A railway fire-fighting pipe network global intelligent leakage detection operation and maintenance system and method

Abstract

This invention belongs to the field of railway intelligent fire protection and pipeline multi-source leakage detection technology, and discloses a comprehensive intelligent leak detection and maintenance system and method for railway fire protection pipeline networks. The invention adopts a three-level hardware architecture: a station-level perception layer, a station-section-level regional aggregation layer, and a group-company-level global control layer, with a supporting execution linkage layer. The perception layer deploys IEEE1588PTP sub-millisecond synchronous multimodal acquisition units to simultaneously acquire pressure, buried acoustic signatures, and inlet / outlet flow time-series signals. The edge end incorporates convolutional FastICA blind source separation, BO-LightGBM leak identification, multi-source topology adaptive fusion positioning, and flow steady-state dual-criteria modules, dynamically adjusting the positioning dimension weights, using dual-condition judgment to reduce false alarms, and is equipped with differentiated maintenance processes for track maintenance windows and train operations. This invention is adaptable to railway scenarios with strong interference and multi-branch pipeline networks, achieving a micro-leakage identification accuracy of ≥97.4%, a normal positioning error of 0.91m, and local operation compliance with railway network security standards at the edge end. It can be widely applied to online monitoring of buried fire protection pipeline networks in various railway stations.

Description

Technical Field

[0001] This invention relates to the fields of intelligent pipeline leak detection, railway smart fire protection IoT, edge signal processing and machine learning, and is particularly applicable to the full-area online leak monitoring and hierarchical closed-loop operation and maintenance of underground fire water supply pipelines in passenger stations, marshalling yards, freight yards, locomotive depots, rolling stock depots, dynamic inspection stations and sections under the jurisdiction of various bureaus of China State Railway Group. Specifically, it is a full-area intelligent leak detection and operation and maintenance system and method for railway fire protection pipelines. Background Technology

[0002] The railway system's fire protection pipeline network is characterized by numerous points, long lines, dispersed layout, hierarchical management, and significant responsibilities. A large railway hub typically includes dozens of independent areas such as passenger stations, EMU depots, freight yards, locomotive depots, oil depots, and material warehouses. The pipeline network exhibits a complex topology combining star-shaped radiation and linear connections. Railway fire protection management is implemented under a three-tiered responsibility system: station, depot, and group company. Each level has significantly different management authority, technical capabilities, and maintenance resources. Existing municipal pipeline leak detection equipment and single acoustic leak listening devices have multiple technical shortcomings that make them unsuitable for railway pipeline network scenarios. First, their interference suppression capabilities are weak. The continuous vibration of train wheels and rails along the railway line, the 50Hz power frequency noise of traction power supply, and the intermittent disturbances of water pumps all overlap, making it impossible for conventional single filtering methods to isolate weak leak characteristics, resulting in a high false alarm rate. Furthermore, the broadband vibrations generated when trains pass by highly overlap with the frequency bands of minute leak sound patterns, making it difficult for traditional frequency domain filtering to effectively distinguish them. Secondly, it cannot effectively handle the problem of concurrent leaks at multiple points in multi-branch pipe networks. The fixed-weight fusion algorithm commonly used in municipal systems does not consider the negative pressure wave reflection offset caused by tees and elbows. After multiple leak signals are superimposed, they cannot be separated and located. However, the number of tees and elbows in railway station pipe networks is far greater than that of municipal direct-buried pipes, and the positioning error of a single negative pressure wave can reach several meters. Thirdly, it does not match the time-sharing water supply characteristics of railways. Railways have two completely different flow conditions: no water use during nighttime maintenance windows and intermittent fire hydrant water release during the day. Fixed flow windows will continuously generate false alarms. Fourthly, the operation and maintenance architecture is not compatible with the railway's three-level management system. Existing equipment can only achieve single-layer local alarms and cannot connect to the three-level fire control process of stations, sections, and groups. There is a lack of linkage mechanism for special maintenance work orders during maintenance windows. Fifthly, network security does not meet the requirements of railway dedicated networks. Data from general equipment is uploaded to the external network, which does not comply with the hard specifications of data isolation and local encrypted storage for railway critical information infrastructure. Sixthly, hydraulic and acoustic parameters are set only based on engineering experience and lack complete calculation support from national standards and metrological regulations. The technical materials required for fire protection and metrological acceptance are incomplete. Among the existing publicly available technologies, some solutions can only complete the collection of tunnel fire protection data and do not have the ability of AI multi-source positioning and full-process operation and maintenance closed loop. Some solutions only use basic machine learning algorithms and have not solved the core pain points of multiple interferences, adaptive weights and blind source noise reduction in railways. None of them can fully adapt to the complex working conditions unique to railway fire protection. Summary of the Invention

[0003] This application provides a comprehensive intelligent leak detection and maintenance system and method for railway fire protection pipelines, aiming to solve one of the following existing technical problems:

[0004] First, there is the problem of weak interference suppression capability. The continuous vibration of train wheels and rails along the railway line, the 50Hz power frequency noise from traction power supply, and the intermittent disturbances from water pumps all overlap, making it difficult for conventional single-channel filters to isolate weak leakage characteristics, resulting in a high false alarm rate. The 20-2000Hz broadband vibration generated when trains pass by highly overlaps with the 100-800Hz frequency band of the sound signature of minor leaks, making it difficult for traditional frequency domain filters to effectively distinguish them. Second, there is the problem of poor positioning accuracy in multi-branch pipe networks. The municipal fixed-weight fusion algorithm does not consider the negative pressure wave reflection offset caused by tees and elbows, making it impossible to separate and locate multiple leakage signals due to their superposition. The number of tees and elbows in railway station pipe networks is far greater than that of municipal direct-buried pipes, and the positioning error of a single negative pressure wave can reach 5-10m. Third, there is the problem of poor adaptability to different time-of-day operating conditions. Railways have two completely different flow conditions: no water use during nighttime maintenance windows and intermittent fire hydrant water release during the day. Fixed flow windows continuously generate false alarms. Fourth, there is the issue of mismatched operation and maintenance architecture. Existing equipment only supports single-layer local alarms and cannot connect to the three-level fire control process of stations, sections, and the group, lacking a linkage mechanism for special maintenance work orders during track maintenance windows. Fifth, there is the issue of insufficient network security compliance. Data from general equipment is uploaded to the external network, which does not comply with the strict specifications for data isolation and local encrypted storage of critical railway information infrastructure. Sixth, there is the issue of insufficient technical basis. Hydraulic and acoustic parameters are set only based on engineering experience, without complete calculation support from national standards and metrological regulations, and fire protection and metrological acceptance materials are missing.

[0005] To address the aforementioned issues and achieve the aforementioned objectives, the technical solution adopted in this application is as follows:

[0006] The hardware of this invention is divided into four layers, from bottom to top: station-level perception layer, station / section-level regional aggregation layer, group company-level global control layer, and a supporting independent execution linkage layer. Specifically, each layer is described in detail below:

[0007] Station-level perception layer: Deploys PTP synchronous multimodal sensing units and ARM Cortex-A7 edge computing terminals. The multimodal sensing unit integrates pressure sensing components (sampling frequency 80Hz–200Hz, reference 100Hz), buried acoustic signature sensing components (sampling frequency 800Hz–2000Hz, reference 1000Hz), inlet / outlet ultrasonic flow sensing components, and an IEEE 1588 PTP synchronous clock with synchronization time difference. <0.001s. The edge computing terminal completes signal preprocessing and AI inference locally, encrypts and stores the collected raw data locally, and does not upload it to the public network. The edge computing terminal has four built-in core algorithm modules: Convolutional FastICA blind source separation module: It judges the degree of multipath convolution of the signal by the multipath delay sampling parameter Nd. When Nd is greater than the set threshold of 0.1, it enables the M-order convolution compensation operation, where M is 8 to 16. It adopts the frequency domain convolution FastICA scheme: First, the time domain observation signal is transformed to the frequency domain through short-time Fourier transform. The convolution mixture of each frequency point is approximated as instantaneous mixture. FastICA separation is performed independently for each frequency point. Finally, the pure leakage feature signal in the time domain is obtained through inverse short-time Fourier transform, realizing the effective separation of train vibration and power frequency mixed interference. BO-LightGBM leakage identification module: It takes pressure time series, sound pattern time frequency, and flow characteristics as input and outputs leakage probability values ​​in the range of 0 to 1, with a preset optimal probability judgment threshold. The model uses a balanced sampling method to process unbalanced leakage samples. Multi-source topology adaptive fusion positioning module: Calculates the leak point coordinates based on negative pressure wave, acoustic cross-correlation, and flow residual correction. It dynamically normalizes the three positioning weights based on the number of pipeline branches and real-time vibration intensity, outputting the final fused leak point location. It includes two error calculation modes for extreme safety verification and normal performance evaluation, respectively. Flow steady-state leakage dual-criteria module: Sets a leakage flow threshold based on instrument measurement error, outputting a valid leak alarm only when both the leakage probability and flow residual conditions are simultaneously met.

[0008] Station-level regional aggregation layer: Deploys regional servers, lightweight BO-LightGBM diagnostic modules, and work order lifecycle management modules to store sensing data from each station, with a typical storage period of ≥3 years, and generates electronic work orders for inspection and maintenance.

[0009] Group-level overall control layer: Build a pipeline network digital twin visualization platform, a global leakage trend prediction engine, and a fire protection comprehensive assessment report module to perform global leakage hotspot analysis and medium- and long-term leakage risk prediction.

[0010] Execution linkage layer: includes the zoned electric valve PLC control unit and the trenchless repair work order generation and file update unit.

[0011] The complete technical chain of the core algorithm of this invention is as follows: in the order of the technical chain, it is: synchronous multimodal acquisition (IEEE1588PTP time alignment), calculation of time delay sampling parameter Nd and adaptive switching of convolution / instantaneous ICA noise reduction, time-frequency feature extraction, BO-LightGBM leakage probability output, parallel solution of three types of localization, specifically including negative pressure wave, acoustic cross-correlation and flow residual, topology adaptive weight fusion, flow-probability dual criterion alarm, time-segmented hierarchical operation and maintenance, adaptive parameter periodic update, and finally, fallback identification of abnormal pipeline network conditions.

[0012] The key mathematical model used in this invention is the convolutional FastICA blind source separation model, and the mixed observation signals acquired by the sensor can be represented as a linear instantaneous mixing model: ;in This is a mixed observation signal acquired synchronously through m channels. It is a mixed matrix (full column rank). It is an independent source vector containing leakage source signals, train vibration noise, power frequency interference, and water pump disturbance. The noise is zero-mean Gaussian white noise. When multipath propagation exists, the observed signal follows a convolutional mixture model: Where M is the convolution order, Let m be the mixing matrix of the m-th order time delay. The source signal vector is a time-delayed sampling point m. The convolutional mixing is transformed into an instantaneous mixing of each frequency point through short-time Fourier transform, and then separated frequency point by frequency point through the FastICA algorithm.

[0013] The formula for calculating the time delay sampling parameter Nd is: Where L is the sensor spacing (m). The speed of sound wave propagation (m / s). The sampling frequency is Hz. M-order convolutional ICA compensation is enabled when Nd > 0.1; instantaneous ICA model is used when Nd ≤ 0.1.

[0014] The FastICA iterative update formula is: Iterative convergence condition: .

[0015] And the BO-LightGBM leak classification identification model:

[0016] Output of a single decision tree: ;

[0017] Overall predictions from the ensemble model: ;

[0018] Leakage probability (Sigmoid mapping): ;

[0019] Model hyperparameters The Gaussian process surrogate model was determined through Bayesian optimization.

[0020] Desired improvements to the acquisition function:

[0021] Optimal Leakage Detection Threshold Determined by maximizing the Youden exponent of the ROC curve in the training set: J(p) = TPR(p) + TNR(p) - 1

[0022] Three-level leakage classification rules: For early warning of micro-leakage; This is a general leak alarm; Emergency alert for a major leak.

[0023] The Joukowsky negative pressure wave velocity formula with water temperature correction used in this invention is as follows: ;

[0024] Where K(T) is the bulk modulus (Pa) of water at temperature T. Let T be the density of water at temperature T (kg / m³), D be the inner diameter of the pipe (m), E be the elastic modulus of the pipe (Pa), and e be the wall thickness (m).

[0025] Dual-ended sensor negative pressure wave positioning formula:

[0026]

[0027] Where L is the total length of the pipeline (m). , The time (s) when the negative pressure wave reaches the sensors at both ends.

[0028] Voiceprint cross-correlation time delay localization model:

[0029] Cross-correlation function of two voiceprint signals: ;

[0030] The peak cross-correlation corresponds to the optimal propagation delay. Voiceprint distance calculation: .

[0031] Multi-source topology adaptive fusion localization model:

[0032] Fusion formula:

[0033] Weight normalization constraint: ,

[0034] Adaptive weight calculation formula:

[0035]

[0036] Baseline weight adjustment rules: Topology dimension: When there are no tee branches (straight pipes) within the monitoring pipe segment. When the number of branches in the tee is ≥3 When there are 1 to 2 branches ;

[0037] Interference dimension: During the quiet period of the track opening (no train traffic, background noise sound pressure level < 60dB) During periods of strong train interference (background noise sound pressure level ≥ 80dB), When there is moderate interference ;

[0038] Metrological dimension: When the instrument has passed metrological verification and the error is within the allowable range. When the instrument drifts out of tolerance or the error exceeds twice the allowable value .

[0039] Background noise sound pressure level is collected and calculated in real time by buried acoustic sensors, and train traffic status can be obtained through the railway traffic dispatching system interface.

[0040] Flow residual correction distance: ;

[0041] in The residual for zonal flow balancing (m³ / s). L represents the rated design flow rate of the pipe section (m³ / s) and L represents the length of the monitoring pipe section (m). This is the flow-distance calibration coefficient, which physically represents the correction coefficient for the proportion of leakage distance corresponding to the proportion of residual flow per unit flow rate. It is calibrated through on-site artificial leakage simulation tests, with a typical value range of 0.6~1.0, 0.8 for straight pipe sections, and 0.65 for multi-branch pipe sections.

[0042] Dual-criteria model for flow balancing:

[0043] Partition mass conservation:

[0044] Flow balance residual:

[0045] Leakage detection threshold: Only when and At that time, it was determined to be a genuine leak.

[0046] This invention provides two error calculation modes, as follows:

[0047] Linear superposition mode (ultimate safety check): ;

[0048] Root mean square synthesis mode (normal performance evaluation): ;

[0049] Beneficial effects:

[0050] 1. This invention, through the deep integration of an adaptive blind source separation algorithm, a topology-aware multi-source fusion positioning mechanism, and a three-tiered railway operation and maintenance architecture, achieves high-precision, low-false-alarm, and highly compliant full-domain intelligent leak detection and maintenance for the unique scenarios of strong interference, multiple branches, and time-segmented management of railway fire protection pipe networks. It possesses several significant advantages. Firstly, it has a strong anti-interference advantage. The convolutional blind source separation algorithm does not require pre-collection of noise samples, and its micro-leakage identification rate under continuous train vibration conditions is no less than 95%, significantly outperforming traditional single-filter and fixed-weight algorithms.

[0051] 2. The adaptive fusion mechanism adopted in this invention can dynamically correct the positioning error caused by tees and elbows, and can simultaneously identify two or more concurrent leaks in a pipe section, solving the problem that existing municipal equipment cannot handle multi-point leaks.

[0052] 3. This invention also has the advantage of being adaptable to all scenarios. Specifically, the adaptive flow window can match different water supply conditions during peak hours and daytime, significantly reducing false alarms generated during normal pipeline operation. Fourthly, it has railway compliance advantages. The complete system of four-layer architecture design, local encrypted storage, three-level work order flow, and annual metering verification fully meets the multiple management regulations of railway fire protection, dedicated network, and metering.

[0053] 4. This invention also boasts the advantage of low-cost edge deployment. By quantifying all artificial intelligence and fluid algorithms, it can run in real time on ARM edge terminals without cloud computing power support. The standardization and development of the equipment is simple, and the deployment cost is controllable. Sixthly, it offers a safety redundancy advantage. An automatic switching fallback judgment logic is set up for abnormal conditions such as airbag and pipeline emptying, avoiding missed detections caused by the failure of a single hydraulic model. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a topology diagram of the three-level full-area leak detection system for railway fire protection pipelines.

[0056] Figure 2 Hardware wiring diagram for PTP synchronization of multimodal sensors;

[0057] Figure 3 Here is a complete flowchart of FastICA signal processing;

[0058] Figure 4 A block diagram for adaptive weight adjustment in multi-source fusion. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0060] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0061] Example 1:

[0062] This embodiment uses a large passenger hub station of a railway bureau as an application scenario. The station's fire protection pipeline network covers 12 independent zones, including the station building, platforms, freight yard, and locomotive maintenance area, with a total length of 18.7 km. It primarily uses DN100 galvanized steel pipes and features numerous tees, bends, and branches. Furthermore, it is constantly affected by train vibrations and traction power supply frequency interference. This system strictly matches the three-tiered railway fire protection management architecture of station-section-group company. The overall solution is as follows:

[0063] Station-level sensing layer: The entire station is deployed in zones, with a total of 36 multimodal synchronous acquisition units and 6 ARM Cortex-A7 edge computing terminals. Each acquisition unit integrates a 0.2-level pressure sensor component with a sampling frequency of 100Hz, a buried MEMS acoustic signature sensor component with a sampling frequency of 1000Hz, an ultrasonic flow sensor component, and an IEEE 1588PTP synchronous clock, enabling simultaneous acquisition of multiple signals with the maximum time difference. <0.001s; In this embodiment, the synchronization time difference introduced by the PTP synchronization clock is converted into distance error and included as a component error of the system hardware in the overall positioning error evaluation system.

[0064] Each edge terminal runs all algorithms locally and has four core modules built-in:

[0065] Convolutional FastICA blind source separation module: The preset calculation delay sampling parameter Nd automatically enables M-order convolution compensation operation when Nd>0.1. It uses tanh nonlinear mapping to iteratively separate the leakage effective signal from train vibration and 50Hz power frequency mixed interference without the need to collect noise samples in advance.

[0066] BO-LightGBM Leakage Identification Module: Input 128-dimensional pressure time series, acoustic frequency, and flow characteristics, output leakage probability in the range of 0 to 1, use SMOTE balanced sampling to process micro-leakage unbalanced samples, and set the optimal judgment threshold to 0.7.

[0067] The multi-source topology adaptive fusion positioning module: It solves in parallel for the coordinates of the leak point due to negative pressure wave, the coordinates of the leak point due to acoustic cross-correlation, and the flow residual correction. Based on the number of tees in the pipeline network and the real-time vibration intensity, it dynamically normalizes three types of positioning weights and outputs the final leak point location. It incorporates two error calculation modes: ultimate safety verification and normal performance evaluation. In this multi-source topology adaptive fusion positioning module, the propagation velocity of the negative pressure wave adopts the Joukowsky fluid calculation formula with water temperature correction.

[0068]

[0069] Where T is the water temperature. Let T be the bulk modulus of water at temperature T. Let T be the density of water at temperature T, D be the inner diameter of the pipe, E be the elastic modulus of the pipe, and e be the wall thickness of the pipe.

[0070] The attenuation of buried acoustic signals was modeled using a bivariate exponential fitting model of soil cover thickness and soil moisture content.

[0071] ;

[0072] Where H is the cover thickness, φ is the soil volumetric water content, a and b are the soil attenuation fitting coefficients, and R0 is the model fit goodness R0. 2 ≥0.94.

[0073] The steady-state leakage dual-criteria module sets a leakage flow threshold based on the flow meter measurement error. It outputs a valid leakage alarm only when the leakage probability and the flow residual simultaneously meet the judgment conditions.

[0074] All raw pipeline data is encrypted using national cryptographic algorithms and stored locally. It is only aggregated to higher-level networks through the railway intranet and is not transmitted to the public network.

[0075] Station-level regional aggregation layer: Deploy regional servers and a work order lifecycle management system to store all station perception data within the jurisdiction for a period of ≥3 years. Equipped with a lightweight BO-LightGBM diagnostic module for secondary verification, it automatically generates electronic work orders for inspection and maintenance, realizing electronic management and control of the entire process of hazard dispatch, handling, and acceptance.

[0076] Group-level overall control layer: Build a digital twin visualization platform for the pipeline network, restore the pipeline network topology and equipment locations of the entire railway bureau in a 1:1 ratio, and have functions such as global leakage hotspot statistics, medium and long-term leakage risk prediction, and automatic generation of fire protection operation and maintenance assessment reports to support global preventive maintenance decisions.

[0077] The execution linkage layer is configured with a zoned electric valve PLC control unit and a trenchless repair work order generation unit. When a major leakage alarm is triggered, the corresponding pipe section valve can be shut down in conjunction with the system, and trenchless repair work orders and file update processes adapted to railway track maintenance windows can be generated simultaneously. The system sets two sets of equipment availability assessment indicators: one for train operation periods and one for the overall annual availability. The system availability during train operation periods is no less than 99.9%, and the overall annual availability is no less than 99.5%.

[0078] Brief Description of Working Principle: This system operates on the core logic of "edge local computation, three-level hierarchical control, and multi-source complementary verification." All sensors achieve time alignment through PTP sub-millisecond clocks, eliminating positioning deviations caused by time misalignment at the hardware level. This provides a unified time reference for negative pressure wave time difference positioning and acoustic cross-correlation positioning, thus completing the foundation for synchronous perception. Secondly, the raw data first enters the convolutional FastICA module, which automatically switches between instantaneous and convolutional ICA modes based on the propagation delay corresponding to the pipe section length. This removes mixed noise such as train vibration and power frequency interference, purifying the leakage characteristic signal. Subsequently, time-frequency features are extracted and fed into the BO-LightGBM model to output the leakage probability, while simultaneously calculating three types of positioning results in parallel. The optimal leak point coordinates are obtained by dynamically allocating weights based on the number of pipe network branches and real-time vibration intensity. Finally, false alarms are filtered through dual-condition verification of flow residual and leakage probability, completing edge intelligent processing. Then, the station level is responsible for local data collection, local early warning, and on-site verification; the section level is responsible for regional data aggregation, AI verification, and work order scheduling; and the group level is responsible for overall situation assessment, big data analysis, and standard control. Data converges from the bottom up, while instructions are issued from the top down, perfectly aligning with existing railway management processes and achieving the technical objective of a three-tiered collaborative architecture. Finally, when a large area of ​​air pockets or abnormal emptying is detected in the pipeline network, the system automatically blocks the negative pressure wave positioning channel, retaining only acoustic signature and flow rate as dual criteria. This avoids missed detections due to the failure of a single hydraulic model, achieving the goal of providing a backup for abnormal operating conditions.

[0079] Example 2:

[0080] This embodiment is based on the full-domain intelligent leak detection and maintenance system provided in Embodiment 1 above. It is adapted to two typical operating conditions: railway maintenance window and daytime train operation. It realizes intelligent management and control of the entire process from data collection to maintenance closed loop. The complete steps are as follows:

[0081] S1 multimodal synchronous acquisition is started. The system uses the IEEE1588PTP unified clock reference to synchronously trigger all pressure, acoustic fingerprint and flow sensors to collect data. The pressure sampling frequency is 100Hz, the acoustic fingerprint sampling frequency is 1000Hz, and the synchronization time difference of multiple signals is controlled within 0.001s. All data are uploaded to the station edge computing terminal after being stamped with high precision.

[0082] S2 edge terminal ICA signal preprocessing: The edge terminal first calculates the time delay sampling parameters. Where L is the sensor spacing (m). The speed of sound wave propagation (m / s). Where is the sampling frequency (Hz). When When >0.1, enable M-order convolution ICA compensation; when When the value is ≤0.1, an instantaneous ICA model is used; the FastICA iterative algorithm is used to separate three types of noise: train vibration, power frequency interference, and water pump disturbance, and outputs a clean leakage characteristic signal; the iterative convergence threshold is set to .

[0083] S3 feature extraction and leakage probability calculation extracts 128-dimensional features from the purified signal, including time domain, frequency domain, statistics, and valve position status. These features are then input into the pre-trained BO-LightGBM model, which outputs real-time leakage probability values ​​in the 0-1 range. The model hyperparameters are determined through Bayesian optimization, with the optimal decision threshold being 0.7.

[0084] S4 multiphysics localization parallel solution and fusion: The system solves three types of localization results in parallel: Negative pressure wave localization: The wave velocity is calculated using the Joukowsky formula with real-time water temperature correction, and the leak point coordinates are calculated using the time difference of dual-end sensors. Acoustic cross-correlation localization: Based on the soil cover thickness and soil moisture content, the acoustic attenuation coefficient is corrected, and the time delay is obtained by calculating the cross-correlation peak value of the two acoustic signature signals to calculate the coordinates of the leak point. Traffic residual correction: Based on the traffic balance residual of the partition, the positioning correction amount is calculated. The final leak location was then calculated using the adaptive weighted fusion formula: Among them, weight , , satisfy The weighting is determined by the number of pipeline branches, real-time vibration intensity, and instrument error status through dynamic normalization. When there are no branches in the straight pipe, the weighting of negative pressure waves is increased, and when there is strong train interference, the weighting of flow rate is increased and the weighting of sound signature is decreased.

[0085] This step is accompanied by two error evaluation modes: a linear superposition mode. Perform limit safety upper limit verification using the root mean square synthesis mode (normal performance evaluation): Perform routine accuracy assessments. Among them, The positioning error introduced into blind source separation signal processing. The residual error introduced by multi-source weighted fusion The hardware error is introduced by the PTP clock synchronization time difference. If the system detects a large area of ​​water accumulation in the pipeline network or abnormal pipeline emptying, it automatically shuts down the negative pressure wave positioning logic and retains only the soundprint and flow rate as backup identification methods.

[0086] S5 dual-condition leakage detection simultaneously verifies two conditions: ① The leakage probability output by BO-LightGBM is ≥0.7; ② The absolute value of the partition flow residual exceeds the leakage flow threshold. A leak is considered valid only if both conditions are met simultaneously.

[0087] The steady-state flow calculation window in this step is adaptively switched: a 1-second baseline window is used for nighttime skylight water discharge conditions; a 0.5-second short window is used for daytime sporadic water discharge conditions; and the flow residual determination is paused for continuous large-flow water discharge, such as fire hydrant tests and pipeline water replenishment.

[0088] S6 hierarchical alarm and work order distribution classifies alarms into three levels based on leakage probability: 0.7≤p<0.8 is a micro-leakage warning, 0.8≤p<0.9 is a general leakage alarm, and p≥0.9 is a major leakage emergency alarm. Alarm information is pushed to the corresponding level according to permissions. During maintenance windows, fixed-point maintenance work orders are directly issued. Major leaks trigger the linkage control of zone valves simultaneously.

[0089] The S7 system periodically and adaptively updates parameters. It updates the negative pressure wave velocity calculation parameters daily based on real-time water temperature and the acoustic attenuation coefficient monthly based on groundwater content monitoring data. It also completes the mandatory verification of all sensing equipment annually according to metrological regulations and updates the instrument error threshold and leakage judgment threshold simultaneously to ensure the long-term operational accuracy of the system.

[0090] The method provided in this embodiment uses "signal purification - intelligent identification - multi-source fusion - dual verification - hierarchical handling - iterative optimization" as its core logic. It performs end-to-end adaptive optimization for the unique operating conditions of railways, including time-segmented operation, strong interference, and multiple branches. Specifically, it uses convolutional FastICA blind source separation to remove strong interference at the signal input, solving the problem that traditional filtering cannot handle overlapping frequency band noise, ensuring effective identification of minor leaks even during train passage. Multi-source fusion is used to offset the defects of single models. The physical mechanisms of the three positioning methods—negative pressure wave, acoustic signature, and flow rate—are independent. Dynamic weights are used to adapt to different operating conditions, leveraging the advantages of negative pressure wave's long propagation, high acoustic signature accuracy, and high flow rate reliability while avoiding the limitations of their respective application boundaries. Dual criteria and adaptive window pressure drop false alarms are employed. Probability and flow rate dual verification eliminate false alarms caused by single signal fluctuations from a mechanistic perspective. Combined with time-segmented flow rate windows matching railway water usage patterns, invalid alarms are significantly reduced. For negative pressure wave failure scenarios such as airbags and air evacuation, the identification path is automatically switched to avoid the risk of systemic missed detections caused by single model failure. By dynamically correcting environmental parameters and conducting annual metrological calibration, the accuracy degradation caused by seasonal changes and instrument drift is offset.

[0091] Example 3:

[0092] This embodiment, based on any of the above embodiments, provides an embodiment for hardware acquisition and timing synchronization parameter calculation. Specifically:

[0093] 1. Pressure sampling frequency: The effective frequency band of the negative pressure wave hydraulic transient signal is 0–40 Hz. According to the Nyquist sampling theorem, Hz. 100Hz is selected, with a 20Hz redundancy reserved to eliminate 50Hz power frequency aliasing interference. This selection is based on GB / T15478-2015 "Performance Test Methods for Pressure Sensors".

[0094] 2. PTP synchronization time difference constraint: The maximum allowable positioning error is 1m, and the conservative wave velocity is 1000m / s. Reverse calculation:

[0095]

[0096] Engineering values <0.001s, hardware clock synchronization can be stably achieved.

[0097] 3. Acoustic sampling frequency: The acoustic signal of buried pipelines attenuates significantly above 400Hz, with an effective signal-to-noise ratio bandwidth of 50–400Hz. According to the Nyquist theorem... The 1000Hz reference setting is selected to suit various backfill sites. The 500Hz setting is only a field-selectable setting and is only applicable to clay-sand mixed backfill soil. This selection constraint is only described in this embodiment and does not constitute a limitation on the scope of protection of this invention.

[0098] 4. ICA Delay Discrimination: Delay Sampling Parameter Calculation Formula: ;

[0099] Where L is the sensor spacing (m). The speed of sound wave propagation (m / s). The speaker sampling frequency (Hz). When >0.1, enable M-order convolution ICA compensation; when For values ​​≤0.1, the instantaneous ICA model is used. Special note: Used to determine the propagation delay of acquired signals, ICA (Internal Correlation Analysis) reduces noise in multiple mixed signals at the same measurement point and does not participate in the cross-correlation localization calculation between two points. Example: L=10m branch pipe, m / s, Hz, Automatically enable convolution compensation.

[0100] Example 4:

[0101] This embodiment fully illustrates the operation flow and parameter definitions of the blind source separation module:

[0102] Signal whitening preprocessing: The mixed observation signal x(t) is centered and whitened to obtain the decorrelated signal z, so that the variance of each component of the signal is 1 and they are uncorrelated, reducing the complexity of subsequent iterative solutions.

[0103] Delay determination: Calculate Nd according to the formula in Example 1, and select the algorithm mode according to the threshold: for short branches with small delay, instantaneous ICA can be used to reduce computing power consumption; for long pipe sections with large delay, convolutional ICA is enabled to ensure separation accuracy.

[0104] Iterative solution: Initialize the separation vector w, update it according to the FastICA iterative formula, and normalize w after each iteration.

[0105] Convergence criterion: When the difference in the L2 norm of the separating vectors before and after iteration is less than 1 / 2... When convergence is reached, the pure leakage characteristic signal after separation is output.

[0106] Source signal sorting: Based on the spectral characteristics of the signal and the prior leakage frequency range, the effective signals corresponding to the leakage are screened from the separated multiple independent sources, and the three types of noise sources, namely train vibration, power frequency interference and water pump disturbance, are eliminated.

[0107] Example 5:

[0108] This embodiment provides a complete explanation of the BO-LightGBM leakage identification model, specifically including:

[0109] 1. Dataset Construction

[0110] The simulated pipeline dataset contains 500 sets, divided into training, validation, and test sets in a 7:2:1 ratio. The SMOTE algorithm was used to balance the micro-leakage samples: when training directly using the imbalanced dataset, the F1 score was 0.82, and after balancing, the F1 score was 0.96.

[0111] 2.128-dimensional time-frequency feature extraction

[0112] Four types of features, totaling 128 dimensions, were extracted from the clean leakage signal after ICA separation:

[0113] Time-domain features: mean, variance, kurtosis, peak factor, waveform factor, etc., approximately 20 dimensions;

[0114] Frequency domain characteristics: MFCC coefficients, power spectral density, spectral centroid, etc., approximately 60 dimensions;

[0115] Statistical characteristics: higher-order moments, information entropy, fractal dimension, etc., approximately 30 dimensions;

[0116] Valve position status characteristics: approximately 18 dimensions.

[0117] 3. Model hyperparameters and inference performance

[0118] After Bayesian optimization and grid search verification, the total number of decision trees T=300 and the maximum depth of a single tree depth=8. At this time, the F1-score of the validation set is the highest (0.96). When the number of trees is further increased, the F1-score improves by less than 0.01 but the inference time increases by 30%.

[0119] The ARM Cortex-A7 platform's INT8 quantization inference takes 287ms per iteration, meeting the requirements for real-time alarms.

[0120] 4. Determination of the optimal leakage probability threshold

[0121] The optimal threshold is determined by maximizing the Youden exponent of the ROC curve in the training set: traversing all boundary thresholds p∈[0,1], the maximum value of J(p) corresponds to p=0.7. Verification: when p=0.6, TNR=0.91, the false positive rate increases; when p=0.8, TPR=0.89, the false negative rate increases. Therefore, 0.7 is selected as the basic decision threshold.

[0122] Example 6:

[0123] This embodiment provides a detailed explanation of the mathematical model for Joukowsky negative pressure waves and acoustic attenuation. The water temperature-corrected wave velocity formula is calculated as follows: The Joukowsky wave velocity formula with water temperature correction is: Step-by-step calculation for a DN100 galvanized steel pipe operating at 20℃ (normal temperature): Known parameters Pa, kg / m³, D=0.1m, Pa, e = 0.004m:

[0124] The numerator of the sound velocity in unconstrained water: m / s

[0125] Pipe wall modification items:

[0126] The denominator as a whole:

[0127] Final wave speed: a = 1483.2 / 1.126 = 1317 m / s

[0128] A correction formula for water modulus and density at temperatures ranging from 5 to 30℃ is included to eliminate wave velocity calculation biases caused by seasonal water temperature variations. A bivariate soil cover attenuation model is employed, with buried acoustic signal attenuation using a bivariate exponential fit between soil cover thickness and soil moisture content. Where H is the cover thickness (m), and φ is the soil volumetric water content (0–1). Model fit goodness. =0.94. The attenuation coefficient can be recalibrated on-site for different backfill soil types. Example calculation: H=2m, φ=0.15, substituting, we get... That is, the sound amplitude value is reduced to 80.2% of its original value.

[0129] Example 7:

[0130] This embodiment, based on Embodiment 1, provides a detailed explanation of multi-source fusion and dual-mode error calculation.

[0131] Fusion positioning calculation: Employing the core formula of fusion positioning: Weight normalization constraint .

[0132] For example: taking the baseline weight =0.4, =1.2, =0.6, the total is 2.2;

[0133] The calculation yields: ≈0.182, ≈0.545, ≈0.273, the sum of the three is strictly equal to 1; assuming =150.5m, =150.2m, =150.8m, substituting the values, the final positioning coordinates are approximately 150.38m.

[0134] Two error calculation modes: linear superposition mode, i.e., limit safety check:

[0135]

[0136] Root mean square synthesis mode, i.e., normal performance evaluation:

[0137]

[0138] in, For ICA signal separation error, For multi-source fusion residuals, Hardware errors introduced by the synchronization clock.

[0139] Example 8:

[0140] This embodiment, based on Embodiment 1, elaborates on the dual criteria for flow rate and the adaptive window. According to JJG1030-2007 "Verification Procedure for Ultrasonic Flowmeters", the comprehensive error of the flowmeter is... Leakage detection threshold:

[0141]

[0142] DN100 main pipeline m³ / s, m³ / s = 0.8 L / s.

[0143] Adaptive window logic:

[0144] Window segmentation function rule: No water usage window condition: Use a 1s baseline window, at which the flow fluctuation is minimal and the detection sensitivity is highest;

[0145] Sporadic water discharge during the day: A 0.5s short time window is used to quickly track changes in flow rate and reduce false alarms caused by normal water discharge;

[0146] Continuous high-flow-rate operation: Suspend flow residual determination to avoid false alarms triggered by high-flow-rate disturbances.

[0147] Example 9:

[0148] This embodiment, based on Embodiment 1, elaborates on the fallback logic for extreme pipeline network conditions as follows: When there is a large area of ​​water-filled air pockets, pipe empties, or large cracks in the pipeline network, the negative pressure wave propagation path is disrupted, and the hydraulic model fails. At this time, the system automatically shields the negative pressure wave positioning output, retaining only the dual-condition discrimination of acoustic signature characteristics and flow residual, avoiding missed detections due to the failure of a single hydraulic model. Air pocket condition judgment conditions: The pressure signal exhibits continuous irregular oscillations, and the negative pressure wave propagation duration exceeds the theoretical upper limit by more than 30%.

[0149] Pipeline purging criteria: The inlet and outlet flow rates are continuously 0 and the pipeline pressure is below 0.05 MPa for more than 5 minutes.

[0150] The fallback logic is triggered by the following conditions: the pressure signal exhibits continuous irregular oscillations, the negative pressure wave propagation time exceeds the theoretical upper limit by more than 30%, and the flow residual continues to increase without a corresponding step change in pressure.

[0151] Example 10:

[0152] This embodiment describes the artificially simulated leakage verification test based on Embodiment 1.

[0153] Test conditions: DN100 galvanized steel pipe, total length 300m, buried at a depth of 2m with clay cover, water temperature 20℃. Leakage gradient: five levels of leakage from trace to moderate, 0.5, 1, 2, 5, and 10 L / min.

[0154] Test results: Accuracy of identifying minute leaks: 97.4%; average error of normal positioning: 0.91m; maximum error: 1.48m < 1.5m; railway safety red line; recognition rate of train vibration simulation conditions: 95.1%.

[0155] All sensing and gateway hardware in this invention uses commercially available, standardized, and mature electronic components, such as pressure transmitters, ultrasonic flow meters, MEMS acoustic sensors, and ARM Cortex-A7 edge computing gateways, eliminating the need for customized special equipment. The entire algorithm can be lightweightly deployed on general-purpose ARM embedded chips using C / C++ language, with a single inference time of less than 300ms after INT8 quantization. Performance verification is completed through laboratory simulations of railway fire protection pipe networks and artificial gradient leakage simulation tests. The technical solution can be proven to be fully manufactured and stably operated without the need for on-site mass production. It can be widely adapted to all buried fire protection pipe network renovation and construction projects in railway bureaus across the country, including stations, marshalling yards, freight yards, locomotive depots, and rolling stock depots, and has the value for large-scale industrial promotion.

[0156] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A comprehensive intelligent leak detection and maintenance system for railway fire protection pipelines, characterized in that, It includes a station-level perception layer, a station-section-level regional aggregation layer, a group company-level global control layer, and an execution linkage layer; The station-level sensing layer deploys a multimodal synchronous acquisition unit and an edge computing terminal. The multimodal synchronous acquisition unit integrates pressure sensing components, buried acoustic signature sensing components, inlet and outlet ultrasonic flow sensing components, and an IEEE 1588PTP synchronous clock. The pressure signal sampling frequency is 80Hz–200Hz, and the acoustic signature signal sampling frequency is 800Hz–2000Hz. The maximum allowable time difference for simultaneous acquisition of multiple signals is [not specified]. <0.001s; The edge computing terminal completes signal preprocessing and AI inference locally, and the collected raw pipeline data is encrypted and stored locally without being transmitted to the external network; The edge computing terminal has a built-in convolutional FastICA blind source separation module, a BO-LightGBM leakage identification module, a multi-source topology adaptive fusion localization module, and a traffic steady-state leakage dual criterion module; The convolutional FastICA blind source separation module is used to calculate the multipath delay sampling parameter Nd of the multi-channel signal at the same measurement point, and switches between instantaneous ICA or M-order convolutional compensation operation based on the comparison result of Nd and the set threshold. When there is a multipath effect of buried sound propagation, convolutional compensation is enabled, and a nonlinear mapping function is used to iteratively separate the effective signal of pipeline leakage from train vibration and power frequency mixed interference. The BO-LightGBM leak identification module is used to extract pressure time series, acoustic frequency and flow characteristics and input them into the model, and output the leak probability value in the range of 0 to 1. The model uses a balanced sampling method to process unbalanced leak samples. The multi-source topology adaptive fusion positioning module is used to calculate the negative pressure wave leak point coordinates, acoustic cross-correlation leak point coordinates and flow residual correction amount respectively. Based on the number of pipeline branches and real-time vibration intensity, the three types of positioning weights are dynamically normalized, and the final fusion leak point location is output. The steady-state leakage dual criterion module is used to set the leakage flow threshold by combining the instrument measurement error, and outputs a valid leakage alarm only when the leakage probability and the flow residual simultaneously meet the judgment conditions. The station-level regional aggregation layer is used to store the sensing data of each station and generate electronic work orders for inspection and maintenance. The group-level global control layer is used to build a pipeline network digital twin visualization platform to perform global leakage hotspot analysis and medium- and long-term leakage risk prediction. The execution linkage layer includes a zoned valve linkage control unit and a trenchless repair work order generation unit.

2. The intelligent leak detection and maintenance system for the entire railway fire protection pipeline network according to claim 1, characterized in that, In the multi-source topology adaptive fusion positioning module, the propagation velocity of the negative pressure wave is calculated using the Joukowsky fluid dynamics formula with water temperature correction. Where T is the water temperature. Let T be the bulk modulus of water at temperature T. Let T be the density of water at temperature T, D be the inner diameter of the pipe, E be the elastic modulus of the pipe, and e be the wall thickness of the pipe.

3. The intelligent leak detection and maintenance system for the entire railway fire protection pipeline network according to claim 1, characterized in that, In the multi-source topology adaptive fusion positioning module, the attenuation of buried acoustic signals adopts a bivariate exponential fitting model of soil cover thickness and soil moisture content: Where H is the cover thickness, φ is the soil volumetric water content, a and b are the soil attenuation fitting coefficients, and the model fit goodness R is... 2 ≥0.

94.

4. The intelligent leak detection and maintenance system for the entire railway fire protection pipeline network according to claim 1, characterized in that, The synchronization time difference introduced by the PTP synchronization clock is converted into distance error and included as a component error of the system hardware in the overall positioning error evaluation system.

5. The intelligent leak detection and maintenance system for the entire railway fire protection pipeline network according to claim 1, characterized in that, The system sets two sets of equipment availability assessment indicators: one for operating hours and one for the whole year. The system availability during operating hours shall not be less than 99.9%, and the overall availability throughout the year shall not be less than 99.5%.

6. A method for intelligent leak detection and maintenance of railway fire protection pipelines based on the system described in any one of claims 1 to 5, characterized in that, Includes the following steps: S1: Synchronous acquisition starts, synchronously acquiring pipeline pressure, buried acoustic signature, and inlet / outlet flow time-series data through the PTP unified clock; S2: Edge signal preprocessing, calculate the time delay sampling parameter Nd, determine whether to enable convolution compensation based on the threshold, and separate mixed interference to obtain a pure leakage characteristic signal; S3: Feature extraction and leakage probability calculation. Extract time-frequency features from the separated signal, input them into the BO-LightGBM model, and output the real-time leakage probability. S4: Parallel solution of multi-physics field localization, calculates the negative pressure wave localization coordinates, acoustic cross-correlation localization coordinates and flow residual correction value respectively, and adaptively weighted fusion to obtain the final coordinates of the leak point; S5: Dual-condition leakage detection, simultaneously verifying whether the leakage probability and flow residual exceed their respective thresholds, generating a valid leakage alarm only when both conditions are met; S6: Tiered alarm and work order distribution, classifying alarm levels into three levels based on leakage probability, and issuing fixed-point maintenance work orders during maintenance windows; S7: Periodic parameter adaptive update, updating wave velocity and acoustic attenuation calculation parameters based on real-time water temperature and groundwater content, and completing the statutory mandatory verification of metering and sensing equipment and updating the instrument error threshold annually.

7. The intelligent leak detection and maintenance method for the entire railway fire protection pipeline network according to claim 6, characterized in that, In step S4, the adaptive weighted fusion calculates the final leak location according to the following formula: ;in Locating the coordinates of the negative pressure wave. For acoustic cross-correlation positioning coordinates, For flow residual correction; weighting coefficient , , satisfy The value is determined by the dynamic normalization of the number of pipeline branches and the real-time vibration intensity: ； in , , These are the baseline weights for the negative pressure wave, acoustic, and flow positioning dimensions, respectively.

8. The intelligent leak detection and maintenance method for the entire railway fire protection pipeline network according to claim 6, characterized in that, Two systems are configured to calculate the total positioning error: the first is a linear superposition mode. The first type is used for safety upper limit verification under extreme operating conditions; the second type is the root mean square synthesis mode. Used for accuracy assessment of pipeline network routine operation; among which, The positioning error introduced into blind source separation signal processing. The residual error introduced by multi-source weighted fusion Hardware error introduced by PTP clock synchronization time difference.

9. The intelligent leak detection and maintenance method for the entire railway fire protection pipeline network according to claim 6, characterized in that, Adaptive switching between steady-state flow calculation window and judgment threshold: In the case of no water usage window, a 1-second reference window and standard leakage threshold are used to ensure detection sensitivity; in the case of sporadic water discharge during the day, a 2-second long-term smoothing window and a relaxed leakage threshold are used to suppress false alarms caused by normal water usage fluctuations; in the case of continuous large flow discharge, the flow residual judgment is suspended and reactivated after the flow returns to steady state; the judgment condition for continuous large flow discharge is: the inlet and outlet flow exceeds 30% of the rated flow of the pipe section for 10 seconds.

10. The intelligent leak detection and maintenance method for the entire railway fire protection pipeline network according to claim 6, characterized in that, When a large area of ​​water-filled air pockets is detected in the pipeline network, the negative pressure wave positioning logic is automatically shut down, and only the soundprint and flow rate dual criteria are retained as a backup leak identification method; when an abnormal pipeline emptying condition is detected, all leak detection logic is automatically suspended, and automatically resumed after the pipeline network is filled with water.

Images