A dangerous chemical pipeline detection and early warning system

Abstract

The present application relates to the technical field of detection in dangerous chemical production pipeline, and particularly relates to a dangerous chemical pipeline detection and early warning system. In order to achieve the above-mentioned purpose, the present application provides a dangerous chemical pipeline detection and early warning system, which comprises a detection system and an early warning platform for evaluating the results of the detection system; the detection system comprises in-pipe detection and out-of-pipe detection; the in-pipe detection comprises pulse eddy current detection, and the out-of-pipe detection comprises unmanned aerial vehicle detection. The present application aims to solve the technical problem that the old pipeline cannot be monitored, managed and early warned in the prior art.

Description

Technical Field

[0001] This invention relates to the technical field of detection within hazardous chemical production pipelines, and particularly to a hazardous chemical pipeline detection and early warning system. Background Technology

[0002] Hazardous chemical production enterprises generally use pressure pipelines for transportation. Due to the production needs and site limitations of these enterprises, the laying of pressure pipelines is often complex, characterized by being overhead, dense, branching, and winding. To inspect these pipelines, inspectors must navigate through numerous interwoven pipes, identifying the direction and welded joints of each pipe, and sometimes climbing to great heights or drilling underground to inspect overhead or buried pipelines. For high-altitude pressure pipelines with demanding inspection spaces, such as those attached to towers, overhead lines, and pipe corridors, manual safety monitoring is difficult. For example, the lack of reliable and convenient inspection tools makes it difficult to effectively inspect pressure pipelines located at high altitudes and covered with insulation. Currently, manual sampling is commonly used to detect corrosion in pressure pipelines under insulation layers; however, working at heights carries a high risk factor, and the work of erecting scaffolding, dismantling insulation, and grinding is arduous, resulting in high maintenance costs, high maintenance difficulty, and low coverage. Therefore, it is necessary to manage pipeline integrity performance to assess pipeline health and schedule inspection and maintenance activities to reduce risks and costs. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a hazardous chemical pipeline detection and early warning system, which aims to solve the technical problem that the prior art cannot monitor, manage and provide safety early warning for old pipelines.

[0004] To achieve the above objectives, this invention proposes a hazardous chemical pipeline detection and early warning system, comprising a detection system and an early warning platform for evaluating the results of the detection system; the detection system includes in-pipe detection and out-of-pipe detection, the in-pipe detection including pulsed eddy current detection, and the out-of-pipe detection including drone detection; the early warning steps of the early warning platform are as follows:

[0005] a. Defect detection and identification: Defect information is obtained by collecting data from in-pipe and out-of-pipe inspections and combining inspection, monitoring, testing and analysis techniques.

[0006] b. Defect growth prediction: Based on the damage prediction model and the collected data, defect growth is predicted.

[0007] c. Based on risk management, recommend the best inspection, maintenance, and repair policies and activities;

[0008] The early warning platform includes a finite element modeling and analysis method, comprising the following steps:

[0009] a. Create a single-line diagram of the piping system;

[0010] b. Establish a node-connection graph;

[0011] c. Establish the finite element model of the pipe fittings;

[0012] d. Establish a finite element model of the piping system;

[0013] e. Calculate the finite element calculation results;

[0014] f. Output stress linearization.

[0015] Preferably, pulsed eddy current detection includes the following steps:

[0016] a. A cylindrical coil is placed vertically inside a metal pipe to create an eddy current field;

[0017] b. By introducing a second-order vector potential, solve the frequency domain analytical expression of the non-axisymmetric vortex field in the pipeline;

[0018] c. Solve the inverse Laplace transform using the residue method to obtain the time-domain analytical expressions for the induced voltage across the detection coil and the eddy current density distribution inside the pipe wall under the excitation of an arbitrary cylindrical coil placed outside the conductive and magnetic metal pipe.

[0019] d. Compare the pulse eddy current diffusion process inside the tube wall and the sensitivity of time-domain induced voltage to wall thickness detection under different coil placement methods.

[0020] Preferably, drone detection uses drone autopilot, including the following steps:

[0021] a. Set the horizontal safety distance Ha and the vertical safety distance Da between the drone and the power distribution line. The specific calculation formula is as follows:

[0022] ;

[0023] ;

[0024] Where Ha represents the horizontal distance between the vehicle center and the drone during autonomous driving, in meters; L represents the longest crossarm length of the distribution network tower, in meters; n represents the safe distance of the distribution network line, in meters; Da represents the vertical safe distance between the drone and the horizontal ground, in meters; H represents the height of the distribution network structure, in meters; and θ represents the range angle between the autonomous drone and the distribution network, in rad.

[0025] b. Based on the inspection route, a control command is also attached. If the drone deviates during its flight, adjustments are made immediately. The adjustment calculation formula is as follows: ;

[0026] Where i represents the number of times the UAV performs a cyclic inspection flight, and h represents the total number of segments or waypoints in the inspection path.

[0027] As a preferred method, deep learning is used to identify pipelines in drone detection. This method includes:

[0028] a. Data augmentation of the dataset using photometric and geometric distortion has higher robustness to images acquired from different sources;

[0029] b. Training samples are generated using the CutMix algorithm to train a model with its original loss function;

[0030] c. Based on YOLOv4, an improved CSPNet is used as the backbone network in the network, and image prediction is performed at multiple scales;

[0031] d. Replace SiLU with ELU to obtain Conv layers with the ELU activation function, apply them to all ConvELU layers in the YOLOv5 architecture, and integrate the convolutional block attention model to find pipeline attention regions in complex scenes.

[0032] As a preferred method, temperature analysis is used in UAV detection. The steps of this method are as follows:

[0033] a. Based on the location information of the drone, the pipeline is divided into multiple segments, and the pipeline region Rp in the image of each segment is extracted;

[0034] b. Extract the maximum stable extreme value region features from Rp to obtain the maximum temperature region Rt(i, j) of the pipeline, where i represents the i-th frame and j represents the j-th region in the frame. Combine the location information of the video taken by the drone to form a pre-position of whether the temperature maximum region is a defect.

[0035] c. Apply SORT to the location region R(i, j) to obtain the tracking information of the temperature maximum region, thereby determining whether the current temperature maximum region location has changed from the historical maximum region location. If the maximum region has changed, mark the anomaly point in the region for on-site confirmation by technicians.

[0036] d. Visualize the temperature changes of thermal pipelines in the same location area. Use color-scale statistical methods to statistically analyze the minimum, maximum and average values ​​of the color scales to understand the temperature changes of the pipelines in the area. Use SE-TCN to pre-judge and mark whether the temperature changes are abnormal for on-site confirmation and analysis by technicians.

[0037] Preferably, an SE attention mechanism module is added to SE-TCN, the SE attention mechanism module including the following steps:

[0038] a. Perform global pooling on the channels to encode the entire feature into a single global feature;

[0039] b. Learn the correlation between feature channels through a fully connected network to obtain the weight coefficients of the importance of each channel;

[0040] c. Weight the coefficients to the corresponding feature channels to redistribute the importance of different channels.

[0041] Preferably, the early warning platform includes a security knowledge graph, which is constructed using the following steps:

[0042] a. Acquire and preprocess data from multiple heterogeneous sources;

[0043] b. Hazardous materials information extraction, including entity extraction, entity relationship extraction, and attribute extraction;

[0044] c. Construct the hazardous materials ontology using a meta-learning model.

[0045] As a preferred approach, the meta-learning model is constructed as follows: A task distribution p(T) is defined on the dataset D, from which tasks T ~ p(T) can be sampled. The training process of meta-learning first relies on sampling M training tasks from the dataset. ,in This is called the support set. This is called a query set, which supports the training and testing processes within a single task, and the training process optimizes parameters. ; , L meta and L task and represent the loss function of the outer meta-learner and the loss function of the inner task algorithm, respectively. In a single task The task parameters that achieve appropriate performance are... The meta-parameters represent the best performance across all tasks; the testing process of meta-learning relies on Q test tasks in the dataset that do not overlap with the training task data. ,use Train each task in Dmeta-test of The formula is: .

[0046] Preferably, the single-line diagram of the piping system includes a method for creating a rasterized diagram, comprising the following steps:

[0047] a. Binarize the grayscale image; remove the background and some impurities and noise from the image, extract the target object, set a certain threshold T, and use T to divide the image data into two parts: the pixel group greater than T and the pixel group less than T, i.e. the target object and the background area, so as to extract the target object from the multi-value image.

[0048] b. Perform image smoothing and denoising, as well as image thinning, transforming the original image, which is more than one pixel wide, into a single-pixel image.

[0049] As a preferred method, the creation of a single-line diagram of a piping system includes text information processing methods, comprising the following steps:

[0050] a. Identify text information in drawings; use a bidirectional RNN to predict the feature sequence, learn each feature vector in the sequence, and output the predicted label distribution; use CTC Loss to transform a series of label distributions obtained from the recurrent layer into the final label sequence;

[0051] b. Perform natural language processing on the text information in the identified drawings; classify the identified text, determine the pipes it is labeled with, thereby achieving the matching of text information and pipes, and finally generating a node-connection relationship graph.

[0052] Compared with existing technologies, the beneficial effects of the hazardous chemical pipeline detection and early warning system provided by this invention are as follows:

[0053] (1) Construct an early warning platform for the safety production of hazardous chemicals in extreme environments, and establish a life-saving guarantee system that reduces personnel, increases efficiency, and ensures safety, in order to meet the needs of the two levels of "ground-air" and "internal-external" in the safety production of hazardous chemicals.

[0054] (2) The fully automated inspection of the operation status of pipelines scattered in the high altitude of the hazardous chemical production area is carried out. Based on the pipeline image edge detection method and temperature analysis algorithm of YOLOv5, the detection accuracy of UAV is greatly improved.

[0055] (3) The method of using pipeline simulation is to first obtain the pipeline size, pipe fitting connection relationship and support information based on the pipeline single line diagram, and then establish a three-dimensional finite element model of the pipeline, and process the stress analysis results after finite element analysis.

[0056] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. Attached Figure Description

[0057] Figure 1 This is the principle of the wall thickness corrosion pulsed eddy current electromagnetic non-destructive testing system according to an embodiment of the present invention.

[0058] Figure 2This is an embodiment of the present invention of a cylindrical coil eddy current detection model placed vertically inside a metal pipe.

[0059] Figure 3 This is the theoretically calculated curve of the time-domain induced voltage of the conductive and magnetically conductive pipe according to an embodiment of the present invention.

[0060] Figure 4 This is the architecture of the unmanned aerial vehicle (UAV) automatic inspection platform system according to an embodiment of the present invention.

[0061] Figure 5 This is a schematic diagram of lidar scanning according to an embodiment of the present invention.

[0062] Figure 6 This is a structural diagram of the inspection path according to an embodiment of the present invention.

[0063] Figure 7 This is a schematic diagram of a drone scanning operation according to an embodiment of the present invention.

[0064] Figure 8 This is a schematic diagram of the YOLOv5 network structure according to an embodiment of the present invention.

[0065] Figure 9 This is a schematic diagram of the ConvELU layer structure according to an embodiment of the present invention.

[0066] Figure 10 This is a schematic diagram of the CBAM structure according to an embodiment of the present invention.

[0067] Figure 11 This is a three-primary-color light mode image according to an embodiment of the present invention.

[0068] Figure 12 This is an edge recognition map of an infrared acquisition image according to an embodiment of the present invention.

[0069] Figure 13 This is a flowchart of the pipeline temperature analysis algorithm according to an embodiment of the present invention.

[0070] Figure 14 This is a residual structure diagram of the SE-TCN according to an embodiment of the present invention.

[0071] Figure 15 This is a schematic diagram of the SE-TCN anomaly detection model according to an embodiment of the present invention.

[0072] Figure 16 This is a flowchart of the knowledge graph research based on multi-source heterogeneous data fusion according to an embodiment of the present invention.

[0073] Figure 17 This is the process of meta-model learning in an embodiment of the present invention.

[0074] Figure 18 This is a diagram of the CRNN network structure according to an embodiment of the present invention.

[0075] Figure 19 This is a diagram of the Chinese method analysis (LAC) network structure according to an embodiment of the present invention.

[0076] Figure 20 This is the Hough line detection method according to an embodiment of the present invention.

[0077] Figure 21 This is a single-line diagram of piping system information feature diagram according to an embodiment of the present invention.

[0078] Figure 22 This is the process-oriented automatic finite element modeling technology architecture of this invention.

[0079] Figure 23 This is a schematic diagram of the hierarchical structure of the evaluation system according to an embodiment of the present invention.

[0080] Figure 24 This is a scale table according to an embodiment of the present invention.

[0081] Figure 25 This is a flowchart illustrating the security warning process according to an embodiment of the present invention.

[0082] Figure 26 This is the overall architecture of the early warning platform according to an embodiment of the present invention. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0084] In the description of this invention, it should be noted that when an element is referred to as being "fixed to" or "set on" another element, it can be directly on or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to or indirectly connected to the other element.

[0085] In the description of this invention, it should be noted that the terms "center," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. "Several" means one or more, unless otherwise explicitly specified.

[0086] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0087] A hazardous chemical pipeline detection and early warning system includes a detection system and an early warning platform for evaluating the results of the detection system. The early warning platform is a management method, tool, and activity procedure for assessing pipeline health and scheduling inspection and maintenance activities to reduce risk and cost. The detection system includes in-pipe detection and out-of-pipe detection; the in-pipe detection includes pulsed eddy current detection, and the out-of-pipe detection includes drone detection.

[0088] The early warning steps of the early warning platform are as follows:

[0089] a. Defect detection and identification: Defect information is obtained by collecting data from in-pipe and out-of-pipe inspections and combining inspection, monitoring, testing and analysis techniques.

[0090] b. Defect growth prediction: Based on the damage prediction model and the collected data, defect growth is predicted.

[0091] c. Based on risk management, recommend the best inspection, maintenance, and repair policies and activities.

[0092] The early warning platform includes a finite element modeling and analysis method, comprising the following steps:

[0093] a. Create a single-line diagram of the piping system;

[0094] b. Establish a node-connection graph;

[0095] c. Establish the finite element model of the pipe fittings;

[0096] d. Establish a finite element model of the piping system;

[0097] e. Calculate the finite element calculation results;

[0098] f. Output stress linearization.

[0099] See Figure 1 In an optional embodiment, the detection system includes in-pipe detection, wherein the pulsed eddy current method (PECT) replaces the traditional sinusoidal current excitation with a pulsed current, generating a pulsed magnetic field outside the conductor, inducing pulsed eddy currents inside the conductor. The degree of corrosion of the wall thickness is measured by detecting the decay process of this pulsed eddy current electromagnetic field. Wall thickness detection can be performed on in-service ferromagnetic pipes outside the cladding layer, with a cladding layer thickness of up to 200 mm. The principle of the eddy current detection mechanism is as follows: Figure 1 As shown, the system mainly consists of a computer host, a DA / AD converter, a pulse power amplifier circuit, and a coil probe. The coil probe comprises a coil frame, an excitation coil, and a detection coil. The two ends of the excitation coil are connected to the output of the power amplifier circuit, the two ends of the detection coil are connected to the input of the AD converter, and the two ends of the current sampling resistor are also connected to the input of the AD converter. The computer host can perform signal acquisition and control, signal display, data storage, and signal processing functions. A cylindrical coil is vertically placed inside a metal pipe to form an eddy current field. Since the eddy current field under the excitation of the coil placed inside the pipe is a non-axisymmetric three-dimensional eddy current field, the frequency domain analytical expression of the non-axisymmetric eddy current field is first solved by introducing a second-order vector potential. Then, the inverse Laplace transform is solved using the residue method to obtain the time domain analytical expressions for the induced voltage at both ends of the detection coil and the eddy current density distribution inside the pipe wall under the excitation of an arbitrarily placed cylindrical coil outside the conductive and magnetic metal pipe. Finally, the pulse eddy current diffusion process inside the pipe wall and the detection sensitivity of the time-domain induced voltage to the wall thickness are compared under different coil placement methods.

[0100] Let the inner radius of the long, straight, conductive, and magnetic metal pipe be... The outer radius is Pipe wall thickness Electrical conductivity σ, magnetic permeability ( The permeability of free space, (where d is the relative permeability). A hollow cylindrical excitation coil (subscript d) and a detection coil (subscript p) of height h are placed perpendicular to the radial normal of the pipe outside the pipe, as shown in Figure 2. The distance between the edge of the coil probe and the outer wall of the pipe is defined as the probe lift-off distance. The placement method of this coil probe is denoted as method. The inner and outer radii of the cylindrical coil are respectively... and The number of turns is N. Establish a cylindrical coordinate system. And make the Z-axis coincide with the pipe axis. For ease of solution, the field is divided into 3 regions: Region 1: Area 2: Area 3: The pulse current flowing through the excitation coil The Laplace transform is Ignoring displacement current in the model, the magnetic vector potential A satisfies the frequency domain form of the vector Helmholtz equation:

[0101] (1)

[0102] In air field regions 1 and 3 In conductor field region 2, For the non-axisymmetric vortex field problem shown in Figure 2, directly solving for the magnetic vector potential becomes difficult. In this case, introducing a second-order vector potential based on the Lorentz gauge of the magnetic vector potential can simplify the solution. Introducing the second-order vector potential W to represent the magnetic vector potential A, we have:

[0103] (2)

[0104] In the formula Let be the unit vector along the z-axis. and These are independent title functions. Substituting equation (2) into equation (1) and simplifying, we get: (3)

[0105] Therefore, the sufficient condition for equation (3) to hold is that the scalar function... and Simultaneously satisfying the homogeneous Helmholtz equations, the components of the magnetic vector potential and the components of the magnetic flux density B in cylindrical coordinates are as follows:

[0106] (4)

[0107] (5)

[0108] Based on the general solution of the equation, consider Given the boundedness of the field quantities, the general solution expression for the second-order vector potential scalar function of each field region can be obtained as follows:

[0109] (6)

[0110] (7)

[0111] (8)

[0112] (9)

[0113] (10)

[0114] In the formula , and These are the first and second type m-order modified Bessel functions, respectively. , , , , , as well as All are undetermined coefficients. According to the superposition principle, the second-order vector potential in field region 3... It is the superposition of the incident field and the scattering field from the pipe, that is... , among which Describe the incident field excited in air when the excitation coil acts alone; use The scattered field excited in field region 3 when describing the reflection from the pipe surface and the effect of eddy currents inside the pipe wall.

[0115] Then, based on the boundary conditions, after deriving the expressions for each undetermined coefficient and the coil coefficient, the frequency domain expression for the induced voltage of the scattered field caused by the eddy current in the pipe at both ends of the detection coil can be obtained. Scattered field induced voltage It can be divided into induced voltage generated by direct reflection from the pipe surface. and the voltage induced by the eddy current field generated by the eddy currents inside the pipe The two parts have the following frequency domain expressions:

[0116] (11)

[0117] (12)

[0118] After obtaining the expression for the field quantity in the complex frequency domain, the inverse Laplace transform is solved by calculating the frequency domain residue, ultimately yielding the time-domain expression for the pulsed eddy current field. Let... Figure 2 In the pulsed eddy current testing model of the conductive and magnetically permeable pipeline shown, the outer diameter of the pipeline under test is 180 mm, the wall thickness d is 8 mm, the conductivity σ is 4.5 MS / m, and the relative permeability is... The coil lift-off distance from the outer wall of the pipe is 10 mm. The amplitude of the pulsed excitation current is 2.0 A, and the fall time is 0.60 ms. The three-dimensional distribution of eddy current density inside the pipe wall after the excitation current is turned off can be calculated. To observe the diffusion process of the eddy current over time, the field point is determined. The eddy current density modulus at is The theoretical time-domain induced voltage curves for the aforementioned conductive and magnetically permeable pipes with wall thicknesses of 6.0 mm and 8.0 mm can be calculated respectively, as follows: Figure 3 As shown.

[0119] In an optional embodiment, drone inspection utilizes drone autopilot. Fully automated inspections are conducted on the operational status of pipelines scattered high above the hazardous chemical production area. First, risk point identification based on the process is performed in the hazardous chemical production area to determine key inspection locations. High-definition visible light or far-infrared information is collected based on the characteristics of the hazardous chemicals, such as their form and temperature. Through fully automated inspection task management, the entire process—including planned drone startup, fully automated patrol, and automatic return—is automated, significantly improving the efficiency and reducing manpower required for production area inspections. The platform architecture and system processes are as follows: Figure 4 As shown in the figure, the drone intelligent inspection platform is a detection module of the early warning platform.

[0120] To improve the accuracy and efficiency of autonomous flight paths for unmanned aerial vehicles (UAVs), a flight path control algorithm for UAV autonomous flight modeling was developed by combining carrier phase differential positioning technology. LiDAR scanning is used as follows: Figure 5 As shown. To improve the accuracy of data acquisition in the control algorithm, the horizontal safety distance Ha and the vertical safety distance Da between the drone and the power distribution line are first set, and the specific calculation formulas are as follows:

[0121] (13)

[0122] (14)

[0123] Where Ha represents the horizontal distance between the vehicle's center and the drone during autonomous driving, in meters (m); L represents the longest crossarm length of the distribution network tower, in meters; n represents the safe distance of the distribution network line, in meters; Da represents the vertical safe distance between the drone and the horizontal ground, in meters; H represents the height of the distribution network structure, in meters; and θ represents the range angle between the autonomous drone and the distribution network, in rad. The inspection path is as follows: Figure 6 As shown.

[0124] Based on the inspection path, a control command is also attached. If the drone deviates during its flight, it will immediately make adjustments to ensure the safety of the drone's autonomous inspection. The calculation formula for the control algorithm is shown below:

[0125] (15)

[0126] Where i represents the number of times the UAV performs a cyclic inspection flight, and h represents the total number of segments or waypoints in the inspection path.

[0127] Information collection and live video transmission technologies primarily rely on WiFi transmission, which is susceptible to distance limitations. This research aims to construct a multi-channel remote video transmission technology based on 4G / 5G / broadband communication, enabling real-time transmission of drone-captured video and sharing among multiple users. The drone carries a high-definition camera to capture video in real time. The inspection video stream is encoded and decoded using H.264 (a digital video compression format) and then transmitted to the drone hardware system (control remote or smart drone hub). It is then transmitted via the 5G network RTMP protocol to a streaming media server built on the open-source architecture Nginx-rtmp-modul. The streaming media server transcodes the video stream. A self-developed drone automatic inspection system uses the HTTP-FLV live streaming protocol to pull the video stream from the streaming media server, enabling front-end playback of the multi-channel remote video transmission. Key technical processes are as follows: Figure 7 As shown.

[0128] In an optional embodiment, since the video images captured during the drone inspection process cover a wide range, and the pipeline only occupies a small part of them, in order to avoid interference from the surrounding environment of the pipeline and improve the accuracy of the algorithm in identifying the location of defects, it is first necessary to segment the ROI region in the infrared image. The pipeline image edge detection method based on YOLOv5 uses deep learning methods to identify and outline the pipeline part, which requires a large amount of data to ensure the accuracy of the algorithm.

[0129] First, data augmentation is performed on the dataset to make the algorithm more robust to images acquired from different sources. Photometric distortion and geometric distortion are two commonly used data augmentation methods. Geometric distortion includes random scaling, cropping, flipping, and rotating of images. Photometric distortion adjusts the hue, saturation, and numerical values ​​of the image. The CutMix algorithm is added to traditional data augmentation methods. CutMix uses a region from another image to cover the occluded area. By combining two training samples (x... A ,y A ) and (x B y B The samples (x, y) are merged to form a new training sample. The generated training sample is used to train a model with its original loss function. The formula for calculating the merging operation is:

[0130] (16)

[0131] Where M For where to remove and fill in the binary mask in the two images, This is based on element-wise multiplication.

[0132] By combining traditional data augmentation methods with the CutMix algorithm, the detection performance of the network can be improved and the interference of different environments on detection can be reduced.

[0133] The proposed deep learning algorithm is YOLOv5, a relatively advanced YOLO object detection and recognition algorithm. Building upon YOLOv4, it uses an improved CSPNet as the backbone network and performs image predictions at multiple scales to improve prediction and recognition accuracy. The recognition accuracy can reach over 90%. The YOLOv5 network structure is as follows: Figure 8 As shown.

[0134] To improve the accuracy of activation functions in pipeline detection tasks and adapt to various environments, SiLU is replaced with ELU. ELU is a variant of ReLU, which not only reduces training time but also improves the network's test set performance. When x < 0, an exponential function is used to connect the differential function without interruption. The ELU formula is shown below:

[0135] (17)

[0136] Both SiLU and ELU activation functions can address the problems associated with ReLU, but SiLU has limitations in its application. Therefore, ELU activation function was used instead, resulting in Conv layers that apply the ELU activation function. This was then applied to all ConvELU layers in the YOLOv5 architecture, such as... Figure 9 As shown. Furthermore, a Convolutional Block Attention Model (CBAM) is integrated to find pipeline attention regions in complex scenes. CBAM is a simple yet effective attention module. It is a lightweight module that can be integrated into most well-known CNN architectures and can be trained end-to-end. Given a feature map, CBAM infers an attention map sequentially along both channel and spatial dimensions, then multiplies the attention map with the input feature map to perform adaptive feature refinement. The structure of the CBAM module is as follows. Figure 10 As shown.

[0137] Integrating CBAM into different models across various classification and detection datasets significantly improved model performance. In drone-captured images, large coverage areas often contain confusing geographic elements. CBAM helps extract attention regions, allowing YOLOv5 to resist cluttered information and focus on pipe targets. The improved YOLOv5 excels particularly at pipe target detection in drone-captured scenes. Similarly, the algorithm enhances pipe detection performance in complex environments. Figure 11 and Figure 12 This is an example of edge recognition from visible light and infrared images of pipelines in the production area.

[0138] In an optional embodiment, the temperature analysis method is used in drone inspection as a method of intelligent analysis of infrared video streams from drone inspections.

[0139] The temperature analysis algorithm consists of four stages, such as... Figure 13 As shown. Specifically:

[0140] ① Based on the drone's location information, the pipeline is divided into multiple segments (the pipeline area within one frame of the video is considered one segment), and the pipeline region R within each segment is extracted. p ;

[0141] ② For R p Extracting the Maximum Stable Extreme Region Feature (MSER) ​​yields the maximum region R of the pipe temperature. t (i, j), where i represents the i-th frame and j represents the j-th region in that frame. Combining the location information of the video taken by the drone, a pre-positioning is formed to determine whether the region with the highest temperature is a defect.

[0142] ③ Apply the SORT algorithm to the location region R(i, j) to obtain the tracking information of the temperature maximum region, thereby determining whether the current temperature maximum region position has changed from the historical maximum region position. If the maximum region has changed, mark the abnormal point in the region for technical personnel to confirm on site.

[0143] ④ The temperature changes of thermal pipelines in the same location area are visualized. The color scale statistical method is used to statistically analyze the minimum, maximum and average values ​​of the color scale to understand the temperature changes of the pipelines in the area. The SE-TCN algorithm is used to pre-judge and mark whether the temperature changes are abnormal, so that technicians can confirm and analyze them on site.

[0144] TCN is a classic one-dimensional convolutional neural network that can be applied to time series data processing. The TCN network structure mainly consists of four parts: causal convolution, dilated convolution, residual convolution, and one-dimensional convolution.

[0145] To capture the correlations between feature channels and improve the TCN's ability to aggregate global information, an SE attention mechanism module is added to the TCN. This module assigns a weight to each channel, causing the model to focus more on channels with key features and suppress channels with non-key features, thereby improving the model's feature extraction capability. The SE module has low computational cost and mainly consists of three steps: Sequence, Excitation, and Reweight.

[0146] ① Sequeeze: Performs global pooling on the channels, encoding the entire feature into a single global feature;

[0147] ② Excitation: Learn the correlation between feature channels through a fully connected network to obtain the weight coefficients of the importance of each channel;

[0148] ③ Reweight: The weight coefficients are added to the corresponding feature channels to redistribute the importance of different channels.

[0149] The TCN network consists of multiple stacked residual blocks. Compared to the classic TCN network, SE-TCN introduces an SE module after each residual block. The SE module incorporates global max pooling on top of global average pooling, which helps the model obtain the maximum and minimum values ​​of the signal and enhances the SE module's ability to represent global features. The SE-TCN residual structure is as follows: Figure 14 As shown.

[0150] As SE-TCN residual blocks are stacked, the output of the last residual block is aggregated by a fully connected layer to obtain the anomaly probability. The structure of the SE-TCN anomaly detection model is as follows: Figure 15 As shown.

[0151] In an optional embodiment, the early warning platform includes a safety knowledge graph. Information extraction during knowledge graph construction includes entity extraction, entity relationship extraction, and attribute extraction. Before extracting hazardous materials information, data is acquired and preprocessed from multiple heterogeneous source data to improve data quality, reduce unnecessary workload, and enhance the effectiveness of the information extraction model.

[0152] The knowledge graph of hazardous materials is constructed using a bottom-up approach, such as... Figure 16As shown, based on the extraction of hazardous materials information, a hazardous materials ontology is constructed. The ontology is a conceptual template abstracted from a structured knowledge base. The knowledge base constructed based on this ontology has a clear hierarchy, strong structure, and relatively low data redundancy. Ontologies exist in various forms and can be categorized into different types, such as domain ontologies and knowledge ontologies. The construction methods for different ontologies also differ. Generally, methods for building ontologies can be divided into manual, semi-automatic, and automatic construction, but currently, most are done manually, and expert participation is required in specific domains. For the construction of the hazardous materials knowledge ontology, a semi-automatic construction method is used, employing cluster analysis under the constraint of domain knowledge.

[0153] For entities with low frequency of occurrence in the knowledge graph of hazardous chemicals, a meta-model learning algorithm for the hazardous chemicals domain is proposed. The meta-model learning process can be defined as follows: Define a task distribution p(T) on the dataset D, from which tasks T~p(T) can be sampled. The training process of meta-learning first relies on sampling M training tasks from the dataset. ,in This is called the support set. This is called a query set, and it supports the training and testing processes within a single task, corresponding to both the set and the query set. The purpose of the training process is to optimize the parameters. :

[0154] (18)

[0155] (19)

[0156] In formulas (18) and (19), L meta and L task and represent the loss function of the outer meta-learner and the loss function of the inner task algorithm, respectively. In a single task The task parameters that achieve appropriate performance are... This represents the meta-parameter that achieves optimal performance across all tasks.

[0157] The testing process of meta-learning relies on Q test tasks in the dataset that do not overlap with the training task data. ,use Train D meta-test Each task of As shown in formula (20):

[0158] (20)

[0159] For the inner task learner, it is possible to utilize exist The above validates the generalization effect of the task model parameters selected by meta-learning on new tasks. In traditional machine learning, manually defined aspects include model structure, initial parameter values, and parameter update methods; these configurations determine a model's ability to solve a task. Meta-learning utilizes an outer meta-learner to optimize these configurations, selecting task models that perform well on new tasks. The meta-model learning process is as follows: Figure 17 As shown.

[0160] In an optional embodiment, complex piping system single-line diagrams are classified into vectorized and rasterized diagrams. For rasterized diagrams, the grayscale diagrams are first binarized to remove the background and some impurities and noise, extracting the target object. A common method is to set a threshold T, dividing the image data into two parts: a pixel group greater than T and a pixel group less than T, i.e., the target object and the background region, thus extracting the target object from the multi-valued image. Common methods include fixed thresholding and adaptive thresholding. Next, image smoothing and denoising are performed, followed by image thinning. Image thinning is an important image preprocessing technique, aiming to extract the skeleton of the source image, that is, to transform the original image, which is more than one pixel wide, into a single-pixel image, which is beneficial for better image analysis and extraction of the main features. Ideally, the thinned image retains the structural features of the original image, the interrelationships between images, and the extracted skeleton is located in the middle of the original image. Image thinning removes redundant pixels from the image while maintaining its basic shape and topological structure, greatly compressing the data volume of the original image and improving the image recognition speed and accuracy.

[0161] To recognize text information in drawings, deep learning-based OCR technology is employed, such as the CRNN network, widely used in text recognition, as the text recognizer. The CRNN structure primarily consists of CNN + RNN + CTC. Their respective functions are: using a deep CNN to extract features from the input image to obtain a feature map; using a bidirectional RNN (BLSTM) to predict the feature sequence, learning each feature vector in the sequence, and outputting the predicted label (true value) distribution; and using CTCLoss to transform the series of label distributions obtained from the recurrent layers into the final label sequence. The CRNN network structure is as follows: Figure 18 As shown.

[0162] Simultaneously, natural language processing is performed on the text information in the identified drawings, that is, the identified text is classified, and its labeled pipelines are determined, thereby achieving the matching of text information and pipelines, and finally generating a node-connection graph. The text classification network here uses a LAC network (Lexical Analysis of Chinese) (a joint lexical analysis model, such as...). Figure 19As shown, it can perform Chinese word segmentation, part-of-speech tagging, and proper noun recognition tasks in a holistic manner. It processes the text by classifying and tagging parts of speech, and selects the key information needed from the text recognition and saves it to the backend database.

[0163] Furthermore, straight line segments can be identified using the Hough transform algorithm. The flowchart for Hough line detection is shown below. Figure 20 As shown.

[0164] The specific method is explained below:

[0165] (1) In Establish the corresponding parameter space and accumulators, and clear all accumulators to zero;

[0166] (2) Perform a Hough transform on each pixel in the image with a grayscale value of 0, corresponding to... The space consists of curves, and 1 is added to the corresponding accumulator;

[0167] (3) Count the values ​​of each accumulator. If the value is less than the threshold 0, clear the accumulator value to zero, find the maximum value in the accumulator, and obtain the corresponding point. , ), representing a straight line in the xoy coordinate system, where ( , Substituting into equation (10), the equation of the line can be obtained.

[0168] (twenty one)

[0169] By analyzing the largest inscribed circle in the local vicinity of a pixel, data such as line type and line width can be obtained, for example... Figure 21 As shown. Further, the system implements parametric automatic modeling of complex pipe fittings, structured mesh generation of complex pipe fittings, and pipe system assembly and matching. In the automated modeling program architecture, the program consists of a human-computer interaction main interface and an Abaqus PDE secondary development module. In the main interface, the node editor, connection relationship editor, and pipe attribute editor input the geometric parameters of the pipe fittings. After the main interface generates the working file, a multi-threaded mode is used to call five modules in the Abaqus Python development environment to respectively implement working file interpretation, pipe fitting modeling, pipe fitting mesh generation, pipe system assembly, load and boundary loading, as shown. Figure 22 As shown.

[0170] In an optional embodiment, based on the hazardous chemicals safety knowledge graph for the fusion of multi-source heterogeneous datasets and the fully automated modeling underlying system, the stress in the local defect region is extracted through the Abaqus secondary development platform PDE, and the stress field at the local defect location is constructed through interpolation. For each point stress... Solving for the three principal stresses The orientation of the principal stresses was studied, and the variation law of the principal stress orientation was investigated. Based on the conservative principle, the average orientation of the principal stresses and the optimal oblique crack projection plane were determined.

[0171] Furthermore, an AHP analysis method was used, combined with relevant regulations and requirements for hazardous chemicals and special equipment, and the safety status assessment results based on detected defect parameters and simulated stress calculations, to establish a hazardous chemical safety assessment model. Specifically:

[0172] ① Establish a hierarchical structure for the evaluation system. Divide the decision-making objectives, considered factors, and decision objects into three layers based on their interrelationships: the objective layer, the criterion layer, and the indicator layer, and draw a hierarchical structure diagram. The objective layer represents the degree of safety risk of hazardous chemicals; the criterion layer represents the risk factors for hazardous chemical production risk assessment; and the indicator layer represents the factors influencing the risk factors. The hierarchical structure is as follows: Figure 23 As shown.

[0173] ② Construct pairwise comparison judgment matrices. In risk assessment, since there are many factors determining the degree of risk, the consistent matrix method can be used to determine the weights between factors at each level. The consistent matrix method compares factors pairwise using a relative scale to minimize the difficulty of comparing factors with different characteristics. Quantitative standards include... Figure 24 As shown. Then, relative importance is calculated for each risk factor to determine its weight.

[0174] ③ Overall hierarchical ranking and consistency test. Calculate the weights of the relative importance of all factors at a given level to the highest level (overall objective) sequentially from the highest level to the lowest level.

[0175] Based on the risk level assessment indicators of the safety assessment model, the safety early warning platform intelligently and automatically provides risk warnings, dynamic risk situation analysis, accident emergency support, dynamic early warning, and risk distribution for major accident risks.

[0176] A warning model is trained using deep learning or classification algorithms. The safety risk type and level information from the aforementioned safety assessment results are used as input to the model, enabling it to quickly determine the corresponding warning indicators. Furthermore, relevant regulations on hazardous chemicals safety and response plans from professional fields are integrated into a warning plan knowledge base. This knowledge base is then matched with the warning indicators, and based on both semantic and text matching algorithms, recommended reasonable and effective warning plans and corresponding safety risk response measures are provided. Figure 25 As shown, this minimizes losses and reduces the accident rate.

[0177] Finally, after implementing specific countermeasures, the results are inspected and verified, and the risk situation is reassessed. If it meets the requirements, the alarm can be lifted; if not, the warning plan will be reimplemented until it meets the requirements. Ultimately, this will form an intelligent and information-based safety early warning platform for extreme production environments of hazardous chemicals, which will be demonstrated and applied in large-scale hazardous chemical production enterprises, such as... Figure 26 As shown.

[0178] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dangerous chemical pipeline detection and early warning system, characterized in that: The application relates to a detection system and a warning platform for evaluating the results of the detection system; the detection system comprises in-pipe detection and out-of-pipe detection; the in-pipe detection comprises pulse eddy current detection; the out-of-pipe detection comprises unmanned aerial vehicle detection; the warning steps of the warning platform are as follows: a. Defect detection and identification: obtaining defect information by collecting in-pipe detection and out-of-pipe detection data and combining inspection, monitoring, testing and analysis technology; b. Defect growth prediction: predicting defect growth according to a damage prediction model and collected data; c. Risk-based management: suggesting optimal inspection, maintenance and repair policies and activities; The warning platform comprises a finite element modeling analysis method, which comprises the following steps: a. Establishing a pipe system single-line diagram; b. Establishing a node-connection relationship diagram; c. Establishing a pipe fitting finite element model; d. Establishing a pipe system finite element model; e. Calculating finite element calculation results; f. Outputting stress linearization.

2. The system of claim 1, wherein: The pulse eddy current detection comprises the following steps: a. Vertically placing a cylindrical coil in a metal pipeline to form an eddy current field; b. Solving a frequency domain analytical expression of a pipeline non-axisymmetric eddy current field by introducing a second-order vector potential; c. Solving a Laplace inverse transform by using the residue method to obtain a time domain analytical expression of induced voltage at both ends of a detection coil and vortex density distribution in a pipe wall under excitation of an arbitrarily placed cylindrical coil outside a conductive and magnetizable metal pipeline; d. Comparing pulse eddy current diffusion processes in the pipe wall under different coil placement modes and time domain induced voltage detection sensitivity to wall thickness.

3. The system of claim 1, wherein: The unmanned aerial vehicle detection uses an unmanned aerial vehicle to automatically drive and automatically patrol and inspect the running state of a pipeline scattered in the air in a dangerous chemical production area, and comprises the following steps: a. Setting a horizontal safety distance Ha and a vertical safety distance Da between the unmanned aerial vehicle and the distribution network line, and the specific calculation formula is as follows: ； ； Wherein, Ha represents the horizontal distance length maintained between the vehicle center and the unmanned aerial vehicle during automatic driving; L represents the longest cross arm length of the distribution network tower; n represents the safety distance of the distribution network line; Da represents the vertical safety distance between the unmanned aerial vehicle and the horizontal ground; H represents the height of the distribution network structure; and theta represents the range angle between the unmanned aerial vehicle and the distribution network during automatic driving; b. According to the patrol path, an auxiliary control instruction is simultaneously attached, and once the unmanned aerial vehicle deviates during driving, adjustment is immediately made, and the calculation formula for adjustment is as follows: ； Wherein, i represents the cycle of the unmanned aerial vehicle; and h represents the total number of segments or the total number of waypoints of the patrol path.

4. The system of claim 1, wherein: In the unmanned aerial vehicle detection, a deep learning method is used to identify the pipeline, and the method comprises the following steps: a. Using photometric distortion and geometric distortion to perform data enhancement on a data set, so that the data set has higher robustness to different acquired images; b. Using a CutMix algorithm to generate training samples for training a model with an original loss function; c. On the basis of YOLOv4, an improved CSPNet is used as a backbone network in the network, and the image is predicted at multiple scales. d. Replace SiLU with ELU to get the Conv layer of the application activation function ELU, and apply it to all ConvELU layers in the YOLOv5 structure, and integrate the convolution block attention model to find the attention area of the pipeline in the complex scene.

5. The system of claim 1, wherein: In the unmanned aerial vehicle detection, a temperature analysis method is used, and the steps of the method are as follows: a. According to the position information of the unmanned aerial vehicle, the pipeline is divided into multiple sections, and the pipeline region Rp in each section is extracted; b. Extract the maximum stable extreme value region feature of Rp to obtain the maximum region Rt(i, j) of the pipeline temperature, wherein i represents the ith frame, and j represents the jth region in the frame, and the position information of the video shot by the unmanned aerial vehicle is combined to form a preliminary positioning of whether the temperature maximum region is a defect; c. Use SORT on the position region R(i, j) to obtain tracking information of the temperature maximum region, so as to judge whether the current temperature maximum region position and the historical maximum region position change, if the maximum region changes, the abnormal point of the region is marked for on-site confirmation by technical personnel; d. The temperature change of the heat pipeline in the same position region is visualized, a color scale statistical method is used to statistically analyze the minimum value, maximum value and average value of the color scale, the temperature change of the pipeline in the region is mastered, and SE-TCN is used to pre-judge whether the temperature change is abnormal and mark, for on-site confirmation and analysis by technical personnel.

6. A dangerous chemical pipeline detection and early warning system as claimed in claim 5, wherein: An SE attention mechanism module is added to the SE-TCN, and the SE attention mechanism module includes the following steps: a. Global pooling is performed on the channel to encode the entire feature into one global feature; b. The correlation between the feature channels is learned through a fully connected network to obtain a weight coefficient of the importance of each channel; c. The weight coefficient is weighted to the corresponding feature channel to complete the redistribution of the importance of different channels.

7. The system of claim 1, wherein: The early warning platform includes a safety knowledge graph, which is constructed by the following steps: a. Data acquisition and preprocessing are completed from multiple heterogeneous source data; b. Hazardous material information extraction, including entity extraction, entity relationship extraction, and attribute extraction; c. Constructing a hazardous material ontology through a meta-learning model.

8. A dangerous chemical pipeline detection and early warning system as claimed in claim 7, wherein: The method for constructing the meta-learning model is: defining a task distribution p(T) on the data set D, in which a task T ~ p(T) can be sampled, and the training process of the meta-learning first relies on sampling M training tasks from the data set wherein called a support set, called a query set, the support set and the query set correspond to the training and testing process in a single task; Optimizing parameters through a training process ; , where L meta and L task represent the loss function of the outer meta-learner and the inner task algorithm respectively, where denotes the task parameters that achieve suitable performance in a single task, denotes the meta-parameters that achieve optimal performance across all tasks; the testing process of meta-learning relies on Q testing tasks in the dataset that do not overlap with the training task data , using to train the meta-parameters of each task in Dmeta-test , the formula is: .

9. The system of claim 1, wherein: The pipeline single-line diagram includes a rasterization method, which includes the following steps: a. Perform binaryzation processing on the gray-scale drawing; remove the background and part of the impurities and noise in the image, extract the target object, and set a threshold T to divide the image data into two parts, the pixel group greater than T and the pixel group less than T, i.e. the target object and the background area, so as to extract the target object from the multi-value image; b. Perform image smoothing and denoising, and image thinning to change the original image pattern with more than one pixel wide into a single-pixel pattern.

10. The system of claim 9, wherein: The pipeline single-line diagram establishment includes a text information processing method, which includes the following steps: a. Recognize the text information in the drawing; use a bidirectional RNN to predict the feature sequence, learn each feature vector in the sequence, and output a predicted label distribution; use CTC Loss to convert a series of label distributions obtained from the cyclic layer into a final label sequence; b. The text information in the identified drawing is subjected to natural language processing; the recognized text is classified to determine the pipeline marked thereby, so as to realize matching of the text information and the pipeline, and finally generate a node-connection relationship graph.

Images