A health monitoring, intelligent prediction and early warning disposal system for a cross-fault oil and gas pipeline

Abstract

The present application relates to a kind of health monitoring, intelligent prediction and early warning treatment system of cross-fault oil and gas pipeline, including sensor module, execution module, early warning module and AI-PLC collaborative module;AI-PLC collaborative module includes AI module and PLC module connected by network cable;PLC module is configured to generate execution command according to the prediction result obtained from AI module, and send execution command to execution module, generate alarm command according to the prediction result, and send alarm command to early warning module;AI module is configured to obtain acquisition data from sensor module, obtain prediction result according to the acquisition data obtained from sensor module, LSTM algorithm and InceptionV3 model;AI module includes: picture preprocessing module, fire image recognition module, fire prediction judgment module, sliding window length calculation module and learning rate calculation module.The present application combines AI and PLC, and constructs the control closed loop from data perception, intelligent analysis to automatic disposal.

Description

Technical Field

[0001] This invention relates to the field of oil and gas pipeline health assessment technology, and in particular to a health monitoring, intelligent prediction and early warning system for cross-fault oil and gas pipelines. Background Technology

[0002] As a critical infrastructure for long-distance oil and gas transportation, cross-fault oil and gas pipelines are directly related to energy supply, the ecological environment, and the safety of life and property of residents along the route. Because oil and gas pipelines crossing active faults are subjected to the combined effects of multiple factors such as fault creep, seismic activity, temperature fluctuations, and third-party damage, they are highly susceptible to accidents such as pipeline strain accumulation, local buckling, weld cracking, and even rupture and leakage. Once an oil or gas leak occurs, it can trigger fires or explosions, causing serious secondary disasters. Traditional monitoring methods for cross-fault pipelines mainly rely on manual inspections, single-point sensor deployment, and experience-based threshold warnings, which have shortcomings such as slow response speed, limited coverage, poor anti-interference capabilities, and difficulty in predicting potential risks.

[0003] A search revealed a pattern recognition system based on a space-time-frequency feature pyramid network disclosed in Chinese patent CN114881085B. This system includes a sequentially connected acoustic vibration data detection module, an adjacent multi-channel space-time-frequency three-dimensional feature map fusion module, a multi-frame space-time-frequency three-dimensional feature pyramid network module, a Bi-LSTM network module, and a pattern recognition module. The acoustic vibration data detection module acquires waveform data of external acoustic vibration signals detected by each sensor channel, obtaining a two-dimensional time-frequency map of adjacent channels. The adjacent multi-channel space-time-frequency three-dimensional feature pyramid network module... The feature map fusion module fuses adjacent multi-channel time-frequency two-dimensional maps into a space-time-frequency three-dimensional feature map, and divides the adjacent multi-channel space-time-frequency three-dimensional feature map into multiple frames of space-time-frequency three-dimensional feature maps in chronological order; the multi-frame space-time-frequency three-dimensional feature pyramid network module is used to extract the space-time-frequency three-dimensional feature information parameters of each frame sequentially; the Bi-LSTM network module is used to mine the temporal correlation feature information of the multi-frame space-time-frequency three-dimensional feature information parameters; the pattern recognition module is used to perform pattern recognition of external vibration sources based on the mined temporal feature information.

[0004] However, the above technical solutions still have at least the following drawbacks: low early warning accuracy, making it difficult to meet the monitoring and early warning needs of cross-fault oil and gas pipelines. Summary of the Invention

[0005] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a monitoring and early warning system for cross-fault oil and gas pipelines, comprising a sensor module, an execution module, an early warning module, and further including:

[0006] AI-PLC collaborative module, the AI-PLC collaborative module includes an AI module and a PLC module connected via a network cable;

[0007] The PLC module is configured to generate an execution command based on the prediction result obtained from the AI ​​module, and send the execution command to the execution module; and to generate an alarm command based on the prediction result, and send the alarm command to the early warning module.

[0008] The AI ​​module is configured to acquire collected data from the sensor module and obtain prediction results based on the collected data acquired from the sensor module, the LSTM algorithm, and the InceptionV3 model.

[0009] The AI ​​module includes: an image preprocessing module, a fire image recognition module, a fire prediction and judgment module, a sliding window length calculation module, and a learning rate calculation module.

[0010] More preferably, the fire image recognition module is configured to divide the image data in the training set and the test set into images with fire and images without fire, and to store the images with fire and images without fire in different subfolders in the training set and the test set respectively.

[0011] The image preprocessing module is configured to unify the size and pixel values ​​of the images with and without fire.

[0012] Remove redundant labels from images with and without fire, and only assign a "fire" label to images with fire and a "no fire" label to images without fire, based on the name of the subfolder.

[0013] More preferably, the fire image recognition module is configured to obtain the number of fire images and non-fire images samples based on the number of fire tags and non-fire tags;

[0014] Weights are assigned to the images with and without fire based on the number of samples in the training set.

[0015] More preferably, the fire prediction and judgment module is further configured to divide the weighted image data into several datasets according to the number of consecutive frames;

[0016] The fire prediction and judgment module is further configured to determine whether a fire has occurred based on the number and weight of the fire images and non-fire images in the dataset.

[0017] More preferably, the collected data includes: annual average creep rate of the fault, critical strain rate threshold for cumulative damage to the pipeline material, sensor sensitivity, and sampling frequency;

[0018] The sliding window length calculation module is configured to determine the theoretical value of the sliding window length based on the sliding window calculation formula, the strain transfer coefficient calculation formula, and the obtained annual average creep rate of the fault, the critical strain rate threshold of the cumulative damage of the pipeline material, the sensor sensitivity, and the sampling frequency.

[0019] More preferably, the formula for calculating the sliding window is shown in equation (1). (1);

[0020] in, This is the theoretical value for the sliding window length. For sensor sensitivity, The strain transfer coefficient, Annual average creep rate of faults This represents the critical strain rate threshold for cumulative damage to the pipeline material. The sampling frequency;

[0021] The formula for calculating the strain transfer coefficient is shown in equation (2).

[0022] (2);

[0023] in, The strain transfer coefficient, The strain direction factor. The angle between the fault and the pipeline. The attenuation coefficient is... The effective length of the pipeline affected by the fault. Spatial decay;

[0024] The formula for calculating the attenuation coefficient is shown in equation (3). (3);

[0025] in, The attenuation coefficient is... For soil reaction modulus, The elastic modulus of the pipe. This represents the cross-sectional area of ​​the pipe.

[0026] More preferably, the fine-tuning interval of the sliding window is determined based on the fine-tuning interval and the theoretical value of the sliding window length;

[0027] If the sliding window length of the monitoring and early warning system for the cross-fault oil and gas pipeline is within the sliding window fine-tuning range, the sliding window length does not need to be adjusted. If the sliding window length is outside the sliding window fine-tuning range, the sliding window length is adjusted to the theoretical value of the sliding window length.

[0028] More preferably, the learning rate calculation module is configured to determine the effective learning rate range based on the collected data, the initial maximum learning rate, and the minimum effective learning rate;

[0029] The learning rate for each training round is dynamically set according to the learning rate decay formula and the effective learning rate range.

[0030] After each fixed training cycle, the learning rate is reset to the initial maximum learning rate.

[0031] More preferably, the learning rate decay formula is shown in equation (4).

[0032] (4);

[0033] in, The learning rate at the current moment. The initial maximum learning rate, To minimize the effective learning rate, This is the current training round number. This represents the total number of training rounds.

[0034] More preferably, the sensor module includes a pressure sensor, a displacement sensor, an acoustic sensor, a vibration sensor, and a camera;

[0035] The execution module includes an emergency shut-off valve and a remote interlocking shutdown system;

[0036] The early warning module includes an alarm;

[0037] The execution module and the early warning module are both electrically connected to the PLC module, and the sensor module is electrically connected to the AI ​​module.

[0038] Compared with existing technologies, the AI-PLC collaborative module provided by this invention significantly improves the overall performance and system stability of the cross-fault oil and gas pipeline monitoring and early warning system by combining AI and PLC. This invention employs the InceptionV3 model and LSTM algorithm for fire video monitoring and pipeline time-series physical quantity prediction, respectively, and suppresses single-frame instantaneous interference through mechanisms such as image preprocessing, category weight allocation, and continuous multi-frame comprehensive judgment, thereby improving the accuracy of fire identification. The sliding window length calculation module of this invention establishes a mapping relationship between the physical parameters of the pipeline's geological environment, such as the annual average creep rate of the fault and the strain transfer coefficient, and the sliding window length of the LSTM model, enabling the LSTM model to operate at a near-optimal window length under different fault conditions, thus improving the accuracy of time-series prediction. The learning rate calculation module of this invention adopts a cosine annealing learning rate decay strategy with hot restart, allowing the model to quickly approach the optimal solution in the early stages of training, achieve fine-tuning in the later stages of training, and escape local optima through periodic resets, further improving the convergence speed and final accuracy of the deep learning model. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall architecture of the cross-fault oil and gas pipeline health monitoring, intelligent prediction and early warning system of the present invention.

[0040] Figure 2 This is a flowchart illustrating the fire image recognition module of the present invention;

[0041] Figure 3 This is a flowchart illustrating the fire prediction and judgment module of the present invention;

[0042] Figure 4 This is a flowchart illustrating the sliding window length calculation module of the present invention;

[0043] Figure 5 This is a flowchart illustrating the learning rate calculation module of the present invention. Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] To facilitate understanding of the methods and systems provided in the embodiments of this application, the background of the embodiments of this application will be introduced before introducing the embodiments of this application.

[0046] As a critical infrastructure for long-distance oil and gas transportation, cross-fault oil and gas pipelines are directly related to energy supply, the ecological environment, and the safety of life and property of residents along the route. Because oil and gas pipelines crossing active faults are subjected to the combined effects of multiple factors such as fault creep, seismic activity, temperature fluctuations, and third-party damage, they are highly susceptible to accidents such as pipeline strain accumulation, local buckling, weld cracking, and even rupture and leakage. Once an oil or gas leak occurs, it can trigger fires or explosions, causing serious secondary disasters. Traditional monitoring methods for cross-fault pipelines mainly rely on manual inspections, single-point sensor deployment, and experience-based threshold warnings, which have shortcomings such as slow response speed, limited coverage, poor anti-interference capabilities, and difficulty in predicting potential risks.

[0047] A search revealed a pattern recognition system based on a space-time-frequency feature pyramid network disclosed in Chinese patent CN114881085B. This system includes a sequentially connected acoustic vibration data detection module, an adjacent multi-channel space-time-frequency three-dimensional feature map fusion module, a multi-frame space-time-frequency three-dimensional feature pyramid network module, a Bi-LSTM network module, and a pattern recognition module. The acoustic vibration data detection module acquires waveform data of external acoustic vibration signals detected by each sensor channel, obtaining a two-dimensional time-frequency map of adjacent channels. The adjacent multi-channel space-time-frequency three-dimensional feature pyramid network module... The feature map fusion module fuses adjacent multi-channel time-frequency two-dimensional maps into a space-time-frequency three-dimensional feature map, and divides the adjacent multi-channel space-time-frequency three-dimensional feature map into multiple frames of space-time-frequency three-dimensional feature maps in chronological order; the multi-frame space-time-frequency three-dimensional feature pyramid network module is used to extract the space-time-frequency three-dimensional feature information parameters of each frame sequentially; the Bi-LSTM network module is used to mine the temporal correlation feature information of the multi-frame space-time-frequency three-dimensional feature information parameters; the pattern recognition module is used to perform pattern recognition of external vibration sources based on the mined temporal feature information.

[0048] However, the above technical solutions still have at least the following drawbacks: low early warning accuracy, making it difficult to meet the monitoring and early warning needs of cross-fault oil and gas pipelines.

[0049] like Figure 1 As shown, some embodiments of the present invention relate to a monitoring and early warning system for cross-fault oil and gas pipelines, including a sensor module, an execution module, and an early warning module, characterized in that it includes:

[0050] AI-PLC collaborative module, the AI-PLC collaborative module includes an AI module and a PLC module connected via a network cable;

[0051] The PLC module is configured to generate an execution command based on the prediction result obtained from the AI ​​module, and send the execution command to the execution module; and to generate an alarm command based on the prediction result, and send the alarm command to the early warning module.

[0052] The AI ​​module is configured to acquire collected data from the sensor module and obtain prediction results based on the collected data acquired from the sensor module, the LSTM algorithm, and the InceptionV3 model.

[0053] The AI ​​module includes: an image preprocessing module, a fire image recognition module, a fire prediction and judgment module, a sliding window length calculation module, and a learning rate calculation module.

[0054] In these embodiments, the AI-PLC collaborative module provided by this invention significantly improves the overall performance and system stability of the cross-fault oil and gas pipeline monitoring and early warning system by combining AI and PLC. This invention employs the InceptionV3 model and the LSTM algorithm for fire video monitoring and pipeline time-series physical quantity prediction, respectively, and suppresses single-frame instantaneous interference through mechanisms such as image preprocessing, category weight allocation, and continuous multi-frame comprehensive judgment, thereby improving the accuracy of fire identification. The sliding window length calculation module of this invention establishes a mapping relationship between the physical parameters of the pipeline's geological environment, such as the annual average creep rate of the fault and the strain transfer coefficient, and the sliding window length of the LSTM model, enabling the LSTM model to operate at a near-optimal window length under different fault conditions, thus improving the accuracy of time-series prediction. The learning rate calculation module of this invention adopts a cosine annealing learning rate decay strategy with hot restart, enabling the model to quickly approach the optimal solution in the early stages of training, achieve fine-tuning in the later stages of training, and escape local optima through periodic resets, further improving the convergence speed and final accuracy of the deep learning model.

[0055] Furthermore, the AI-PLC collaborative module consists of an AI module and a PLC module physically connected via a network cable. The two modules interact with each other and coordinate commands through a communication protocol (such as the Mitsubishi MC protocol). This design combines the deep learning computing power of the AI ​​module with the high stability and reliability of the PLC module's industrial control capabilities, thus improving the overall performance and stability of the system.

[0056] In one specific implementation, the PLC module (e.g., a Mitsubishi PLC controller) serves as the lower-level machine and is the core control unit of the system. The functions of the PLC module include: first, receiving prediction results and analysis conclusions from the AI ​​module; second, generating and sending execution commands to the execution module based on the prediction results, such as performing emergency shutdown and cut-off operations when a fire or serious leakage risk is predicted; and third, generating and sending alarm commands to the early warning module (e.g., an alarm device) based on the prediction results to provide different levels of audible and visual alarms.

[0057] The AI ​​module is responsible for data processing and intelligent analysis. Its functions include: first, acquiring real-time sensor data from the sensor module, particularly video image data from cameras and time-series data from other sensors; second, based on the acquired data, using the InceptionV3 model to perform fire identification on the video image data, and using the Long Short-Term Memory (LSTM) network algorithm to predict the trends of time-series physical quantities; and third, sending the analyzed prediction results (e.g., fire probability, concentration prediction, strain prediction, etc.) back to the PLC module for decision-making. Internally, the AI ​​module is further divided into: an image preprocessing module, a fire image recognition module, a fire prediction and judgment module, a sliding window length calculation module, and a learning rate calculation module.

[0058] Furthermore, in the AI-PLC collaborative module, the communication between the PLC module and the AI ​​module is implemented using the Mitsubishi MC protocol, and a communication connection with the PLC is established by installing the HslCommunication library function. In a specific embodiment, the network address of the PLC module is preferably 192.168.5.250, and the port number is preferably 5005. After obtaining the prediction result, the AI ​​module directly writes the control instructions or predicted values ​​into the designated data register of the PLC module through this protocol. For example, it writes the predicted abnormal values ​​(such as the predicted value of carbon dioxide concentration or temperature at the next moment) into data register D107, and writes the control execution instructions (such as the status word for starting / stopping the actuator) into data register D100. The ladder diagram program inside the PLC calls the data in this register, thereby driving the execution module to perform abnormal handling.

[0059] Furthermore, the system of this invention supports two operating modes: the first is a real-time mode, in which the AI ​​module continuously reads real-time sensor data collected by the PLC module and performs predictions sequentially; the second is a batch mode, in which the AI ​​module performs batch backtracking analysis and model verification on historical data. Users can switch between the two modes through a graphical interface on a host computer.

[0060] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the fire image recognition module is configured to divide the image data in the training set and the test set into images with fire and images without fire, and to store the images with fire and images without fire in different subfolders in the training set and the test set respectively.

[0061] The image preprocessing module is configured to unify the size and pixel values ​​of the images with and without fire.

[0062] Remove redundant labels from images with and without fire, and only assign a "fire" label to images with fire and a "no fire" label to images without fire, based on the name of the subfolder.

[0063] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the fire image recognition module is configured to obtain the sample number of fire images and non-fire images based on the number of fire tags and non-fire tags.

[0064] Weights are assigned to the images with and without fire based on the number of samples in the training set.

[0065] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the fire prediction and judgment module is further configured to divide the weighted image data into several datasets according to the number of consecutive frames.

[0066] The fire prediction and judgment module is further configured to determine whether a fire has occurred based on the number and weight of the fire images and non-fire images in the dataset.

[0067] In these implementations, the fire monitoring function of this system is mainly completed collaboratively by the image preprocessing module, the fire image recognition module, and the fire prediction and judgment module in the AI ​​module, and its specific implementation is as follows:

[0068] First, the image preprocessing module and the fire image recognition module work together to prepare data and train the model. To build a high-precision fire recognition model, a large amount of image data is needed as both the training and testing sets.

[0069] In terms of data classification and storage, the collected pipeline monitoring video frame image data are divided into two main categories: "images with fire" and "images without fire." At the file system level, these two types of images are stored in different subfolders (for example, in "... / train / fire" and "... / train / nofire" respectively). This allows the image data to be automatically labeled "with fire" or "without fire" based on the subfolder name, eliminating the need for tedious manual label cleaning and setting.

[0070] In terms of image preprocessing, to meet the model's input requirements and improve model performance, the image preprocessing module performs standardization operations on all images. Specifically, this includes: resizing all images to a fixed size (e.g., 300×300 pixels) to fit the input requirements of the InceptionV3 model; and normalizing the pixel values ​​of the images, for example, scaling the pixel values ​​from the range [0, 255] to the range [0, 1], which helps accelerate model convergence. Furthermore, data augmentation operations can be performed, such as randomly flipping the training images horizontally to increase data diversity and improve the model's generalization ability.

[0071] Furthermore, considering that in real-world scenarios, the number of images without fire is usually much greater than that with fire, leading to sample imbalance, the model tends to predict all images as "without fire." Before training, the fire image recognition module automatically calculates and assigns different weights to the two categories based on the number of images with and without fire in the training set. For example, assigning a higher weight to the "with fire" category, which has fewer samples, makes the model focus more on correctly classifying "with fire" images during training, thereby effectively reducing the false negative rate of fires. In a specific embodiment, the calculation of the weights for images with and without fire can be automatically completed using the "balanced" mode of the scikit-learn library, and the generated category weight dictionary is passed into the model's fit function.

[0072] Furthermore, the fire prediction and judgment module is responsible for analyzing the real-time video stream and making a final judgment. A flash of flame in a single frame may be a false alarm, such as from vehicle headlights or reflections. To improve the accuracy of the warning, this module adopts a continuous frame-based judgment strategy. It maintains a fixed-length frame buffer, preferably 5 consecutive frames. Only when the number of frames identified as "on fire" in the buffer reaches or exceeds a set threshold—for example, more than 4 out of 5 frames are identified as on fire, and the average fire probability also exceeds a preset alarm threshold (e.g., 0.7)—will the system determine that a real fire has occurred. The PLC module then sends an execution command to control the execution module's actions, sending an alarm command to trigger the alarm. This method integrates time and spatial information, effectively filtering out transient interference and achieving more reliable fire warnings.

[0073] In one specific embodiment, the fire image recognition module is built based on a pre-trained InceptionV3 model. To achieve lightweight and high accuracy, the basic model structure is simplified: the original top fully connected layer is removed, and after obtaining the output of the basic model, a global average pooling layer (GlobalAveragePooling2D) is added sequentially to reduce the number of parameters, a fully connected layer with 128 neurons (with ReLU activation function) is added, and finally an output layer (with Sigmoid activation function) is added for binary classification.

[0074] Furthermore, the fire image recognition module employs a two-stage fine-tuning training strategy: The first stage is a rapid transfer learning stage, where all layers of the InceptionV3 base model are frozen and not used in training. Only a custom classification head network is trained, using the Adam optimizer with a large learning rate (e.g., 0.001) to accelerate convergence, for several training epochs (e.g., 3 epochs). The second stage is a fine-tuning stage, where the last few layers (preferably the last 10 layers) of the base model are unfrozen and trained. The Adam optimizer is used with a small learning rate (e.g., 0.00001) for fine-tuning to further improve detection accuracy, for several training epochs (e.g., 3 epochs). Simultaneously, during training data generation, bilinear interpolation is used for image scaling, and the scikit-learn library is used to automatically calculate class weights (in "balanced" mode). The generated class weight dictionary is then input into the model, effectively solving the problem of missed detections caused by the imbalance between fire and non-fire sample numbers.

[0075] Furthermore, to improve the processing speed of the video stream and reduce computational overhead, the fire prediction and judgment module introduces a frame skipping mechanism when processing real-time video streams. Specifically, the system does not predict frame by frame, but performs fire probability prediction every fixed number of frames (e.g., extracting 1 frame every 3 frames). Simultaneously, the system maintains a queue of continuously detected frames of length 5. When the predicted probability of more than 80% of the frames in the queue (i.e., 4 or more frames) is greater than the fire alarm threshold (e.g., 0.7), a fire alarm state is triggered; if the predicted probability is greater than the warning threshold (e.g., 0.4) but the alarm condition is not met, a safety warning state is triggered.

[0076] Furthermore, the fire prediction and judgment module, based on InceptionV3 single-frame recognition, may also include an LSTM temporal discrimination submodule. This submodule is used to input the InceptionV3 output probability sequence (e.g., a fire probability sequence of 5 or 30 consecutive frames) from multiple consecutive frames into the LSTM temporal discrimination network, which then provides the final fire discrimination result. The LSTM temporal discrimination network preferably includes: an input layer (receiving a 3D tensor of shape [continuous frame length, 1], with the single-frame fire probability as a unique feature), a first LSTM layer (preferably 64 neurons, returning a sequence), a first Dropout layer (preferably with a dropout rate of 0.2), a second LSTM layer (preferably with 32 neurons), a second Dropout layer (preferably with a dropout rate of 0.2), and a final Dense output layer, where the activation function is Sigmoid. This LSTM temporal discrimination submodule, by capturing the temporal evolution of flame probability, can further suppress false alarms caused by instantaneous interference in a single frame (e.g., headlight reflection, screen display, etc.), thus improving the accuracy of the warning.

[0077] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the collected data includes: the annual average creep rate of the fault, the critical strain rate threshold for cumulative damage to pipeline materials, sensor sensitivity, and sampling frequency.

[0078] The sliding window length calculation module is configured to determine the theoretical value of the sliding window length based on the sliding window calculation formula, the strain transfer coefficient calculation formula, and the obtained annual average creep rate of the fault, the critical strain rate threshold of the cumulative damage of the pipeline material, the sensor sensitivity, and the sampling frequency.

[0079] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the sliding window calculation formula is shown in equation (1). (1);

[0080] in, This is the theoretical value for the sliding window length. For sensor sensitivity, The strain transfer coefficient, Annual average creep rate of faults This represents the critical strain rate threshold for cumulative damage to the pipeline material. The sampling frequency;

[0081] The formula for calculating the strain transfer coefficient is shown in equation (2).

[0082] (2);

[0083] in, The strain transfer coefficient, The strain direction factor. The angle between the fault and the pipeline. The attenuation coefficient is... The effective length of the pipeline affected by the fault. Spatial decay;

[0084] The formula for calculating the attenuation coefficient is shown in equation (3). (3);

[0085] in, The attenuation coefficient is... For soil reaction modulus, The elastic modulus of the pipe. This represents the cross-sectional area of ​​the pipe.

[0086] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the fine-tuning range of the sliding window is determined based on the fine-tuning range and the theoretical value of the sliding window length.

[0087] If the sliding window length of the monitoring and early warning system for the cross-fault oil and gas pipeline is within the sliding window fine-tuning range, the sliding window length does not need to be adjusted. If the sliding window length is outside the sliding window fine-tuning range, the sliding window length is adjusted to the theoretical value of the sliding window length.

[0088] This system can not only monitor fires but also use the LSTM algorithm to predict trends in time-series physical quantities. In one or more embodiments of this invention, the LSTM algorithm is applied to, but is not limited to, single-variable time-series signals such as pipe stress (from strain sensors), pipe displacement (from displacement sensors), carbon dioxide concentration (from carbon dioxide concentration detectors), and temperature (from temperature sensors). The length of the sliding window is a key parameter affecting the prediction accuracy of the LSTM model.

[0089] In terms of data preprocessing, the AI ​​module reads historical CSV data using the Pandas library, uses linear interpolation combined with forward / backward filling (interpolate+bfill+ffill) to handle missing values, and uses Scikit-learn's MinMaxScaler to normalize and scale the original data (preferably fitting the scaler to the first 60% of the training set to avoid data leakage); the AI ​​module loads the pre-trained LSTM model and the corresponding scaler file into memory to predict the value of the same physical quantity at the next time step.

[0090] Regarding the LSTM model architecture, the specific network topology of the LSTM prediction model includes: an input layer (receiving a 3D tensor of shape [sliding window length, number of features]), a first LSTM layer (preferably containing 64 neurons and returning a sequence), a first Dropout layer (preferably with a dropout rate of 0.2 to prevent overfitting), a second LSTM layer (preferably containing 32 neurons), a second Dropout layer (preferably with a dropout rate of 0.2), and a final Dense output layer (outputting one predicted value). During model training, the Adam optimizer is used, employing differentiated training learning rate configurations for different physical quantities (e.g., a learning rate of 0.001 for carbon dioxide concentration prediction and 0.0005 for slowly varying physical quantities such as temperature), combined with a mean squared error (MSE) loss function, and an early stopping mechanism (preferably with a patience value of 10) is introduced. Training is automatically stopped and the optimal weights are restored when the validation set loss no longer improves.

[0091] In one specific embodiment, the training-related hyperparameters of the LSTM model can be set as follows: window_size=30, epochs=100, batch_size=32, learning_rate=0.0005-0.001, lstm_units=[64,32], dropout_rate=0.2, validation_split=0.1, patience=10.

[0092] Furthermore, the sliding window length (window_size) can be set as a fixed hyperparameter in conventional time series prediction applications (e.g., empirical values ​​such as 10, 30, etc.); in the special geological condition of cross-fault oil and gas pipelines, since parameters such as fault creep rate, pipeline-fault angle, and soil reaction force vary significantly with geographical location, this invention provides the following sliding window length optimization calculation method, enabling the LSTM model to work at a near-optimal window length in different pipeline segments.

[0093] The sliding window length calculation module is used to scientifically determine the theoretical value of the sliding window length. The data acquired by this module includes: the annual average creep rate of the fault, the critical strain rate threshold for cumulative damage to the pipeline material, sensor sensitivity, sampling frequency, etc. The module determines the theoretical value of the sliding window length using the following formula, which is shown in equation (1):

[0094] (1);

[0095] in, This is the theoretical value for the sliding window length. For sensor sensitivity, The strain transfer coefficient, Annual average creep rate of faults This represents the critical strain rate threshold for cumulative damage to the pipeline material. The sampling frequency is given by equation (1). The physical meaning of equation (1) is that the theoretical value of the sliding window length is determined by the larger of the first characteristic quantity characterizing the fault creep strain accumulation process and the second characteristic quantity characterizing the material damage-sensitive sampling process. The first characteristic quantity is the time required for the pipeline to reach the strain that the sensor can recognize under fault creep. Secondly, it is based on the minimum sampling window time corresponding to the cumulative material damage threshold and the sampling frequency. This ensures that the LSTM input sequence can cover both the strain accumulation process caused by fault creep and the time scale that is sensitive to material damage.

[0096] Among them, strain transfer coefficient It reflects the degree to which fault displacement is transmitted to pipeline strain, and its calculation formula is shown in equation (2): (2);

[0097] in, The strain transfer coefficient, This is the strain direction factor (dimensionless, reflecting the coupling coefficient between the pipe axis and the principal strain direction). The angle between the fault and the pipeline. The attenuation coefficient is... The effective length of the pipeline affected by the fault, factor This is the spatial attenuation term, reflecting the exponential spatial attenuation law as strain propagates along the axial direction of the pipe.

[0098] The attenuation coefficient reflects the influence of the pipe-soil interaction on the strain transmission along the pipe axis, and its calculation formula is shown in equation (3): (3);

[0099] in, The attenuation coefficient is... Soil reaction modulus (unit: N / m³ or equivalent) (Consistent equivalent dimensions) The elastic modulus of the pipe. For the cross-sectional area of ​​the pipe. Specifically, the specific unit system of each physical quantity in equation (3) can be selected according to the actual engineering parameters using a unified International System of Units (SI) to ensure that the dimensions on both sides are consistent.

[0100] Using the above formulas (1) to (3), the system can calculate a theoretically optimal sliding window length based on the specific geological environment of the pipeline (fault creep rate, angle between the fault and the pipeline, soil reaction force, etc.), the pipeline's own properties (elastic modulus, cross-sectional area, etc.), and the monitoring equipment parameters (sensor sensitivity, sampling frequency, etc.).

[0101] Furthermore, the system also possesses the ability to dynamically fine-tune the sliding window length. The system presets a reasonable fine-tuning range (e.g., ±10% of the theoretical value of the sliding window length). During system operation, if the currently used sliding window length is within this fine-tuning range, it is considered valid and requires no adjustment. If changes in certain parameters cause the current sliding window length to exceed this fine-tuning range, the system will automatically adjust it back to the theoretical value of the sliding window length calculated according to formula (1) to ensure that the LSTM model always makes predictions under conditions close to optimal.

[0102] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the learning rate calculation module is configured to determine the effective learning rate range based on the collected data, the initial maximum learning rate, and the minimum effective learning rate.

[0103] The learning rate for each training round is dynamically set according to the learning rate decay formula and the effective learning rate range.

[0104] After each fixed training cycle, the learning rate is reset to the initial maximum learning rate.

[0105] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the learning rate decay formula is shown in equation (4).

[0106] (4);

[0107] in, The learning rate at the current moment. The initial maximum learning rate, To minimize the effective learning rate, This is the current training round number. This represents the total number of training rounds.

[0108] In these implementations, to ensure that the InceptionV3 model and LSTM algorithm in the AI ​​module can be trained quickly and efficiently and achieve high accuracy, the AI ​​module also includes a learning rate calculation module specifically designed to optimize the training process.

[0109] Furthermore, the learning rate calculation module is configured to dynamically adjust the learning rate during model training. The learning rate is a crucial hyperparameter in deep learning training, determining the step size for updating model weights. A suitable learning rate strategy can help the model converge to the optimal solution more quickly, avoiding training stagnation or divergence.

[0110] In one specific implementation, the module operates as follows: First, the effective learning rate range is determined. Before training begins, based on historical data analysis and expert experience, an initial maximum learning rate and a minimum effective learning rate can be set or obtained. For example, it can be set... =0.001, = This range defines the dynamic range of the learning rate during the training process.

[0111] Second, the learning rate is dynamically set. During training, the module does not use a fixed learning rate, but rather calculates and sets the learning rate dynamically for each round of training according to the learning rate decay formula. This invention preferably adopts a decay strategy with cosine annealing, the formula of which is shown in equation (4):

[0112] (4);

[0113] in, The learning rate at the current moment. The initial maximum learning rate, To minimize the effective learning rate, This is the current training round number. This represents the total number of training rounds.

[0114] The decay strategy with cosine annealing is adopted so that the learning rate is large in the early stage of training to quickly approach the optimal solution, and gradually decreases in the later stage of training to achieve fine adjustment.

[0115] Third, to help the model escape local optima, this module also introduces a periodic reset mechanism. Specifically, after each fixed training cycle, the learning rate does not stop at the minimum effective learning rate, but is reset to the initial maximum learning rate, and then the cosine annealing of the next cycle begins. By adopting this mechanism, it helps to find a better global solution, further improving the model's performance.

[0116] In another specific embodiment, the learning rate calculation module can be used in conjunction with the aforementioned fine-tuning training strategy when training the InceptionV3 model: In the first stage (freezing the base model and training only the classification head), a relatively large initial maximum learning rate (e.g., 0.001) is used for cosine annealing; in the second stage (unfreezing the last few layers of the base model), a relatively small initial maximum learning rate (e.g., ...) is used. or Cosine annealing is then performed to allow the two-stage fine-tuning and dynamic learning rate adjustment to work together to further improve training effectiveness.

[0117] In some implementations of monitoring and early warning systems for cross-fault oil and gas pipelines, the sensor module includes a pressure sensor, a displacement sensor, an acoustic sensor, a vibration sensor, and a camera.

[0118] The execution module includes an emergency shut-off valve and a remote interlocking shutdown system;

[0119] The early warning module includes an alarm;

[0120] The execution module and the early warning module are both electrically connected to the PLC module, and the sensor module is electrically connected to the AI ​​module.

[0121] The sensor module is responsible for comprehensively collecting status information of oil and gas pipelines crossing faults. Specifically, it includes: pressure sensors installed along the pipeline route to monitor internal pressure changes, a crucial basis for leak detection; displacement sensors deployed in the fault-crossing areas to monitor minute displacements and deformations caused by geological creep; acoustic and vibration sensors to capture high-frequency acoustic or abnormal vibration signals generated during pipeline leaks; and cameras deployed along the pipeline route, particularly at critical locations such as valve chambers and stations, to provide real-time video feeds for visual fire identification and routine inspections.

[0122] The execution module is responsible for implementing physical intervention measures and controlling the development of the situation after receiving instructions from the PLC module. Specifically, it includes: an emergency shut-off valve, installed on the pipelines on both sides of the fault, which can be remotely controlled by the PLC module to shut down in the event of a serious leak or fire, quickly cutting off the oil and gas supply and preventing the accident from escalating; and a remote interlocking shutdown system, linked with the control systems of upstream and downstream pumping stations or stations, which can remotely interlock and shut down the entire pipeline when emergency conditions are triggered.

[0123] The early warning module is responsible for sending alarms to on-site personnel and the remote monitoring center. Specifically, it includes an alarm, which can be a high-decibel audible and visual alarm, to issue a warning on-site and alert nearby personnel; simultaneously, the alarm signal is also uploaded to the monitoring center via the PLC module.

[0124] All of the aforementioned execution and early warning modules are reliably electrically connected to the PLC module via industrial fieldbus or hardwiring. The PLC module, as the core of field control, continuously collects sensor data and sends it to the AI ​​module. Simultaneously, it receives intelligent decisions from the AI ​​module and transforms them into precise control commands for the execution and early warning modules, thus completing a full closed loop from data acquisition and intelligent analysis to response and handling.

[0125] In one or more embodiments of the present invention, the early warning module combines the intelligent analysis results of the AI-PLC collaborative module to adopt a two-layer hierarchical early warning mechanism.

[0126] The first layer is the sub-module level early warning (taking the fire sub-module as an example): The fire prediction and judgment module divides the system state into three levels based on the prediction probability of consecutive frames: "Safe state" (the prediction probability of consecutive frames is lower than the warning threshold WARNING_THRESHOLD, preferably 0.4), "Warning state" (the prediction probability of consecutive frames exceeds the warning threshold but does not reach the alarm threshold), and "Fire state" (more than 80% of the prediction probability in consecutive frames reaches or exceeds the alarm threshold ALARM_THRESHOLD, preferably 0.7).

[0127] The second layer is a system-level comprehensive early warning: Based on the real-time status, AI prediction results, and historical trends of each sub-module of health monitoring and environmental monitoring, the AI-PLC collaborative module dynamically generates a system-level comprehensive early warning signal according to the requirements of relevant regulations and emergency plans. The system-level comprehensive early warning signal preferably includes four risk levels: blue / attention level (low risk), yellow / warning level (medium risk), orange / severe level (high risk), and red / emergency level (extremely high risk). The early warning signal includes, but is not limited to, the location, nature, level, and preliminary response guidelines of the risk, and is pushed to the monitoring center and relevant personnel through the early warning network to provide a basis for emergency decision-making.

[0128] Furthermore, this system also includes a data display and visualization module. For time-series prediction data (such as strain, displacement, concentration, temperature, etc.), the system uses a chart rendering framework, such as the Dash framework based on Plotly, to maintain a fixed-length (preferably the most recent 500 records) circular data queue in memory and starts a local web server based on the HTTP protocol (port is configurable, such as 8050), dynamically rendering a hyperbolic trend chart of actual and predicted values ​​on the web page. For fire video monitoring, the system uses OpenCV and Tkinter to build a local graphical user interface, drawing red alarm borders, current status text, confidence level, and time on video frames in real time, and simultaneously generating a real-time dashboard image containing fire probability trend lines, warning threshold lines, and alarm threshold lines, achieving intuitive visualization of monitoring results.

[0129] Furthermore, in addition to generating a primary alarm based on the comparison between the predicted value and the threshold, the AI-PLC collaborative module can also perform secondary anomaly monitoring based on the slope difference between the predicted value and the actual value. Specifically, the AI ​​module continuously calculates the slope of the LSTM predicted value sequence and the actual value sequence within the same time window, and calculates the absolute value of the slope difference between the two. When the absolute value of the slope difference exceeds a set threshold (preferably 30 units per time step) within several consecutive time windows, the AI-PLC collaborative module triggers a "model instability" warning, prompting the management personnel to retrain the LSTM model, thereby ensuring the reliability of the prediction model during long-term operation.

[0130] In one specific implementation, the internal ladder diagram program of the PLC module implements at least the following control logic: First, manual / automatic mode switching, allowing operators to manually take over control in emergencies, or the system to operate automatically under normal conditions; Second, sensor data threshold judgment, comparing various collected quantities in real time according to preset thresholds, triggering corresponding alarm states when the threshold is exceeded; Third, actuator linkage control, for example, when a fire is detected, the actuator (such as sprinkler system) is activated and related equipment such as non-fire-fighting power supplies are shut down; Fourth, alarm state reset, allowing the alarm state to be cleared and the system to re-enter the normal monitoring cycle after operator confirmation.

[0131] In a specific embodiment, the key operating parameters of this system can be configured as follows:

[0132] (1) Threshold configuration: The carbon dioxide concentration threshold is preferably 430ppm to 1000ppm (adjusted according to the specific monitoring environment), the temperature threshold is preferably 23℃ to 50℃, the carbon monoxide concentration threshold is preferably 1000ppm, and the smoke concentration threshold is preferably 50ppm;

[0133] (2) PLC communication configuration: TCP / IP communication is used between the host computer and the PLC module (preferably PLC network address 192.168.5.250, port 5005). The prediction results of the AI ​​module are written to the PLC data register D107, and the PLC control instructions are output to the execution module through the data register D100.

[0134] (3) Operating modes: This system supports two operating modes: realtime and batch. Users can switch between the two modes through the graphical interface of the host computer.

[0135] (4) Host computer human-machine interaction: The host computer interface can be implemented using graphical programming software (such as KingSCADA). Operators can use this interface to switch modes, view real-time data curves, confirm alarm status, query historical records and export reports. The interface exchanges data bidirectionally with the PLC module through OPC or a custom communication protocol.

[0136] In a specific embodiment of the present invention, the system of the present invention was tested and verified, and the relevant performance indicators are as follows:

[0137] (1) Fire Image Recognition Module: Using the InceptionV3 model with a two-stage fine-tuning training strategy, under lightweight training conditions with a moderate sample size and fewer training rounds (preferably 3+3=6 rounds, combined with an early stopping mechanism), the training set loss decreased from 0.4351 to 0.0901, and the training set accuracy increased from 0.8140 to 0.9680. The final validation set results were: val_loss=0.0848, val_accuracy=0.9688, and the final training set results were: loss=0.0846, accuracy=0.9741. These results demonstrate that the InceptionV3 + two-stage fine-tuning strategy of this invention can achieve high-precision fire recognition with a validation set accuracy of over 96% without relying on large-scale labeled datasets and high-end GPUs, achieving significantly better detection performance than existing fire recognition methods based on single-stage end-to-end training or shallow networks.

[0138] (2) LSTM time series prediction module: Using a sliding window length of 30 (the theoretical value of the sliding window calculated by formula (1) under normal working conditions and under cross-fault working conditions), a double-layer LSTM (64+32 neurons), a dropout rate of 0.2, an Adam optimizer combined with the MSE loss function and an early stop mechanism, prediction is made for environmental monitoring data (such as carbon dioxide concentration and temperature) along the pipeline. The root mean square error (RMSE) between the measured predicted value and the actual value is 9.53, and the mean absolute percentage error (MAPE) is 0.98%. The results show that the LSTM time series prediction module of the present invention can accurately capture the dynamic evolution law of physical quantities, and the prediction accuracy is sufficient to support industrial-level early warning decision-making.

[0139] (3) System response speed: The fire video monitoring of the present invention adopts a frame skipping mechanism (FRAME_SKIP=3), combined with a judgment queue of 5 consecutive frames, and can realize real-time video processing on a medium-sized industrial computer (e.g., Intel Core i7-11370H processor, 8GB memory, NVIDIA GeForce RTX 3050 graphics card); the host computer writes the prediction result to the PLC every 0.5 seconds through the Mitsubishi MC protocol, and the end-to-end delay from the occurrence of the abnormality to the response of the actuator is preferably no more than 2 seconds, which meets the response requirements of industrial-grade real-time control.

[0140] Both the execution module and the early warning module are electrically connected to the PLC module in the AI-PLC collaborative module. The sensor module is responsible for collecting real-time data on pipeline status and the environment; the execution module is responsible for performing physical operations according to instructions, such as closing valves and stopping transmission; and the early warning module is responsible for issuing audible and visual alarms and other prompts.

[0141] The sensor module includes a pressure sensor, a displacement sensor, an acoustic wave sensor, a vibration sensor, and a camera. The pressure sensor, displacement sensor, acoustic wave sensor, and vibration sensor are used to monitor the pipeline's structural condition, leaks, and the risk of damage from external third parties. The camera is used for real-time video monitoring to capture signs of fire and abnormal activity along the pipeline route.

[0142] In one specific implementation, the sensor module can be further divided into two categories: a health monitoring submodule and an environmental monitoring submodule. The health monitoring submodule is used to collect physical quantity signals characterizing the stress and deformation state of the cross-fault oil and gas pipeline structure in real time. The health monitoring submodule includes at least one or more of strain sensors, displacement sensors, acceleration sensors, and tilt sensors. The strain sensors are deployed at key pipe sections crossing the fault to measure the axial / circumferential strain caused by fault creep. The displacement sensors are deployed on both sides of the fault to measure the relative displacement of the pipeline. The acceleration sensors are used to capture transient vibration signals caused by earthquakes or third-party damage. The environmental monitoring submodule is used to collect signals characterizing fire and leakage in the environmental medium along the pipeline in real time. The environmental monitoring submodule includes at least one or more of a camera, a combustible gas concentration detector, a smoke concentration sensor, and a temperature sensor. The combustible gas concentration detector is used to detect changes in carbon monoxide and / or carbon dioxide concentrations as indicators of oil and gas leakage and incomplete combustion. The camera is deployed along the pipeline route to provide real-time video images for fire identification. In a specific embodiment, the environmental monitoring submodule may include at least four analog input channels: channel 1 is a smoke sensor with a range of 0~2000ppm; channel 2 is a temperature sensor with a range of -50~200℃; channel 3 is a carbon dioxide concentration detector with a range of 0~2000ppm; channel 4 is a carbon monoxide concentration detector with a range of 0~1000ppm; the analog quantities are converted into digital quantities by the analog-to-digital converter module, and then read by the PLC module and restored to engineering quantities by a linear transformation formula, as shown in equation (5):

[0143] (5);

[0144] in, This refers to the raw analog value output by the sensor. This is the lower limit of the range of the analog quantity. This represents the upper limit of the analog quantity's range. This corresponds to the lower limit of the range for the project quantity. This corresponds to the upper limit of the range for the amount of work completed. This represents the restored quantities of the work.

[0145] The execution module includes an emergency shut-off valve and a remote interlocking shutdown system. In one specific embodiment, the execution module can be considered as a collective term for field actuators that respond to control signals from the PLC module. These field actuators include, but are not limited to, one or more of the following: emergency shut-off valves, remote interlocking shutdown systems, electric valves, water pumps, sprinkler systems, fans, and AMD (Active Mass Damper) control systems. In the typical implementation scenario of a cross-fault oil and gas pipeline, the execution module preferably includes an emergency shut-off valve and a remote interlocking shutdown system. The former is installed on the pipeline on both sides of the fault and is remotely shut off by the PLC module to cut off the oil and gas source when a serious leak or fire is detected. The latter is linked with upstream and downstream pumping stations to achieve emergency shutdown of the entire pipeline. In equivalent transformations of applying the AI-PLC collaborative control architecture of this invention to other industrial scenarios or building structure safety monitoring scenarios, the execution module may further include alternative field actuators such as water pumps, sprinkler systems, fans, and AMD control systems. Water pumps and sprinkler systems are used for automatic sprinkler fire suppression in the event of a fire; fans are used for forced ventilation and smoke extraction when carbon dioxide concentration exceeds the standard or smoke accumulates; and the AMD control system is used for active damping regulation when structural vibration exceeds a threshold. All of the above execution modules are driven by control signals from the PLC module (e.g., through the assignment of values ​​to data register D100), and they are conventional equivalent substitutions in the art.

[0146] The early warning module includes an alarm, preferably an integrated sound and light alarm, which can issue sound and light warnings on site and simultaneously upload the alarm signal to a remote monitoring center via a PLC module.

[0147] It should be noted that the "pressure sensor, displacement sensor, acoustic sensor, vibration sensor and camera" described in this specification are preferred embodiments, but the implementation of the present invention is not limited to the specific types of sensors mentioned above; those skilled in the art can replace or supplement the above sensors with other types of sensors such as strain sensors, acceleration sensors, tilt sensors, combustible gas concentration detectors, smoke concentration sensors, temperature sensors, and temperature and humidity sensors, depending on the specific monitoring object, monitoring environment and monitoring purpose. The corresponding replacement or supplementary solutions all fall within the scope of equivalent transformations of the present invention.

[0148] Similarly, the "emergency shut-off valve and remote interlocking shutdown system" described in this specification is a preferred embodiment of the execution module. Those skilled in the art can replace it with other types of actuators such as water pumps, sprinkler systems, fans, AMD control systems, and electric valves according to specific application scenarios. The corresponding replacement schemes all fall within the scope of equivalent transformations of this invention.

[0149] The embodiments and functional operations of the subject matter described in this specification can be implemented in the following ways: digital electronic circuits, tangibly implemented computer software or firmware, computer hardware, including the structures disclosed in this specification and their equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, that is, one or more modules of computer program instructions encoded on one or more tangible non-transitory program carriers, for execution by a data processing device or to control the operation of the data processing device.

[0150] Alternatively or additionally, program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are then generated as coded information to be transmitted to an appropriate receiver device executed by data processing equipment. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or one or more combinations of the above.

[0151] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather as descriptions of features that can embody specific embodiments of a particular invention. Specific features described in this specification within the context of an independent embodiment may also be implemented in combination with a single embodiment. Conversely, various features described within the context of a single embodiment may also be implemented independently in multiple embodiments, or in any suitable sub-combination. Furthermore, while features may be described for combination and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and the claimed combination may be redirected to a sub-combination or a variation thereof.

[0152] Similarly, although operations are described in the accompanying drawings in a specific order, it should not be construed as requiring that such operations be performed in the specific order shown or in sequential order, or that all illustrated operations be performed, in order to achieve the desired result. In certain cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that program components and systems can generally be integrated into a single software product or packaged into multiple software products.

[0153] Specific implementations of the subject matter have been described. Other implementations are within the scope of the following claims. For example, the activities described in the claims can be performed in a different order and still achieve the desired result. As an example, the processes described in the drawings do not necessarily require a specific order or sequence to be shown in order to achieve the desired result. In certain implementations, multitasking and parallel processing may be advantageous.

Claims

1. A monitoring and early warning system for cross-fault oil and gas pipelines, comprising a sensor module, an execution module, and an early warning module, characterized in that, Also includes: AI-PLC collaborative module, the AI-PLC collaborative module includes an AI module and a PLC module connected via a network cable; The PLC module is configured to generate an execution command based on the prediction result obtained from the AI ​​module, and send the execution command to the execution module; and to generate an alarm command based on the prediction result, and send the alarm command to the early warning module. The AI ​​module is configured to acquire collected data from the sensor module and obtain prediction results based on the collected data acquired from the sensor module, the LSTM algorithm, and the InceptionV3 model. The AI ​​module includes: an image preprocessing module, a fire image recognition module, a fire prediction and judgment module, a sliding window length calculation module, and a learning rate calculation module.

2. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 1, characterized in that, The fire image recognition module is configured to divide the image data in the training set and the test set into images with fire and images without fire, and to store the images with fire and images without fire in different subfolders in the training set and the test set respectively. The image preprocessing module is configured to unify the size and pixel values ​​of the images with and without fire. Remove redundant labels from images with and without fire, and only assign a "fire" label to images with fire and a "no fire" label to images without fire, based on the name of the subfolder.

3. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 2, characterized in that, The fire image recognition module is configured to obtain the number of fire images and non-fire images samples based on the number of fire tags and non-fire tags. Weights are assigned to the images with and without fire based on the number of samples in the training set.

4. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 3, characterized in that, The fire prediction and judgment module is also configured to divide the weighted image data into several datasets according to the number of consecutive frames. The fire prediction and judgment module is further configured to determine whether a fire has occurred based on the number and weight of the fire images and non-fire images in the dataset.

5. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 1, characterized in that, The collected data includes: annual average creep rate of the fault, critical strain rate threshold for cumulative damage to pipeline materials, sensor sensitivity, and sampling frequency; The sliding window length calculation module is configured to determine the theoretical value of the sliding window length based on the sliding window calculation formula, the strain transfer coefficient calculation formula, and the obtained annual average creep rate of the fault, the critical strain rate threshold of the cumulative damage of the pipeline material, the sensor sensitivity, and the sampling frequency.

6. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 5, characterized in that, The formula for calculating the sliding window is shown in equation (1). (1); in, This is the theoretical value for the sliding window length. For sensor sensitivity, The strain transfer coefficient, Annual average creep rate of faults This represents the critical strain rate threshold for cumulative damage to the pipeline material. The sampling frequency; The formula for calculating the strain transfer coefficient is shown in equation (2). （2）； in, The strain transfer coefficient, The strain direction factor. The angle between the fault and the pipeline. The attenuation coefficient is... The effective length of the pipeline affected by the fault. Spatial decay; The formula for calculating the attenuation coefficient is shown in equation (3). (3); in, The attenuation coefficient is... For soil reaction modulus, The elastic modulus of the pipe. This represents the cross-sectional area of ​​the pipe.

7. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 6, characterized in that, The sliding window length calculation module is further configured to determine the sliding window fine-tuning interval based on the fine-tuning interval and the theoretical value of the sliding window length; If the sliding window length of the monitoring and early warning system for the cross-fault oil and gas pipeline is within the sliding window fine-tuning range, the sliding window length does not need to be adjusted. If the sliding window length is outside the sliding window fine-tuning range, the sliding window length is adjusted to the theoretical value of the sliding window length.

8. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 1, characterized in that, The learning rate calculation module is configured to determine the effective learning rate range based on the collected data, the initial maximum learning rate, and the minimum effective learning rate. The learning rate for each training round is dynamically set according to the learning rate decay formula and the effective learning rate range. After each fixed training cycle, the learning rate is reset to the initial maximum learning rate.

9. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 8, characterized in that, The learning rate decay formula is shown in equation (4). （4）； in, The learning rate at the current moment. The initial maximum learning rate, To minimize the effective learning rate, This is the current training round number. This represents the total number of training rounds.

10. The monitoring and early warning system for cross-fault oil and gas pipelines according to claim 1, characterized in that, The sensor module includes a pressure sensor, a displacement sensor, an acoustic sensor, a vibration sensor, and a camera; The execution module includes an emergency shut-off valve and a remote interlocking shutdown system; The early warning module includes an alarm; The execution module and the early warning module are both electrically connected to the PLC module, and the sensor module is electrically connected to the AI ​​module.

Images