A Pipeline Leak Detection and Alarm Filtering Method Based on Isolated Forests

By constructing high-frequency and low-frequency data streams based on the isolated forest method, and combining them with a pipeline leak detection and alarm filtering method based on transfer entropy and adaptive threshold, the problems of high false alarm rate and high computing load under complex operating conditions are solved. This enables accurate identification of real leaks and filtering of false anomalies, thereby improving the system's detection reliability and accuracy.

CN123116038APending Publication Date: 2026-11-17HUANENG CLEAN ENERGY RES INST +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611025152.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing pipeline monitoring systems suffer from high false alarm rates and heavy computational loads under complex operating conditions due to common-mode interference and sensor hardware failures. They also struggle to accurately distinguish between real leaks and other pseudo-anomalies. Furthermore, traditional methods are prone to causing near-range missed alarms or long-range false alarms in long-distance pipelines.

Method used

An isolation forest-based approach is adopted to construct a high-frequency raw data stream and a low-frequency feature stream. The global anomaly score is calculated using the isolation forest model, and alarm filtering is performed by combining the transfer entropy and adaptive threshold. The core feature set is extracted using the spatial cutting path of the isolation forest model to generate a dimensionality reduction matrix, and the adaptive threshold is calculated based on the actual physical distance and theoretical wave speed for judgment.

Benefits of technology

It reduces the computing load on edge computing devices, improves the accuracy of pipeline leak alarms, can accurately identify real leaks and filter out false anomalies, ensures the reliability of the system under complex operating conditions, and avoids the false alarm problem in traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116038A_ABST
    Figure CN123116038A_ABST
Patent Text Reader

Abstract

This invention relates to the field of pipeline monitoring technology and discloses a pipeline leak detection and alarm filtering method based on isolated forests. The method includes: collecting multi-source sensor data to construct a low-frequency spatiotemporal feature vector; inputting the feature vector into an isolated forest model to calculate anomaly scores, triggering primary anomalies and extracting core feature sets when scores exceed a threshold; extracting corresponding high-frequency raw data sequences to generate a dimensionality-reduced matrix; calculating the propagation entropy between adjacent nodes, searching for the maximum propagation entropy and its corresponding optimal delay; combining the actual physical distance between nodes and the theoretical wave velocity of the fluid to calculate the propagation delay and construct an adaptive threshold; and executing multi-branch logic to determine and output classification results. This invention combines lightweight isolated forest initial screening with physical causal verification of propagation entropy to filter out common-mode interference and sensor hardware faults, reducing the computational load on edge computing while lowering the system's false alarm rate and achieving accurate identification of real leaks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of pipeline monitoring technology, specifically to a pipeline leak detection and alarm filtering method based on isolated forests. Background Technology

[0002] As a core infrastructure for energy transmission, the safe operation of long-distance pipelines is of paramount importance. Currently, industrial sites typically deploy multi-source sensors at pipeline nodes for real-time status monitoring. With the increase in monitoring dimensions and sampling frequency, edge computing devices face immense data processing pressure. Traditional anomaly detection methods mostly process all high-frequency, multi-dimensional data centrally, requiring extremely high hardware computing power and being susceptible to noise interference from redundant dimensions under complex operating conditions, leading to a decrease in the real-time performance of the system response. To alleviate the computing power bottleneck, some existing solutions employ global downsampling or setting simple static thresholds, but this approach loses the transient fluctuation characteristics of pipeline fluid changes, making it difficult to accurately capture early, weak leakage signals.

[0003] Furthermore, existing technologies, after detecting abnormal fluctuations, often lack in-depth verification by incorporating underlying physical propagation mechanisms. Most anomaly detection models only perform statistical correlation analysis at the data level, failing to distinguish between genuine pipe leaks and simultaneous common-mode interference caused by single sensor hardware failures, localized electromagnetic interference, or heavy machinery construction. This leads to frequent false alarms in complex external environments. Simultaneously, in cross-node correlation analysis and alarm verification, traditional solutions generally employ fixed time tolerances to determine the propagation logic of abnormal fluctuations between adjacent nodes. This fixed threshold mechanism does not adequately consider the dynamic delay shifts caused by waveform dispersion, pipe wall friction, and system timing accuracy errors during the propagation of physical sound waves over long distances in pipelines. This fixed time tolerance is highly susceptible to technical defects that can easily lead to near-range missed alarms or long-range false alarms in practical engineering applications. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a pipeline leak detection and alarm filtering method based on isolated forests. This method solves the problems of high false alarm rates, high computational load, and difficulty in accurately distinguishing between real leaks and other pseudo-anomalies caused by common-mode interference and sensor hardware failures in general pipeline monitoring systems under complex operating conditions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a pipeline leak detection and alarm filtering method based on isolated forests, comprising the following steps:

[0006] Collect multi-source sensor data from monitoring nodes, construct a high-frequency raw data stream and a low-frequency feature stream, and construct a low-frequency spatiotemporal feature vector based on the low-frequency feature stream; The low-frequency spatiotemporal feature vector is input into the isolated forest model to calculate the global anomaly score. The global anomaly score is compared with the alarm threshold. If it is less than or equal to the alarm threshold, the normal monitoring state is maintained. If it is greater than the alarm threshold, the primary anomaly identification is triggered. The contribution weight of the feature dimension is calculated based on the spatial cutting path of the isolated forest model to extract the core feature set. In response to the primary anomaly identifier, a data sequence is extracted from the high-frequency raw data stream based on the core feature set to generate a dimensionality reduction matrix; Based on the reduced-dimensional matrix, the transmission entropy of adjacent monitoring nodes is calculated, and the maximum transmission entropy value and the corresponding optimal data flow delay are obtained by traversing within the sliding time delay window. The theoretical propagation delay is calculated based on the actual physical distance and theoretical wave speed. An adaptive threshold is generated by combining spatial distance compensation. The judgment logic is executed based on the maximum transmission entropy value, the optimal data flow delay, the theoretical propagation delay, and the adaptive threshold, and the alarm classification result is output.

[0007] Furthermore, the step of collecting multi-source sensor data from the monitoring nodes and constructing a high-frequency raw data stream and a low-frequency feature stream includes: An absolute timestamp is added to the collected multi-source sensor data through a time synchronization module; The data sequence with an absolute timestamp is continuously written into a first-in-first-out (FIFO) circular buffer to construct the high-frequency raw data stream; The high-frequency raw data stream is downsampled or a moving average is calculated to obtain a low-frequency data sequence, which is used to construct the low-frequency feature stream. By establishing a high- and low-frequency dual-stream processing architecture, the continuous computational efficiency of anomaly detection and the accuracy requirements of causal quantization on the high-frequency raw data are both considered.

[0008] Furthermore, the step of constructing a low-frequency spatiotemporal feature vector based on the low-frequency feature stream includes: Calculate the temporal difference gradient and bidirectional spatial difference gradient between the current monitoring node and its neighboring monitoring nodes; The low-frequency spatiotemporal feature vector is constructed by concatenating the basic state observations, the temporal difference gradient, and the bidirectional spatial difference gradient.

[0009] The above steps integrate the fluctuations and abrupt changes in the pipeline network in the time domain with the differences in propagation attenuation in the spatial dimension, thereby eliminating common-mode background interference caused by changes in normal operating conditions.

[0010] Furthermore, the step of calculating the contribution weight of the feature dimension based on the spatial cutting path of the isolated forest model for extracting the core feature set includes: The spatial cutting path of the abnormal samples within the isolated forest model is traversed in reverse to determine the hierarchical depth of each feature dimension as the split node. Define a hierarchy weight gain that is inversely proportional to the hierarchy depth, and calculate the contribution weight of each feature dimension based on the hierarchy weight gain; All feature dimensions are sorted in descending order according to their contribution weights, and feature dimensions that meet the set ratios are selected and combined to generate the core feature set.

[0011] By mining the intrinsic topological information of the model tree structure, the core physical parameters that trigger the current anomalous event are dynamically separated.

[0012] Furthermore, the step of extracting data sequences from the high-frequency raw data stream based on the core feature set to generate a dimensionality-reduced matrix includes: The analysis time window is set based on the time when the primary anomaly flag is triggered. The length of the analysis time window is determined by the ratio of the maximum actual physical distance between adjacent monitoring nodes to the sound speed of the fluid medium plus the communication redundancy time tolerance. Generate a binary channel mask vector based on the core feature set; The high-frequency raw data stream within the analysis time window is extracted using the channel mask vector through column indexing. Redundant data not belonging to the core dimensions is removed, and the resulting data is combined to generate the dimensionality reduction matrix. This dual pruning mechanism of time and feature dimensions effectively targets the analysis and reduces computational load.

[0013] Furthermore, the step of traversing within the sliding time delay window to obtain the maximum transfer entropy value and the corresponding optimal data flow delay includes: The sliding time delay window is set as a bidirectional sliding time delay window, and the boundary of the bidirectional sliding time delay window is determined by dividing the actual physical distance between adjacent monitoring nodes by the theoretical minimum wave velocity. According to the set time step, the high-frequency data sequence of adjacent monitoring nodes is gradually shifted within the bidirectional sliding time delay window and the transmission entropy under each offset is calculated. The maximum transmission entropy value is obtained by comparison, and the time offset corresponding to the maximum transmission entropy value is extracted as the optimal data flow delay. The direction of fluctuation transmission is determined based on the sign of the optimal data flow delay: When the optimal data flow delay is positive, it is determined that the data fluctuation is transmitted from the current monitoring node to the adjacent monitoring node. When the optimal data flow delay is negative, it is determined that the data fluctuation is transmitted from the adjacent monitoring node to the current monitoring node.

[0014] By calculating the propagation entropy, the strength of the causal relationship between adjacent node sequences and the actual propagation time of abnormal signals are quantified.

[0015] Furthermore, the step of generating an adaptive threshold by combining spatial distance compensation includes: The theoretical propagation delay is obtained by dividing the actual physical distance between adjacent monitoring nodes by the theoretical wave velocity of the fluid medium under the current operating conditions; the dispersion coefficient related to the pipe material and fluid characteristics, as well as the inherent system time error determined by the hardware sampling rate and clock synchronization accuracy, are obtained; the product of the dispersion coefficient and the actual physical distance is calculated, and the product is added to the inherent system time error to obtain the adaptive threshold. A compensation mechanism for distance factors and dispersion characteristics is introduced to eliminate near-range missed alarms or long-range false alarms caused by fixed time tolerance.

[0016] Furthermore, the step of executing the judgment logic based on the maximum propagation entropy value, the optimal data flow delay, the theoretical propagation delay, and the adaptive threshold, and outputting the alarm classification result, includes: The maximum transfer entropy value is compared with a preset causal strength threshold. If the maximum transfer entropy value is less than or equal to the causal strength threshold, the current anomaly is determined to be a false alarm caused by sensor failure or isolated external interference, and the primary anomaly flag is revoked. If the maximum transfer entropy value is greater than the causal strength threshold, the physical causality verification step is entered.

[0017] Furthermore, the physical causality verification step includes: The absolute value of the optimal data flow delay is extracted as the actual physical transmission time; the absolute difference between the actual physical transmission time and the theoretical propagation delay is calculated; if the absolute difference is less than or equal to the adaptive threshold, the primary anomaly is confirmed as a real pipeline network anomaly event and a pipe leakage alarm result is output; if the absolute difference is greater than the adaptive threshold, the current anomaly is determined to be simultaneous common-mode interference and the current alarm result is filtered. Dual verification is performed based on the three-dimensional dynamic constraints of the pipeline network space to eliminate false data associations that violate physical wave velocity limits.

[0018] A second aspect of the present invention provides a pipeline leak detection and alarm filtering system based on isolated forests, for performing the above-described method, the system comprising: The data acquisition module is used to collect multi-source sensor data from monitoring nodes and construct high-frequency raw data streams and low-frequency feature streams. The feature construction module is used to construct a low-frequency spatiotemporal feature vector based on the low-frequency feature stream; An anomaly detection module is used to input the low-frequency spatiotemporal feature vector into the isolated forest model to calculate the global anomaly score, compare the global anomaly score with the alarm threshold, maintain the normal monitoring state if it is less than or equal to the alarm threshold, and trigger the primary anomaly identification if it is greater than the alarm threshold. The module also calculates the contribution weight of the feature dimension based on the spatial cutting path of the isolated forest model to extract the core feature set. The data extraction module is used to respond to the primary anomaly identifier, extract data sequences from the high-frequency raw data stream based on the core feature set, and generate a dimensionality reduction matrix; The causal quantization module is used to calculate the propagation entropy of adjacent monitoring nodes based on the dimensionality reduction matrix, and to traverse within the sliding time delay window to obtain the maximum propagation entropy value and the corresponding optimal data flow delay. The alarm verification module is used to calculate the theoretical propagation delay based on the actual physical distance and the theoretical wave speed, generate an adaptive threshold by combining spatial distance compensation, execute judgment logic based on the maximum transmission entropy value, the optimal data flow delay, the theoretical propagation delay and the adaptive threshold, and output the alarm classification result.

[0019] This invention provides a pipeline leak detection and alarm filtering method based on isolated forests. It has the following beneficial effects: 1. This invention constructs a dual-stream processing architecture of high-frequency data stream and low-frequency feature stream, and uses the spatial cutting path of the isolated forest model to extract the core feature set to generate a dimensionality reduction matrix, thereby reducing the computing load on edge computing devices. After detecting a primary anomaly, the algorithm's intrinsic structure is used to lock the data dimension that contributes the most to the anomaly, and redundant dimensions in the high-frequency data source are eliminated in this way. This solves the problem of computing resource bottleneck caused by the full participation of multi-source high-frequency data in the operation in industrial sites, and ensures the real-time response of the system.

[0020] 2. This invention calculates the transfer entropy and optimal data flow delay between adjacent nodes based on a dimension-reduced matrix, and performs physical causality verification by combining actual physical distance and theoretical wave speed, thereby improving the accuracy of pipeline leakage alarms. By combining the statistical causal correlation of data dimensions with the fluid dynamics propagation law in the pipeline space, it can accurately identify and filter isolated false alarms caused by the failure of a single sensor, as well as spatial simultaneous common-mode interference caused by large mechanical resonance or external electromagnetic environment, ensuring that the final output alarm result corresponds to the real abnormal events in the pipeline.

[0021] 3. This invention introduces a dispersion coefficient related to the fluid characteristics of the pipe and the inherent system time error of the hardware, and constructs an adaptive threshold for causality determination by combining the actual physical distance. This enhances the detection reliability of the system under complex working conditions of long-distance pipelines, realizes dynamic adaptive compensation of the time tolerance range with the physical distance, overcomes the time attenuation deviation caused by waveform dispersion and pipe wall friction during the long-distance propagation of abnormal physical waves, and avoids the problems of near-distance missed detection and long-distance false detection caused by the use of fixed time tolerance in traditional methods. Attached Figure Description

[0022] Figure 1 This is a physical architecture diagram of a pipeline leak detection and alarm filtering system based on an isolated forest according to an embodiment of the present invention; Figure 2 This is an overall workflow diagram of a pipeline leak detection and alarm filtering method based on isolated forests according to an embodiment of the present invention; Figure 3 This is a global anomaly score evolution curve diagram of the present invention; Figure 4 This is the transfer entropy time delay spectrum of the present invention; Figure 5 Macroscopic trend diagram for physical causality verification of the present invention; Figure 6 This is a magnified view of the adaptive envelope of the present invention. Detailed Implementation

[0023] The technical solutions in 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.

[0024] See attached document Figure 1 This invention provides a pipeline leak detection and alarm filtering system based on isolated forests, deployed at various monitoring nodes along a long-distance pipeline. The system's underlying hardware topology includes a multi-source sensor group, a time synchronization module, and an edge computing gateway. The multi-source sensor group is installed in the pipeline body and key valve chambers, encompassing pressure transmitters, temperature transmitters, flow meters, acoustic emission sensors, and fiber optic strain gauges. The time synchronization module uses GPS or a precise time protocol to provide unified time synchronization for all monitoring nodes along the line, ensuring that the distributed data sequences have absolute timestamps ranging from nanoseconds to milliseconds. The edge computing gateway communicates with both the multi-source sensor group and the time synchronization module, and is used to execute local data caching, feature extraction, and real-time edge analysis algorithms.

[0025] The system provided by this invention may include: The data acquisition module is used to acquire continuous observation data from a multi-source sensor group at a set sampling frequency, continuously write high-frequency sequences with timestamps into a first-in-first-out circular buffer, and synchronously output the downsampled low-frequency data stream.

[0026] The feature construction module is used to receive low-frequency data streams, calculate the temporal and spatial differential gradients of each data dimension, and concatenate the basic state values ​​and differential gradients to generate low-frequency spatiotemporal feature vectors.

[0027] The anomaly detection module inputs low-frequency spatiotemporal feature vectors into a pre-trained isolated forest model to calculate a global anomaly score. The global anomaly score is compared to an alarm threshold. If the score is less than or equal to the alarm threshold, normal monitoring is maintained. When the anomaly score exceeds the set alarm threshold, a primary anomaly flag is triggered, and the spatial cutting path of the anomaly sample in the isolated tree model is analyzed in reverse. The contribution weights of each data dimension are calculated to lock down the core feature set.

[0028] The data extraction module is used to extract the analysis time window based on the anomaly trigger time after receiving the initial anomaly identifier, generate a channel mask based on the core feature set, extract the high-frequency raw data of the corresponding dimension from the first-in-first-out circular buffer, and generate a dimensionality reduction matrix.

[0029] The causal quantization module is used to receive the dimensionality reduction matrix, calculate the time delay propagation entropy between adjacent monitoring node sequences, and traverse the physical delay space to obtain the maximum propagation entropy and its corresponding optimal time delay.

[0030] The alarm verification module extracts the actual physical distance between adjacent monitoring nodes and the theoretical wave velocity of the pipeline fluid, calculates the theoretical propagation delay, and constructs an adaptive threshold with physical attenuation compensation. It compares the maximum propagation entropy, optimal delay, adaptive threshold, and theoretical propagation delay, uploads the verification parameters to the business scheduling layer, and, combined with the adaptive threshold and physical rule base, performs the final judgment through a multi-branch logical classifier, outputting the classification result to the scheduling terminal.

[0031] See attached document Figure 2 This invention provides a pipeline leak detection and alarm filtering method based on isolated forests, comprising the following steps: Each monitoring node synchronously collects data from multiple sensor sources. A dual-stream processing architecture of high-frequency raw data stream and low-frequency feature stream is established in the memory of the edge computing gateway. Based on the low-frequency feature stream, the temporal and spatial difference gradients between the current node and its neighboring nodes are calculated to construct a low-frequency spatiotemporal feature vector. Low-frequency spatiotemporal feature vectors are input into a pre-trained isolated forest model in real time for processing. Global anomaly scores are calculated. When the anomaly score exceeds the threshold, the path reduction of the current input sample in the branch structure of the isolated forest is analyzed, the contribution weights of each feature dimension are calculated, and the core feature set is extracted by sorting them in descending order of weight. Using the initial anomaly trigger moment as the anchor point, set the time window before and after, issue a data freeze command to the edge computing gateway, and combine the extraction channel determined by the core feature set to extract the data sequence of the specified time period and the specified dimension from the circular buffer of the high-frequency raw data stream to form a dimensionality reduction matrix. Based on the dimension reduction matrix, the propagation entropy of the time difference between high-frequency data sequences of adjacent monitoring nodes is calculated. The search is performed within the set sliding time delay window to obtain the maximum propagation entropy value and the optimal data flow delay that maximizes the propagation entropy. The system obtains the actual physical distance parameters between adjacent nodes and the theoretical wave velocity of the fluid medium under the current operating conditions, calculates the theoretical propagation delay, introduces a spatial distance compensation coefficient to generate an adaptive threshold, and executes multi-branch decision logic based on the relationship between the maximum transmission entropy value, the optimal data flow delay and the adaptive threshold and the theoretical propagation delay. The system outputs classification results for sensor hardware fault interception, normal operating condition fluctuation interception or real pipe leakage alarm.

[0032] In this embodiment, each monitoring node synchronously collects multi-source sensor data and constructs feature vectors in the edge computing gateway.

[0033] First, each monitoring node acquires continuous observation data of the pipeline's operating status through a configured multi-source sensor group. This multi-source sensor group specifically includes pressure transmitters, temperature transmitters, flow meters, acoustic emission sensors, and fiber optic strain gauges installed on the pipe body or key valve chambers. Since the acoustic emission sensors and dynamic pressure transmitters need to capture transient acoustic and negative pressure wave signals, the data sampling rate for the aforementioned continuous observation data is set to 100Hz to 10kHz. Each monitoring node obtains an absolute timestamp through a GPS timing module or a precise time protocol timing module, and appends the absolute timestamp to the continuous observation data to ensure time synchronization of data sequences between adjacent monitoring nodes.

[0034] The edge computing gateway receives continuous observation data with absolute timestamps and constructs a dual-stream processing architecture in memory for both high-frequency raw data streams and low-frequency feature streams. For the high-frequency raw data stream, the edge computing gateway receives the raw data sequence at the aforementioned high sampling rate and continuously writes it into a circular buffer configured in a first-in-first-out (FIFO) mode. The circular buffer has a fixed dwell time window, the length of which is greater than the maximum delay time for the theoretical propagation of fluid wave speeds between adjacent monitoring nodes, typically set to 3 to 10 minutes. When the duration of written data exceeds the dwell time window, newly written data will overwrite the oldest written data. For the low-frequency feature stream, the edge computing gateway performs downsampling or moving average calculations on the high-frequency raw data stream, obtaining the time interval as... Low-frequency data sequences, The value range is set to 1s to 5s, and the low-frequency data sequence is used to reflect the macroscopic operating conditions of the pipeline.

[0035] After acquiring the low-frequency data sequence, this embodiment further calculates the temporal and spatial difference gradients between the current monitoring node and its neighboring monitoring nodes. Let the current monitoring node be denoted as node. The upstream adjacent monitoring node is denoted as node. Downstream adjacent monitoring nodes are denoted as nodes. .node exist The first moment Dimensional observation data is denoted as In the physical operation mechanism of pipelines, time-difference gradients are used to capture transient abrupt changes in local node characteristics caused by leakage or valve operation. The time-difference gradient is defined as the observed data at the current moment. Compared to the previous time interval The difference between observation data at any given time. In the spatial dimension, spatial differential gradients are used to eliminate common-mode interference caused by normal operating conditions such as global pipeline peak shaving, highlighting the spatial differences in the propagation and attenuation of local abnormal signals. Node With upstream nodes The formula for calculating the spatial difference gradient between them is: ; In the formula, Represents the upstream spatial difference gradient. Indicates upstream adjacent nodes exist The first moment Dimensional observation data. Similarly, nodes. With downstream nodes The downstream spatial difference gradient is defined as the difference between the observation data of the current node and the observation data of the downstream neighboring node.

[0036] Then, the node By concatenating the basic state observations, temporal difference gradients, and bidirectional spatial difference gradients across each data dimension, a low-frequency spatiotemporal feature vector is constructed. Let the total dimension of the expanded feature vector be... , The value of is equal to the sum of the dimensions of the basic observation data, the temporal difference gradient, and the two-way spatial difference gradient. The generated low-frequency spatiotemporal feature vector is denoted as . The expression is: ; In the formula, to These represent the various feature components after cascading. Constructing a low-frequency spatiotemporal feature vector fuses the temporal fluctuation characteristics and spatial correlation characteristics of the pipeline network, serving as input data for the anomaly detection model.

[0037] In this embodiment, the edge computing gateway inputs the constructed low-frequency spatiotemporal feature vector into the pre-trained isolated forest model and extracts the core feature dimensions based on the model topology.

[0038] First, the edge computing gateway inputs the current low-frequency spatiotemporal feature vector into the isolated forest model. In the theoretical mechanism of anomaly detection, the isolated forest recursively partitions the data space by randomly selecting features and split values. Normal samples are typically located in data-dense regions, requiring more spatial cuts to be isolated, thus their traversal paths in the tree structure are longer. Conversely, anomaly samples, due to their abnormal feature behavior, are more easily and quickly isolated in shallow branches near the root node, resulting in shorter paths. Based on this principle, after inputting the low-frequency spatiotemporal feature vector, the model calculates the average path length of the feature vector across all isolated trees in the forest and derives the global anomaly score accordingly. The formula for calculating the global anomaly score is: ; In the formula, Indicates time Feature vector Global anomaly score, This represents the average path length traversed by the feature vector across all isolated trees. This indicates the size of the training sample subset used when constructing a single isolated tree. Indicates the number of samples is The average length of unsuccessful search paths in a binary search tree is used to normalize the score.

[0039] After obtaining the global anomaly score, the system compares it with the set alarm threshold. In this embodiment, the alarm threshold is typically set between 0.6 and 0.8, with the specific value determined based on the statistical quantile of the score distribution under historical normal operating conditions. If the global anomaly score is less than or equal to the alarm threshold, it indicates that the pipeline network is within the normal fluctuation range, and the system maintains normal monitoring. When the global anomaly score is greater than the alarm threshold, it indicates that the current pipeline network state deviates from normal operating conditions, and the system triggers a primary anomaly flag. Triggering a primary anomaly flag indicates that a potential abnormal event has been detected, requiring further in-depth analysis of the core cause of the anomaly.

[0040] Upon receiving the initial anomaly identifier, this embodiment does not directly perform causal inference on the full set of multidimensional data. Instead, it utilizes the intrinsic topology of the isolated forest for dynamic feature dimensionality reduction. Specifically, it performs a reverse traversal of the spatial cutting paths of the current anomaly sample within the branch structure of the isolated forest. Since anomaly samples are more easily isolated at shallow levels, the feature dimensions selected at shallow nodes to partition the anomaly samples represent the core physical factors causing the current anomaly. By statistically analyzing the depth and frequency of each data dimension as a splitting node when isolating the current sample, the contribution of each feature dimension to identifying the current anomaly is quantified. The contribution weight calculation formula for each feature dimension is as follows: ; In the formula, Indicates the first Contribution weights of dimensional features This represents the total number of isolated trees in an isolated forest. Indicates the first Among the trees, using the first The hierarchical weight gain corresponding to branching and cutting of dimensional features.

[0041] To establish a quantitative mathematical relationship between weights and path depth, this embodiment defines hierarchical weight gain. It is inversely proportional to the node depth where the feature is invoked. Specifically, if in the... The tree locates the path of the current abnormal sample, the first... 3D features at depth of If a node is selected as the split attribute, then set... If on the corresponding path, the first If a feature is called multiple times, the corresponding reciprocal is incremented; if it is not called, then... The closer a node is to the root of the tree, the deeper it is. The smaller the value, the larger the hierarchical weight gain value.

[0042] The contribution weights of each feature dimension were calculated. Then, the edge computing gateway sorts all feature dimensions in descending order of their weight values. In this embodiment, the system selects the features whose cumulative weight sum reaches a set percentage (e.g., 80% to 90%). Each feature dimension, or directly extracting the top-ranked features by weight. (For example, the feature dimensions ranked from the top 5 to the top 10) are combined to generate a core feature set. Through the aforementioned feature extraction actions, the system isolates redundant dimensions that contribute very little to the current anomaly.

[0043] In this embodiment, after detecting a primary anomaly, the system extracts a dimensionality-reduced matrix from the high-frequency raw data stream for subsequent causal inference through a dual pruning mechanism of time and dimension.

[0044] In this embodiment, when the isolated forest model is When the primary exception flag is triggered, the system will As a time anchor, a data freeze command is synchronously issued to the edge computing gateways of the local and adjacent monitoring nodes. Because high-frequency raw data streams are continuously written to the circular buffer at a high sampling rate, without control, transient high-frequency data before and after an anomaly will be overwritten by subsequently acquired new data. Upon receiving the data freeze command, the edge computing gateway locks the data in the memory circular buffer to... Centered analysis time window High-frequency data segments within the pipeline. In the acoustic propagation mechanism of pipelines, there is a physical time difference in the transmission of pressure waves generated by pipe leakage or valve operation to adjacent upstream and downstream nodes. The moment when the current node triggers a primary anomaly... The timing of the capture of corresponding wave characteristics by adjacent nodes is not completely synchronized. To ensure that the analysis time window can fully cover the transmission process of pressure waves between nodes, This parameter represents the length of the time window extending forward and backward. The value of is determined based on the physical distance between adjacent monitoring nodes and the sound velocity of the fluid medium in the pipeline. The calculation formula is as follows: ; In the formula, This represents the maximum actual physical distance between the current monitoring node and its upstream and downstream adjacent monitoring nodes. This represents the speed of sound of the fluid medium inside the pipeline under the current operating conditions. This represents the redundancy time tolerance introduced to eliminate system communication delays. In this embodiment, The specific value range is usually set to 10s to 60s. By clipping and locking the time dimension, the system focuses its attention on the transient process before and after the anomaly occurs.

[0045] After completing the data locking in the time dimension, this embodiment further combines the extracted core feature set. The system performs dimensionality clipping. High-frequency data from multiple sensors has numerous dimensions; including all of it in subsequent causal quantization would increase the computing load on the edge computing gateway and introduce irrelevant background noise. Therefore, the system uses a core feature set... Generate channel mask vectors. Assume the total dimension of the high-frequency raw data acquired by the system is... The specific implementation of the channel mask vector is that its length is equal to A one-dimensional binary sequence. In the channel mask vector, if a certain data dimension belongs to the core feature set... If the corresponding element in the channel mask vector is 1, then the value of the element at that position is 1; if a certain data dimension does not belong to the core feature set... If the value is 0, then the element at the corresponding position is assigned the value 0.

[0046] After generating the channel mask vector, the edge computing gateway uses the channel mask vector to lock the analysis time window. The high-frequency data within the matrix is ​​extracted using matrix column indexing. Since the element value in the non-core dimension positions of the channel mask vector is 0, the column indexing extraction operation directly deletes the data column corresponding to the element value of 0, retaining only the high-frequency sequence corresponding to the element value of 1, thereby eliminating redundant dimension data irrelevant to this anomaly. The edge computing gateway combines the extracted high-frequency time series of each core dimension to form a dimensionality-reduced matrix. Dimensionality reduction matrix The number of rows corresponds to the number of data sampling points within the analysis time window, and the number of columns corresponds to the core feature set. The system extracts a large number of features. By extracting both time and dimension, the system addresses the computational resource bottleneck between the high-frequency sampling needs of industrial sites and the limited computing power of edge computing gateways, providing a high-frequency data source for subsequently identifying dynamic causal relationships among multiple data sources.

[0047] In this embodiment, after obtaining the dimensionality reduction matrix, the edge computing gateway further quantifies the information transmission direction and transmission time between the high-frequency data sequences of adjacent monitoring nodes that have time delays.

[0048] Because abnormal physical waves generated by pipeline leaks or valve actions require time to propagate within the pipeline network, adjacent monitoring nodes receive the same abnormal signals in a strict temporal order. To quantify causal relationships, this embodiment extracts the high-frequency data sequence of the current monitoring node and the corresponding high-frequency data sequences of adjacent monitoring nodes, focusing on the core feature dimension of the dimensionality reduction matrix. Specifically, let the high-frequency data sequence of the current monitoring node be denoted as... The high-frequency data sequences of adjacent monitoring nodes are denoted as , Represents a sequence exist The sampled value at time 10:00. Represents a sequence exist The sampled value at a given time. The edge computing gateway calculates the high-frequency data sequence. For high-frequency data sequences The time delay propagation entropy. In information theory, the essence of time delay propagation entropy is to measure the propagation entropy of the source sequence. To what extent can historical information reduce the impact on the target sequence? The uncertainty of the current state prediction. If the sequence The change physically led to the sequence Changes in introduce a sequence. Historical data can lead to an entropy reduction effect. In this embodiment, the formula for calculating the time delay propagation entropy is: ; In the formula, The time delay parameter is Time from sequence to sequence The propagation entropy, Represents a sequence exist The sampled value at time 10:00. Represents a sequence exist The sampled value at time 10:00. Represents a sequence exist The sampled value at time 10:00. This represents the joint probability distribution function or the conditional probability distribution function. Indicates to , as well as Summing the state space for all possible discrete values.

[0049] In a physical pipeline network, the order in which anomaly signals arrive at the current node and its neighboring nodes depends on the location of the anomaly. Therefore, this embodiment sets a bidirectional sliding time delay window. Limit time delay parameters The search space. When the optimal data flow is delayed. When the value is positive, it indicates that data fluctuations are propagated from the current node to adjacent nodes; when the optimal data flow is delayed... A negative value indicates that data fluctuations are propagated from adjacent nodes to the current node. Parameter The specific values ​​are determined based on the maximum actual physical distance between adjacent monitoring nodes and the slowest propagation wave velocity of the fluid medium inside the pipeline. The specific calculation formula is as follows: ; In the formula, This represents the actual physical distance between the current monitoring node and a specific neighboring monitoring node that is currently performing causality quantization. This represents the theoretical minimum wave velocity of the fluid medium under the current operating conditions. To determine the specific delay time, the edge computing gateway iterates through the time delay parameters at set step sizes within a bidirectional sliding time delay window. To ensure accurate data indexing in discrete-time series, the set step size is strictly equal to the sampling time interval of the high-frequency data series. The edge computing gateway calculates the corresponding time delay propagation entropy values, obtaining the maximum propagation entropy value and the optimal data flow delay that maximizes the propagation entropy. The mathematical expressions for the maximum propagation entropy value and the optimal data flow delay are as follows: ; In the formula, Indicates the optimal data flow delay. This represents the maximum propagation entropy value. This indicates the search for variable parameters that maximize the objective function. By traversing a bidirectional sliding time delay window, this embodiment not only quantifies the causal correlation strength between abnormal states of adjacent monitoring nodes, but also accurately extracts the actual propagation time and direction of abnormal fluctuations in space.

[0050] In this embodiment, the edge computing gateway performs causality verification based on the obtained optimal data flow delay and maximum transmission entropy value, combined with the network spatial topology, to confirm real anomalies and filter false alarms.

[0051] Specifically, the edge computing gateway will calculate the obtained maximum transfer entropy value. The data is compared with a preset causal strength threshold. The causal strength threshold is used to determine whether there is a substantial information transmission relationship between the high-frequency data sequences of adjacent monitoring nodes. In this embodiment, the specific value of the causal strength threshold is determined based on the statistical mean and standard deviation of the background noise propagation entropy under normal operating conditions. It is usually set to the statistical mean of the propagation entropy under normal operating conditions plus two to three times the standard deviation to exclude small fluctuations in information volume caused by random noise. When the maximum propagation entropy value is less than or equal to the causal strength threshold, it indicates that the data fluctuations of the current monitoring node and the adjacent monitoring nodes are independent of each other. The system determines that the currently triggered primary anomaly is a false alarm caused by local sensor failure or isolated external interference event, and the system directly cancels the primary anomaly label. When the maximum propagation entropy value is greater than the causal strength threshold, it indicates that there is a data correlation between the two nodes, and the system enters the physical causality verification stage.

[0052] In the general principles of causal analysis, the optimal data flow delay derived from transfer entropy only represents a mathematical correlation at the statistical level. To eliminate common-mode interference caused by external factors such as lightning strikes and synchronous vibrations of large machinery, it is necessary to introduce fluid dynamics constraints from the real space to verify whether the delay of data fluctuations on the time axis matches the physical propagation time of sound waves in the three-dimensional space of the pipeline network. In the pipeline network physical system, the waveform propagation generated by actual leaks or valve operations strictly follows the laws of fluid acoustic propagation. The edge computing gateway extracts the actual physical distance between the current monitoring node and adjacent monitoring nodes. and the current sound velocity of the fluid medium in the pipeline Calculate the theoretical propagation delay. The formula for calculating the theoretical propagation delay is: ; In the formula, This represents the theoretical propagation delay of the abnormal physical wave between two monitoring nodes. Due to waveform dispersion, pipe wall friction, and velocity fluctuations during fluid propagation within the pipe, the actual propagation time will deviate, and the magnitude of this deviation increases with the propagation distance. Therefore, this embodiment constructs a distance-adaptive threshold. The formula for calculating the adaptive threshold is: ; In the formula, Indicates an adaptive threshold. This represents the dispersion coefficient related to pipe material and fluid properties. This represents the inherent system time error determined by the hardware sampling rate and the system clock synchronization accuracy. In this embodiment, The value is determined based on the on-site pipe wall friction coefficient and fluid viscosity calibration, and is usually set within a range of 10. -5 s / m to 10 -4 Between s / m; Primarily affected by the timing error of the Global Positioning System (GPS), it is typically set to 1ms to 5ms. This is achieved by incorporating actual physical distance... By dynamically adjusting the adaptive threshold, the system achieves distance-adaptive tolerance, avoiding the problems of near-range missed detections or long-range false detections caused by using a fixed time tolerance.

[0053] After calculating the above parameters, the edge computing gateway extracts the absolute value of the optimal data flow delay. As the actual physical transmission time of abnormal fluctuations between nodes, calculate the actual physical transmission time and the theoretical propagation delay. The absolute difference between them, and the absolute difference with the adaptive threshold. A comparison is performed. The specific logical judgment formula is as follows: ; If the absolute difference is less than or equal to the adaptive threshold, it indicates that the transmission time of abnormal fluctuations between nodes fully conforms to the laws of physical fluid dynamics. In this case, the system confirms the primary anomaly as a real pipeline anomaly event, generates a high-level alarm command, and uploads it to the central control center. The alarm command includes the anomaly occurrence time, causal transmission direction, and related node information. If the absolute difference is greater than the adaptive threshold, it indicates that although data fluctuations occur at both nodes, the time difference violates the physical wave velocity limit. The system determines the corresponding situation to be simultaneous common-mode interference caused by large-scale electromagnetic interference or heavy machinery construction, thereby filtering the current alarm and resetting the monitoring status of the edge computing gateway. Through physical causality classification and alarm filtering, this embodiment eliminates mathematical correlations that do not conform to physical laws, improving the overall accuracy of anomaly detection.

[0054] Specific application examples: To further aid in understanding the technical solution of this invention, a specific application example is given below, taking into account the actual operating conditions of a long-distance natural gas pipeline.

[0055] The target monitoring segment is set to include upstream monitoring node A (denoted as node A). ) and downstream monitoring node B (denoted as node) The actual physical distance between the two nodes. m. The medium transported inside this pipeline section is natural gas, and the actual speed of sound under the current operating conditions is... 400m / s, theoretical minimum wave speed m / s. Dispersion coefficient of the pipe. s / m, inherent timing error of the system s. Communication redundancy time tolerance The system is configured with a raw data sampling rate of 1000Hz, and the time interval of the low-frequency characteristic stream is... s. At a certain moment Mechanical damage to the pipe and a minor leak occurred near node A due to third-party construction work. The system's workflow is as follows.

[0056] At some point Mechanical damage to the pipe and a minor leak occurred near node A due to third-party construction work. The system's workflow is as follows: The first step is for the edge computing gateway at node A to calculate the spatial differential gradient based on the low-frequency feature flow. Formula for generating low-frequency spatiotemporal feature vectors Then input it into the Isolation Forest model. Calculate the global anomaly score according to the formula: ; The model calculations show that the average path length is shortened, resulting in... The score exceeded the set alarm threshold of 0.65, triggering a primary anomaly flag in the system.

[0057] The second step involves the system traversing the spatial cutting paths of outlier samples in the isolated tree model in reverse order, and calculating the contribution weights of each feature dimension according to the formula: ; Statistical analysis revealed that the cumulative weights of the three dimensions—high-frequency acoustic emission components, dynamic pressure differential, and local strain—reached 85%. The system extracted and combined these weights to generate a core feature set. .

[0058] The third step involves the system calculating the length parameter of the time window extending forward and backward based on the formula. : ; The system issues a data freeze command to the edge computing gateway, combined with the core feature set. Generate a channel mask vector and extract it from the circular buffer. Centered on, with a time span of The high-frequency sequences of these three core dimensions are used to generate a dimensionality reduction matrix.

[0059] Fourth step: The system uses a dimensionality reduction matrix to calculate the sequence of node A. With node B sequence The transfer entropy between them. First, calculate the boundary of the bidirectional sliding time delay window according to the formula. : s Then, within the window [-39.47, 39.47], the time delay propagation entropy is calculated iteratively using the following formula, with a step size of 0.001s (i.e., the reciprocal of the 1000Hz sampling rate): ; Obtain the maximum transfer entropy value and its corresponding optimal data flow delay: ; The maximum transit entropy value was calculated. bits (greater than the causality strength threshold of 0.15 bits), and the corresponding optimal data flow delay. s (positive value indicates that data fluctuations are transmitted from node A to node B).

[0060] The fifth step is to perform a physical causality check. This involves calculating the theoretical propagation delay based on the formula. : s; An adaptive threshold with physical attenuation compensation is constructed based on the formula. : ; The system will calculate the actual physical transfer time (i.e. Substituting the absolute value of the theoretical propagation delay into the logical judgment formula: ; The calculated absolute difference is 0.15s. Since 0.15 ≤ 0.302, the condition of the above judgment formula is met, indicating that the transmission time of abnormal fluctuations between nodes strictly conforms to the laws of physical fluid dynamics. The system ultimately confirms that the primary anomaly is a real pipeline network anomaly event, successfully filters out irrelevant common-mode interference, and outputs the pipe leakage alarm result.

[0061] Experimental verification and effect comparison: To verify the actual technical effectiveness of the present invention, offline playback tests were conducted using historical operating data from the aforementioned pipeline over six consecutive months. The test dataset contained 100 abnormal event labels, including 20 actual leakage events, 50 single sensor hardware failures, and 30 instances of synchronous vibration (common-mode interference) during large-scale mechanical construction. The comparison was made with a traditional benchmark method that uses a globally fixed threshold for judgment, calculates the cross-correlation function using full high-frequency data, and performs fixed-time tolerance verification.

[0062] See attached document Figure 3 In the figure, the horizontal axis represents the span before and after the time reference anchor point, and the vertical axis represents the global anomaly score output by the isolated forest. The horizontal dashed line in the figure represents the set alarm threshold (0.65), and the solid line represents the score evolution trajectory calculated over time. As shown in the figure, when the system is in normal operating condition, the solid line remains at a low and stable level; when mechanical damage to the pipe occurs, the score solid line jumps at the anchor point and breaks through the horizontal dashed line, reaching a peak of approximately 0.78, which intuitively demonstrates the timing and sensitivity of the determination of the primary anomaly flag in this invention.

[0063] After triggering the primary anomaly flag and extracting the dimensionality reduction matrix, the system performs causal quantization. (See attached document.) Figure 4 , Figure 4 The graph shows the transfer entropy time delay spectrum calculated within a set bidirectional sliding time delay window. The horizontal axis represents the time delay parameter, and the vertical axis represents the corresponding transfer entropy value. The horizontal dashed line represents the preset causal strength threshold (0.15), the solid line represents the actually calculated transfer entropy spectrum, and the triangle marker represents the vertex of the maximum transfer entropy value. As shown, the solid line forms a single peak at a specific positive delay position. The vertical coordinate (1.25) of this peak's triangle marker far exceeds that of the horizontal dashed line, indicating a substantial information transfer relationship between adjacent nodes. The horizontal coordinate (37.65s) corresponding to this vertex is the optimal data flow delay used to extract the actual physical time consumption.

[0064] To verify whether the data transmission time described above conforms to the laws of fluid physics, the system performs a physical causality check. (See attached document.) Figure 5 In the figure, the horizontal axis represents the actual physical distance within the spatial range of 0 to 20000m, and the vertical axis represents the time delay. The solid line in the figure represents the macroscopic linear baseline trend of the theoretical propagation delay, and the square markers represent the actual measured scatter points (15000m, 37.65s) calculated in the example. As shown in the figure, the actual measured scatter points closely follow the theoretical baseline solid line on a macroscopic scale.

[0065] Further refer to the appendix Figure 6 In the figure, the solid line in the center represents the theoretical propagation reference (37.50s), the long dashed line above represents the upper boundary of the envelope (+0.302s), and the dotted line below represents the lower boundary of the envelope (-0.302s). The upper and lower boundaries together constitute the legal physical error space, and the circular markers represent the measured scatter points. As shown in the figure, the circular marker (37.65s) falls accurately within the envelope formed by the long dashed line and the dotted line, which conforms to the laws of fluid dynamics propagation. Based on this, the system confirms that the anomaly is a real pipeline leak.

Claims

1. An isolated forest-based pipeline leak detection and alarm filtering method, characterized in that, Includes the following steps: Collect multi-source sensor data from monitoring nodes, construct a high-frequency raw data stream and a low-frequency feature stream, and construct a low-frequency spatiotemporal feature vector based on the low-frequency feature stream; The low-frequency spatiotemporal feature vector is input into the isolated forest model to calculate the global anomaly score. The global anomaly score is compared with the alarm threshold. If it is less than or equal to the alarm threshold, the normal monitoring state is maintained. If it is greater than the alarm threshold, the primary anomaly identification is triggered. The contribution weight of the feature dimension is calculated based on the spatial cutting path of the isolated forest model to extract the core feature set. In response to the primary anomaly identifier, a data sequence is extracted from the high-frequency raw data stream based on the core feature set to generate a dimensionality reduction matrix; Based on the reduced-dimensional matrix, the transmission entropy of adjacent monitoring nodes is calculated, and the maximum transmission entropy value and the corresponding optimal data flow delay are obtained by traversing within the sliding time delay window. The theoretical propagation delay is calculated based on the actual physical distance and theoretical wave speed. An adaptive threshold is generated by combining spatial distance compensation. The judgment logic is executed based on the maximum transmission entropy value, the optimal data flow delay, the theoretical propagation delay, and the adaptive threshold, and the alarm classification result is output.

2. The pipeline leakage detection and alarm filtering method based on isolation forest according to claim 1, characterized in that, The steps for constructing a low-frequency spatiotemporal feature vector based on the low-frequency feature stream include: Calculate the temporal difference gradient and bidirectional spatial difference gradient between the current monitoring node and its neighboring monitoring nodes; The low-frequency spatiotemporal feature vector is constructed by concatenating the basic state observations, the temporal difference gradient, and the bidirectional spatial difference gradient.

3. The pipeline leakage detection and alarm filtering method based on isolation forest of claim 1, wherein, The steps for calculating the contribution weights of feature dimensions based on the spatial cutting paths of the isolated forest model, and for extracting the core feature set, include: The spatial cutting path of the abnormal samples within the isolated forest model is traversed in reverse to determine the hierarchical depth of each feature dimension as the split node. Define a hierarchy weight gain that is inversely proportional to the hierarchy depth, and calculate the contribution weight of each feature dimension based on the hierarchy weight gain; All feature dimensions are sorted in descending order according to their contribution weights, and feature dimensions that meet the set ratios are selected and combined to generate the core feature set.

4. The method of claim 1, wherein, The steps of extracting data sequences from the high-frequency raw data stream based on the core feature set and generating a dimensionality-reduced matrix include: The analysis time window is set based on the time when the primary anomaly flag is triggered. The length of the analysis time window is determined by the ratio of the maximum actual physical distance between adjacent monitoring nodes to the sound speed of the fluid medium plus the communication redundancy time tolerance. Generate a binary channel mask vector based on the core feature set; The high-frequency raw data stream within the analysis time window is subjected to column index extraction using the channel mask vector. Redundant data that does not belong to the core dimension is removed, and the resulting data is combined to generate the dimensionality reduction matrix.

5. The method of claim 1, wherein, The steps for obtaining the maximum propagation entropy value and the corresponding optimal data flow delay by iterating through the sliding time delay window include: The sliding time delay window is set as a bidirectional sliding time delay window, and the boundary of the bidirectional sliding time delay window is determined by dividing the actual physical distance between adjacent monitoring nodes by the theoretical minimum wave velocity. According to the set time step, the high-frequency data sequence of adjacent monitoring nodes is gradually shifted within the bidirectional sliding time delay window and the transmission entropy under each offset is calculated. The maximum transmission entropy value is obtained by comparison, and the time offset corresponding to the maximum transmission entropy value is extracted as the optimal data flow delay. The direction of fluctuation transmission is determined based on the sign of the optimal data flow delay: When the optimal data flow delay is positive, it is determined that the data fluctuation is transmitted from the current monitoring node to the adjacent monitoring node. When the optimal data flow delay is negative, it is determined that the data fluctuation is transmitted from the adjacent monitoring node to the current monitoring node.

6. The method of claim 1, wherein the method further comprises: The steps for generating an adaptive threshold using spatial distance compensation include: The theoretical propagation delay is obtained by dividing the actual physical distance between adjacent monitoring nodes by the theoretical wave velocity of the fluid medium under the current operating conditions. Obtain the dispersion coefficients related to pipe and fluid properties, as well as the inherent system time error determined by the hardware sampling rate and clock synchronization accuracy; The product of the dispersion coefficient and the actual physical distance is calculated, and the product is added to the inherent system time error to obtain the adaptive threshold.

7. The method of claim 1, wherein the method further comprises: The steps for executing the judgment logic and outputting the alarm classification result based on the maximum propagation entropy value, the optimal data flow delay, the theoretical propagation delay, and the adaptive threshold include: The maximum transit entropy value is compared with a preset causal strength threshold. If the maximum transfer entropy value is less than or equal to the causal strength threshold, the current anomaly is determined to be a false alarm caused by sensor failure or isolated external interference, and the primary anomaly identifier is revoked. If the maximum propagation entropy value is greater than the causal strength threshold, then the physical causality verification process begins.

8. The method according to claim 7, wherein, The physical causality verification process includes: Extract the absolute value of the optimal data flow delay as the actual physical transmission time; Calculate the absolute difference between the actual physical transmission time and the theoretical propagation delay; If the absolute difference is less than or equal to the adaptive threshold, the primary anomaly is confirmed as a real pipeline network anomaly event and a pipe leakage alarm result is output. If the absolute difference is greater than the adaptive threshold, the current anomaly is determined to be simultaneous common-mode interference and the current alarm result is filtered.

9. The method of claim 1, wherein the method further comprises: The steps to construct the high-frequency raw data stream and the low-frequency feature stream include: An absolute timestamp is added to the collected multi-source sensor data through a time synchronization module; The data sequence with an absolute timestamp is continuously written into a first-in-first-out (FIFO) circular buffer to construct the high-frequency raw data stream; The high-frequency raw data stream is downsampled or a moving average is calculated to obtain a low-frequency data sequence, which is used to construct the low-frequency feature stream.

10. An Isolation Forest based pipeline leak detection and alarm filtering system, characterized in that, A pipeline leak detection and alarm filtering method based on isolated forests as described in any one of claims 1-9, comprising: The data acquisition module is used to collect multi-source sensor data from monitoring nodes and construct high-frequency raw data streams and low-frequency feature streams. The feature construction module is used to construct a low-frequency spatiotemporal feature vector based on the low-frequency feature stream; An anomaly detection module is used to input the low-frequency spatiotemporal feature vector into the isolated forest model to calculate the global anomaly score, compare the global anomaly score with the alarm threshold, maintain the normal monitoring state if it is less than or equal to the alarm threshold, and trigger the primary anomaly identification if it is greater than the alarm threshold. It also calculates the contribution weight of the feature dimension based on the spatial cutting path of the isolated forest model to extract the core feature set. The data extraction module is used to respond to the primary anomaly identifier, extract data sequences from the high-frequency raw data stream based on the core feature set, and generate a dimensionality reduction matrix; The causal quantization module is used to calculate the propagation entropy of adjacent monitoring nodes based on the dimensionality reduction matrix, and to traverse within the sliding time delay window to obtain the maximum propagation entropy value and the corresponding optimal data flow delay. The alarm verification module is used to calculate the theoretical propagation delay based on the actual physical distance and the theoretical wave speed, generate an adaptive threshold by combining spatial distance compensation, execute judgment logic based on the maximum transmission entropy value, the optimal data flow delay, the theoretical propagation delay and the adaptive threshold, and output the alarm classification result.