters and other sensors are installed to monitor the stress, strain and crack development of the structure in real time.

[0085] S2, preprocessing the original structural health monitoring data in S1;

[0086] S21. Clean the data to remove redundant, erroneous or missing information. For missing values, use mean filling or median filling according to the data distribution law and business logic.

[0087] S22, after the S21 process is completed, identifying and processing abnormal values ​​of the data;

[0088] S23. After processing the outliers in the data, perform data conversion and standardization.

[0089] S3, converting the pre-processed structural health monitoring data into images through time-frequency transformation and feature extraction technology, so that the images can intuitively reflect the mechanical state of the structure;

[0090] The images mentioned here refer to two-dimensional and three-dimensional images generated by transformation and visualization of raw sensor data. They are used as input for multi-task deep learning and are divided into the following categories: time series images, grayscale matrix images, heat maps, and point cloud images.

[0091] S4, convert the image of S3 into a standard format that can be processed by the multi-task learning algorithm;

[0092] S41. Adjust the size to ensure that all images have the same width and height; this is to maintain data consistency and avoid dimension mismatch issues during model training;

[0093] S42, after resizing, normalization is performed to scale the pixel values ​​of the image from the range of [0,255] to the range of [0,1] or [-1,1], which helps to speed up the convergence of the model and improve the stability of training;

[0094] S43. After normalization, adjust the color channel order of the image according to the requirements of the deep learning framework. If the framework requires (red, green, blue) order, while the original image is (blue, green, red) order (e.g. The image read) needs to be converted accordingly;

[0095] S44. After adjusting the color channel order, appropriate data enhancement, such as rotation, scaling, cropping, flipping, etc., can effectively improve the generalization ability of the model. Data enhancement should be performed before or after conversion to a standard format, depending on the enhancement method and the compatibility of the framework;

[0096] S45. After data enhancement is completed, batch processing is performed to form batches. In order to efficiently utilize parallel computing resources such as GPU or TPU, multiple images are packaged into a batch for processing. Therefore, when converting images to a standard format, it is also necessary to consider how to organize these images to form batches;

[0097] After S46 and S45 are completed, format conversion is performed: according to the requirements of the deep learning framework, the processed image data is converted into a file format that can be directly read and processed by the deep learning framework.

[0098] S5. Establish a risk prediction model for the pipeline corridor construction process based on the structural health monitoring data in S2;

[0099] S51, define data matrix: the monitoring data collected from one sensor is regarded as a time series, thereby obtaining a data matrix X composed of monitoring data collected by multiple sensors,

[0100] X∈R n×L；

[0101] Where n represents the number of sensors, L represents the length of each time series, and R represents the pipeline corridor monitoring data matrix;

[0102] S52. After obtaining the data matrix, define the indicator type:

[0103] , ;

[0104] Where m is the type of monitoring indicator, and is the number of sensors corresponding to monitoring indicator i. The prediction results of x and X are assigned to and , Represents the data matrix of the i-th category monitoring indicator, Represents scalar data stored by the sensor;

[0105] S53. After the indicator type is defined, the time scale is defined. In the empirical process, the time series will be gradually input into the model for training. The input length is called the observation window, denoted as p. On this basis, the information of the next time q is predicted, which is called the prediction time scale (20 days).

[0106] S54. After the above process is completed, the prediction implementation begins. The data is imported into the recurrent neural network algorithm to realize the prediction of the monitoring data of the pipeline corridor construction process. The recurrent neural network algorithm has a loop structure, which enables the network to maintain an internal state, thereby capturing the information in the data that evolves over time. Among them, each node (or unit) considers the hidden state of the previous time step while processing the current input. This structure can model the temporal dependency in the data in tasks such as time series analysis, language model, and text generation, so it can be used for time series prediction.

[0107] S55. Before specific implementation, conduct risk prediction: Based on the prediction results of monitoring data, set thresholds and define risks at the tunnel construction site.

[0108] Monitoring types and thresholds:

[0109] Displacement monitoring: The alarm value of the displacement monitoring point is set to a rate greater than 2.5mm / d.

[0110] Stress monitoring: When the stress value exceeds the maximum stress allowed by the design (set according to actual conditions, the value can be changed), an alarm should be issued.

[0111] Crack monitoring: The maximum crack width limit of structural components should be less than or equal to 0.2mm, and an alarm will be issued if it exceeds the limit.

[0112] Vehicle load: The warning threshold for vehicle load monitoring is set as the ratio of the vehicle's gross weight or axle weight to the design vehicle load equal to 2.

[0113] When the monitoring data exceeds the above threshold, it can be defined as a risk. For example, when the rate of the vertical displacement monitoring point of the building is greater than 2mm / d, it can be defined as a risk. Once the above risk threshold is reached, the emergency response plan should be immediately initiated, including but not limited to increasing the monitoring frequency, conducting on-site inspections, taking reinforcement measures or suspending construction.

[0114] S6. Based on the standard format image converted by S4, a multi-task learning model for predicting the mechanical behavior process during the construction of the pipeline corridor is established on the basis of S5. The process is as follows:

[0115] S61. Input the monitoring data of different indicators into the model and perform parallel prediction of multiple time series. The process is as follows:

[0116] ;

[0117] ;

[0118] in is the function that calculates the minimum error, Represents the time series monitoring data from time t-p+1 to time t (time interval p-1), Represents the time series monitoring data from time t+1 to time t+p (time interval p-1), represents a function for calculating the minimum error between sensor data and corresponding predicted data;

[0119] S62. After the parallel prediction is completed, a multi-task learning model is established. The working process is as follows: the original monitoring data is input into the encoder module for spatiotemporal correlation learning. After the encoding process of multiple spatiotemporal attention blocks, the output of the last step is used as input and transferred to the decoder module for final prediction.

[0120] Traditional machine learning models need to learn each monitoring indicator separately, but cannot capture the correlation between these indicators. For example, obtaining the characteristics of the subsidence segment strain is one learning task, and obtaining the characteristics of the Y-direction change is another learning task. Multi-task learning can complete these two tasks in parallel and learn the correlation between them.

[0121] The multi-task learning model adopts and based on The algorithms are used to perform spatial correlation learning and temporal dependence learning respectively, and The framework is built on the basis of two symmetrical encoder modules and decoder modules, and each module is composed of several layers called spatiotemporal attention blocks. The network is the core algorithm for extracting spatiotemporal correlations. is a and A spatiotemporal attention block can be viewed as a coupled network composed of Multiple calculations are superimposed, and then their encoded results are input into the subsequent spatiotemporal attention block for high-jump information learning.

[0122] It is a topological network used to capture node features in space from structural information. By fusing the information of all adjacent sequences in the adjacent matrix A, the state of each sequence is updated at a certain time as follows:

[0123] ;

[0124] in , , , D represents the number of features of each time series, F is the dimension of the target feature, and the output , calculated on this basis , the process is as follows:

[0125] ;

[0126] ;

[0127] in and yes The features captured by the layers A and L, A and W are shared by all nodes in all layers, and E1 and E2 are two nodes with learnable parameters;

[0128] yes A variant of the model that avoids the problem of gradient vanishing and gradient exploding when processing long time series. It uses two gate elements, reset gate and update gate, to capture the long-term dependency of the sequence. The reset gate is used to limit the amount of information in the update part, where the historical hidden state is compressed and the current information is mixed. The update gate is used to generate new information using the information of the reset gate and activation function. The specific process is:

[0129] ;

[0130] ;

[0131] ;

[0132] ;

[0133] in and represents the reset gate and update gate, and represents weight parameters and bias, , and Respectively represent the historical state (indicating the state at the previous moment), the current state (indicating the output state at each time step) and the candidate state at the current moment, represents the input data of the current time step, σ and is the activation function; thus determined The model generation process is as follows:

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] in Represents the weight parameters of graph convolution.

[0139] S7: Train and test the multi-task learning model in S6, and make risk predictions on the construction site to identify potential safety hazards. If there are potential safety hazards, an alarm will be issued. The process is as follows:

[0140] S71. Determine risk indicators and indicators that can reflect the health status of the structure;

[0141] S72. After determining the risk indicators, for each risk indicator, a weight is assigned according to its impact on structural safety, and a basic score is set for each risk indicator; the weights and basic scores are shown in Table 1:

[0142] Table 1 Weights and basic scores

[0143] ;

[0144] S73. The comprehensive score is obtained by multiplying each indicator by its weight and adding the basic score;

[0145] S74. Set thresholds based on the safety requirements of the structure and historical data. When the comprehensive score exceeds the threshold, the system issues an early warning and initiates corresponding early warning measures defined as risks. Early warning measures may include increasing monitoring frequency, on-site inspections, and taking measures to mitigate risks.

[0146] S75. Continue to issue risk warnings to the system and adjust weights, basic scores and thresholds based on new data.

[0147] Therefore, the present invention adopts the above-mentioned structure to systematically integrate structural health monitoring and environmental monitoring in the underground tunnel construction process for the first time, and comprehensively evaluates the construction safety; at the same time, the original data is standardized and enhanced to ensure the consistency of the data and the efficiency of model training, and improve the accuracy of model prediction; and a GCN-GRU coupling network is proposed, which can deeply learn spatiotemporal features, capture dynamic changes in the construction process, realize efficient prediction of complex construction scenes, and improve the generalization ability of the model; a risk prediction model based on a graph network is also constructed, which can analyze monitoring data in real time and issue safety alarms, greatly improving the intelligent level of construction safety management; combined with risk assessment at each stage of construction, full-process safety monitoring and management from the beginning to the end of construction is realized, forming a closed-loop control.

[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A risk prediction method for pipeline corridor construction process based on multi-task learning, characterized in that: The following steps are involved: S1. Pipe gallery construction and on-site structural health monitoring to obtain original structural health monitoring data; S2, preprocessing the original structural health monitoring data in S1; S3, converting the preprocessed structural health monitoring data into images through time-frequency transformation and feature extraction technology; S4, converting the image of S3 into a standard format that can be processed by the multi-task deep learning algorithm; S5. Establish a risk prediction model for the pipeline corridor construction process based on the structural health monitoring data in S2; S6, based on the standard format image converted by S4, a multi-task learning model for predicting the mechanical behavior process during the construction of the pipeline corridor is established on the basis of S5; S7, train and test the multi-task learning model established in S6, and predict risks at the construction site, identify potential safety hazards, and issue an alarm when potential safety hazards are found; In S1, the specific steps for constructing the pipe gallery and conducting on-site structural health monitoring to obtain the original structural health monitoring data are as follows: S11. During the deep foundation pit excavation stage, high-precision displacement sensors and strain gauges are deployed to monitor and analyze data in real time, providing timely warning of collapse risks; S12. After the deep foundation pit construction is completed, during the main structure construction phase of the tunnel, strain gauges and crack gauge sensors are installed to monitor the stress, strain and crack development of the structure in real time; The process of data preprocessing in S2 is as follows: S21. Clean the data to remove redundant, erroneous or missing information. For missing values, use mean filling or median filling according to the data distribution law and business logic. S22, after the S21 process is completed, identifying and processing abnormal values ​​of the data; S23, after processing the outliers in the data, perform data conversion and standardization; The images in S3 are two-dimensional or three-dimensional images generated by transforming and visualizing the original sensor data, which serve as the input for multi-task deep learning; The process of establishing the risk prediction model for the pipeline corridor construction process in S5 is as follows: S51, define a data matrix: regard the monitoring data collected from one sensor as a time series, and obtain a data matrix X composed of the monitoring data collected by multiple sensors. X∈R n×L ; Where n represents the number of sensors, L represents the length of each time series, and R represents the pipeline corridor monitoring data matrix; S52. After obtaining the data matrix, define the indicator type: ， ； Where m is the type of monitoring indicator, is the number of sensors corresponding to the monitoring index i, and the prediction results of x and X are assigned to and , Represents the data matrix of the i-th category monitoring indicator, Represents scalar data stored by the sensor; S53. After the indicator type is defined, the time scale is defined. In the empirical process, the time series will be gradually input into the model for training. The input length is called the observation window, denoted as p. On this basis, the information of the next time q is predicted; S54, after the above process is completed, the prediction implementation is started, the data is imported into the recurrent neural network algorithm, and the monitoring data of the pipe gallery construction process is predicted; S55. Perform risk prediction; set thresholds and define risks at the tunnel construction site based on the prediction results of monitoring data; The specific steps of establishing a multi-task learning model for predicting the mechanical behavior process during the construction of the pipeline corridor in S6 are as follows: S61. Input the monitoring data of different indicators into the model and perform parallel prediction of multiple time series. The process is as follows: ； ； Where argmin() is the function that calculates the minimum error. Represents the time series monitoring data from time t-p+1 to time t, Represents the time series monitoring data from time t+1 to time t+p, represents a function for calculating the minimum error between sensor data and corresponding predicted data; S62, after the parallel prediction is completed, a multi-task learning model is established, specifically: the original monitoring data is input into the encoder module to perform spatiotemporal correlation learning, and after the encoding process of multiple spatiotemporal attention blocks, the output of the last step is used as input and transferred to the decoder module for final prediction; The multi-task learning model uses GCN-based and RNN-based algorithms for spatial correlation learning and temporal dependency learning respectively, and is built on the basis of the seq2seq framework, which consists of two symmetrical encoder modules and decoder modules, and each module consists of several layers of spatiotemporal attention blocks. At the same time, the GCN-GRU network is used as the core algorithm for extracting spatiotemporal correlation. GCN-GRU is a coupled network composed of GCN and GRU algorithms. A spatiotemporal attention block is the superposition of multiple calculations of GCN-GRU, and then its encoding results are input into the subsequent spatiotemporal attention blocks for high-jump information learning; GCN is a topological network used to capture node features in space from structural information. GCN updates the state of each sequence at a certain time by fusing the information of all adjacent sequences in the adjacent matrix A. The process is as follows: ； in , , , D represents the number of features of each time series, F is the dimension of the target feature, and the output , on this basis, the multi-hop GCN is calculated as follows: ； ； in and yes The features captured by the layers A and L, A and W are shared by all nodes in all layers, and E1 and E2 are two nodes with learnable parameters; GRU is a variant of the RNN model. It uses two gate elements, the reset gate and the update gate, to capture the long-term dependency of the sequence. The reset gate is used to limit the amount of information in the update part, where the historical hidden state is compressed and the current information is mixed. The update gate is used to generate new information using the information of the reset gate and the activation function. The specific process is: ； ； ； ； in and represents the reset gate and update gate, and represents weight parameters and bias, , and Respectively represent the historical state, current state and candidate state at the current moment, represents the input data of the current time step, σ and tanh are activation functions; thus, the generation process of the GCN-GRU model is determined as follows: ； ； ； ； Where GCN[] represents the weight parameters of graph convolution.

2. According to a multi-task learning-based risk prediction method for pipeline corridor construction process according to claim 1, it is characterized in that: The specific steps in S4 to convert images into a standard format that can be processed by multi-task deep learning algorithms are as follows: S41. Adjust the size so that all images have the same width and height; S42, performing normalization processing after resizing, scaling the pixel values ​​of the image from the range of [0, 255] to the range of [0, 1] or [-1, 1]; S43, after the normalization process, adjusting the color channel order of the image according to the requirements of the deep learning framework; S44, after adjusting the color channel order, enhancing the data; S45. After data enhancement is completed, batch processing is performed to form batches; After S46 and S45 are completed, format conversion is performed: according to the requirements of the deep learning framework, the processed image data is converted into a file format that can be directly read and processed by the deep learning framework.

3. According to a multi-task learning-based risk prediction method for pipeline corridor construction process according to claim 2, it is characterized in that: The steps to identify potential safety hazards and sound an alarm are as follows: S71. Determine risk indicators and indicators that can reflect the health status of the structure; S72. After the risk indicators are determined, a weight is assigned to each risk indicator according to its impact on the structural safety, and a basic score is set for each risk indicator; S73. Obtain a comprehensive score based on the sum of the weight of each indicator and the basic score; S74. Set thresholds based on the structural safety requirements and historical data. When the comprehensive score exceeds the threshold, the system issues an early warning, defining the risk as the trigger for the corresponding early warning measures. S75. Continue to issue risk warnings to the system and adjust weights, basic scores and thresholds based on new data.

Images