A multi-source data fusion underground pipe network risk investigation system

Abstract

The application discloses a multi-source data fusion underground pipe network risk investigation system, which is suitable for industrial parks. Without relying on satellite remote sensing, the system collects pressure, flow, acoustic, settlement, strain and humidity data by using Internet of Things sensors, inputs the data into a risk fusion analysis module after edge-side reordering and abnormal buffering, calculates a comprehensive risk index according to a "smoothing-robust normalization-sub-index and topological characteristics-weighted fusion and quantile threshold" process and gives early warning, and adopts bandpass and pre-whitening cross-correlation to realize positioning in a leakage scene. The system introduces online weight learning, sensor drift self-calibration and active sampling gain arrangement, improves long-term robustness and resource efficiency, and realizes early discovery, accurate positioning and closed-loop disposal of leakage, collapse and blockage risks.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring and operation and maintenance management of underground infrastructure pipelines in industrial parks, and in particular to a multi-source data fusion system for risk assessment of underground pipelines. Background Technology

[0002] Industrial parks have diverse underground pipe networks that are frequently upgraded and operate under complex conditions. Existing and newly added pipelines coexist, and signals such as pressure, flow, soil, and vibration exhibit strong non-stationary and cross-source asynchronous characteristics. Traditional methods rely heavily on manual inspections or single-threshold alarms, making it difficult to detect minor leaks, gradual blockages, and structural collapses induced by foundation problems in a timely manner. Furthermore, network jitter leading to out-of-order arrival, seasonal and load variations causing distribution drift, and inconsistencies in the dimensions and noise levels of different sensors all amplify the risks of false alarms and missed alarms. Existing systems generally lack synchronous reordering and robust normalization of multi-source data, lack structural priors that integrate historical assets and graph topology, and lack online self-calibration weight update mechanisms that can adapt to changes in field distribution. Therefore, it is difficult to simultaneously achieve the comprehensive goals of "early detection, accurate location, interpretability, and closed-loop control." Especially when relying solely on locally available data sources without the constraints of satellite remote sensing, a systematic technical approach is needed, encompassing sampling, preprocessing, feature construction, fusion discrimination, positioning, and verification, to achieve highly reliable monitoring under conditions of limited bandwidth and computing power. Summary of the Invention

[0003] This invention proposes a multi-source data fusion underground pipeline network risk assessment system. The goal is to achieve early detection, accurate location, interpretability, and closed-loop management of three typical risks—leakage, collapse, and blockage—based solely on locally available sensor information and historical asset data, without relying on satellite remote sensing. The overall system structure consists of IoT sensor node units (S), a data transmission module (N), a historical database module (HDB), and a risk fusion analysis module (RFA). Unit S continuously collects pressure values ​​at key points in the pipeline network. Flow value Acoustic signal amplitude Settlement amount is Pipe wall strain and soil moisture Multi-source time-series data; N, after performing arrival reordering and anomaly buffering at the edge, is stably fed into RFA; HDB, with its long-term accumulated equipment ledgers, operating trajectories, accident records, maintenance logs, geological structures, and graphical model topology, provides statistical benchmarks and structural priors for RFA. RFA uses a fusion model as its core to calculate the comprehensive risk index. , ; Among them, the spatial location index is The time index is The leakage risk sub-index is: The collapse risk sub-index is The congestion risk sub-index is The non-negative fusion weight is ,satisfy and ;when Triggering an alert, including risk thresholds These are configurable parameters.

[0004] The present invention adopts the following processing sequence in the process arrangement: "arrival reordering → exponential sliding smoothing → robust normalization → sub-index and topological feature calculation → weighted fusion and threshold discrimination → (leakage scenario) cross-correlation localization → closed-loop verification" to ensure that the input and output interfaces of each step are verifiable and replaceable.

[0005] During the data acquisition phase, S enables an adaptive sampling strategy for each monitoring point. ; Among them, the baseline sampling rate is The adjustment coefficient is Sliding window variance operator In length Calculation within a time window, spatial location index Time Index The amount detected at the location is expressed as , , For in position The detected value u over all time periods. This formula uses the sliding window variance. To be self-incrementing, reflecting the intuition that "the greater the fluctuation, the more frequent the sampling should be": the optimal benchmark sampling rate is selected. Hz (adjustable within the range of 1–5 Hz), adjustment coefficient (Can be selected within the range of 0.2–2.0), Window length Seconds (selectable within 10–120 seconds). In the heterogeneous network environment of a complex industrial park, a timestamp buffer queue is configured on the N side. and maximum tolerable jitter Arrival reordering is performed on multi-source data to ensure that the same fusion batch meets the requirements. , Indicate time, preferred The parameter can be adjusted within the range of 200–500 ms; to balance real-time performance and completeness, an upper limit for late data is also set. Late samples within this limit are still accepted and trigger a local recalculation of the batch, thereby avoiding the systematic bias of the fusion result caused by "soft packet loss" due to transient link congestion.

[0006] Before entering the RFA (Radio Frequency Analysis), the system performs consistent and robust preprocessing on all sensor sequences. First, exponential sliding smoothing is employed. , For time index t The exponential moving average, For time index time The exponential moving average, preferred (Can be finely adjusted within 0.2–0.4) The sampling interval is the current sampling interval or an integer multiple thereof; then robust normalization is performed. , The robust normalized result of u at time index t and spatial location index x. Let be the exponential moving average of u at time index t and spatial location index x, where The median, The median absolute deviation is calculated by HDB under segmented statistical conditions (seasonal / shift / load) to prevent "spurious fluctuations" caused by the mixing of different operating conditions. From this point on, all subsequent calculations are performed uniformly in the Z-domain (i.e., using...). (Replacing the original quantity) not only breaks down the dimensional barriers across sensors, but also significantly reduces interference from extreme values ​​and slow drift.

[0007] At the risk modeling level, this invention provides engineering definitions for three sub-indices: leakage, collapse, and blockage, supplemented by topology enhancement terms to reflect structural priors. The leakage sub-indices employ a dual-source fusion of pressure mutation and acoustic energy. ; Short-time acoustic energy m is the window length; Total window length , Z A For time index Robust normalization result of pressure value A at spatial location index x; Positive part operator Weight and Preferred (Parameters can be adjusted between 0.6–0.8 and 0.4–0.2), short-time energy window Optimal value 32 (parameters can be adjusted within the range of 16–64). Sampling interval. Collapse sub-index fusion of structural response and foundation environment: ; Among them, preferred , Robust normalization result for ε at time index t and spatial location index x. Robust normalization result of Sa at time index t and spatial location index x. Let H be the exponential moving average of H at time index t and spatial location index x. Humidity normalization , Statistics are compiled by HDB, segmented by season and soil type. The blockage sub-index is combined with pressure differential and flow rate attenuation. ; In the formula, For the robust normalized result of F at time index t and spatial location index x, the preferred value is... (Parameters can be adjusted within the ranges of 0.5–0.8 and 0.2–0.5). Meanwhile, to reflect the engineering fact that "blockages have path dependence and upstream cascading," RFA utilizes the HDB's network topology. Calculate topology-weighted path features ; in For location index The set of reachable paths For path Average pressure difference on the edge, path weight Based on a weighted average of pipe diameter, material, and historical failure rate; preferred selection. As The addition of enhancements increases the weight. (Parameters can be adjusted within the range of 0–0.4) to improve sensitivity to "hidden upstream sedimentation".

[0008] In terms of alarm strategy, this invention adopts a quantile adaptive threshold. ,in For HDB history The sample set, The sample quantile operator is defined as follows: , Indicator functions, For the infimum operator, To scan candidate quantile values, the preferred method is... To balance stability and sensitivity, the preferred option is... (0.97 is preferred for steady-state conditions, and 0.90 is preferred for sensitive conditions). Fusion weights. Online updates with simplex projection ; in, The weight at time t+1, Let be the weight at time t. Standard probability simplex Projection operator on, loss function: Sigmoid function event tags The learning rate was obtained from verification of known alarms. Preferred (can be found) (Internal selection). This learning mechanism ensures that the system can "self-calibrate" as the on-site statistical distribution slowly evolves, avoiding the accumulation of bias caused by long-term fixed weights.

[0009] When the fusion criterion indicates a leak, the system automatically enters the location process: first, it checks... Apply 1–5 kHz bandpass and pre-whitening to suppress narrowband noise and pipe reverberation, then calculate the cross-correlation. ; Obtain the time delay difference , For time index Spatial location index Robust normalization results at point A For time index Spatial location index The robust normalization result for point A, and used with Estimate the distance from the leak point to the i-th sensor, where For sensor spacing, Equivalent sound velocity; preferably in an aqueous medium. =1480 m / s (calibrated within 1350–1550 m / s), calibrated in a gaseous environment through controlled-leak tests or historical event regression. The range of values.

[0010] To improve the long-term robustness and resource efficiency of the system, this invention preferably introduces two closed-loop engineering details. First, sensor drift self-calibration: RFA continuously maintains the steady-state baseline at each point. and When the long window (Preferred adjustment factor) This is identified as slow drift, and the normalized reference is updated without reducing the extraction rate. The new normalized reference is based on the coefficient. Perform exponential sliding updates; secondly, active sampling gain orchestration: when a certain pipe section It is in the historical percentile (preferably ≥0.9) and When the trend is upward, the relevant resistance value will be... Flow value The sampling rate is multiplied by the gain coefficient g (preferably 2, which can be adjusted between 1.5 and 3), and symmetrical downsampling is performed on non-hotspot segments to maintain the total bandwidth conservation, thereby prioritizing communication and computing resources to areas that are "more likely to have problems".

[0011] Pressure and flow nodes should be deployed at the junctions of main inlets / outlets and branches, with a preferred spacing of 300 m (adjustable within 200–500 m); acoustic nodes should be deployed more densely in sections with a high incidence of historical accidents, with a preferred spacing of 150 m (adjustable within 100–200 m); settlement and strain nodes should be deployed more densely in sections crossing roads, in soft soil, and in large-scale excavation and backfill areas; humidity nodes should be deployed in areas adjacent to drainage and chemical pipe corridors. Edge gateways are recommended to have a local cache of ≥8 GB and to have TLS link encryption and device-level certificate authentication; HDBs are recommended to adopt a hybrid architecture of "time-series database + relational database + graph database," as shown in the official website topology. Generated jointly by GIS and BIM and automatically aligned and updated after maintenance. To ensure consistency across multiple sources, all device times are aligned to PTP or GPS timing, with a maximum allowable clock drift preferably ≤100 ms.

[0012] The system is not only statistically adaptive, structurally interpretable, and engineeringally closed-loop, but also maintains a low false alarm rate, good localization, and resource-friendly performance in long-term operation and complex working conditions.

[0013] Beneficial effects This invention offers the following comprehensive effects and engineering advantages: First, by using edge-side timestamp buffering and maximum jitter constraints to achieve cross-source alignment, it significantly reduces misjudgments caused by out-of-order processing. Exponential smoothing and robust normalization based on median / median absolute deviation unify the dimensions and suppress extreme values ​​and slow drift, enabling the fusion of different sensors under the same statistical scale. Second, the three sub-indices of leakage, collapse, and blockage weave mechanistic characteristics such as pressure mutations, acoustic energy, strain and settlement, pressure differentials, and flow attenuation into calculable indicators. Furthermore, it introduces topologically weighted path features based on pipeline maps, enhancing the perception of upstream cascading siltation and structurally vulnerable paths, thus providing stronger interpretability. Third, by employing quantile adaptive thresholds and online weight learning with simplex projection, the system can automatically adapt to seasonal / shift / load changes, maintaining low false alarms and low false negatives over the long term. Fourth, in leakage scenarios, by combining bandpass and pre-whitening cross-correlation localization, it can achieve quantifiable distance estimation and practical on-site handling guidance in complex noise and reverberation environments. In summary, this invention, without relying on satellite remote sensing, achieves early detection, accurate location, mechanism explanation, and closed-loop operation and maintenance of three typical risks through operable data sources and closed-loop methods, significantly improving the inherent safety level and operation and maintenance efficiency of underground pipe networks in industrial parks. Attached Figure Description

[0014] Figure 1 The present invention provides a flowchart of the execution steps of a multi-source data fusion underground pipeline network risk assessment system. Figure 2 This invention presents a working diagram of a multi-source data fusion underground pipeline network risk assessment system. Detailed Implementation

[0015] This embodiment takes a mixed industrial park of chemical and equipment manufacturing as an example, and provides a directly deployable implementation plan for the multi-source data fusion underground pipeline network risk assessment system of the present invention. The system is as follows: Figure 2 As shown, the system consists of IoT sensor node units S, data transmission module N, historical database module HDB, and risk fusion analysis module RFA. It operates in the following sequence: "arrival reordering → exponential sliding smoothing → robust normalization → sub-index and topological feature calculation → weighted fusion and quantile threshold discrimination → (leakage scenario) cross-correlation localization → closed-loop verification." Ultimately, it achieves early detection, accurate localization, and interpretable closed-loop management of three types of risks: leakage, collapse, and blockage. The execution steps are as follows... Figure 1 As shown.

[0016] Step 1: Equipment deployment and initialization. Install S devices at key nodes of the underground pipeline network for water supply and drainage, process water, and gas in the park to continuously collect pressure values. Flow value Acoustic signal amplitude Settlement amount is Pipe wall strain and soil moisture Multi-source time-series data is used. Pressure and flow nodes are preferentially located at the junctions of main pipeline inlets / outlets and branches, with a spacing of 300 m (adjustable within the range of 200–500 m). Acoustic nodes are densely deployed in areas with high historical accident incidence, with a spacing of 150 m (adjustable within the range of 100–200 m). Settlement and strain nodes are densely deployed in road crossings, soft soil, or backfill sections. Humidity nodes are deployed in areas adjacent to drainage and chemical pipe corridors. All nodes are connected to the edge gateway via PoE or local power supply, and time synchronization uses PTP or GPS timing, with a maximum allowable clock drift of ≤100 ms. The N-side uses industrial Ethernet or a 5G private network, enabling TLS link encryption and device-level certificate authentication. The HDB adopts a hybrid architecture of "time-series database + relational database + graph database," importing GIS base maps and BIM models to generate topology maps. , where the node set Indicates the location of monitoring points and pipe fittings, edge set This indicates pipe segment connections and configures a data dictionary and unit dimension table. RFA is deployed in a campus data center or private cloud, providing real-time computing and visualization services.

[0017] Step 2: Sensor Calibration and Adaptive Sampling. After installation, perform 48 hours of static and operational calibration, recording the zero point, span, and noise floor. Then, enable the adaptive sampling strategy. ; in The baseline sampling rate is The adjustment coefficient is Sliding window variance operator In length Calculation within a time window, spatial location index Time Index The amount detected at the location is expressed as , , For in position ; Hz, Window length s, and set a hard bound for the sampling rate: , This strategy increases sampling density when fluctuations increase, and reduces frequency during steady-state periods to save bandwidth and energy, thus avoiding uneven data volume in subsequent fusion.

[0018] Step 3: Data transmission and arrival reordering, with N sides maintaining a timestamp buffer queue for each signal. Perform arrival reordering and deduplication on multi-source data in the same fusion batch to satisfy... Alignment constraints, ms (can be finely adjusted in the range of 200–500 ms) To balance real-time performance and data integrity, a limit is set for late data. s, late samples within this limit are still accepted and the batch is partially recalculated; exceeding this limit... The samples are entered into the database but are only used for offline statistics and do not participate in batch fusion. To enhance link robustness, ACK and retransmission mechanisms are enabled, with a maximum of 3 retransmissions per packet.

[0019] Step 4: Exponential sliding smoothing, all time series entering RFA are first processed by... Smooth the surface. For time index t The exponential moving average, For time index time Exponential moving average , It is the current sampling interval or an integer multiple thereof, and for Boundary protection (first sample) Initialization). This step reduces the impact of transient spikes on subsequent derivatives and energy-related features.

[0020] Step 5: Robust normalization and Z-domain transformation. Perform robust normalization on the smoothed sequence. , Robust normalization result for u at time index t and spatial location index x. Let be the exponential moving average of u at time index t and spatial location index x.

[0021] in The median, To determine the median absolute deviation, statistics are compiled in HDB using a rolling 14-day window, segmented by operating condition (season / shift / load), and refreshed daily at 0:00. From this point forward, all calculations will uniformly use the Z-domain. .

[0022] Step 6: Leakage index calculation; leakage risk is assessed using a dual-source fusion of pressure mutation and acoustic energy. ; Among them, the positive part operator Weight and Preferred (Parameters can be adjusted between 0.6–0.8 and 0.4–0.2). Short-term energy window Optimal value 32 (parameters can be adjusted within the range of 16–64). The sampling interval is used. The derivative is approximated by the backward difference. Short-time acoustic energy Where m is the window length; Total window length Z A for To improve acoustic discrimination, the acoustic channels are bandpass filtered to 1–5 kHz and normalized to unity power before energy calculation.

[0023] Step 7: Calculate the collapse sub-index, integrating collapse risk with structural response and foundation environment. in, Robust normalization result for ε at time index t and spatial location index x. Robust normalization result of Sa at time index t and spatial location index x. Let H be the exponential moving average of H at time index t and spatial location index x. Humidity normalization , The data is segmented by HDB according to season and soil type, with a statistical window of 14 days and updated every 6 hours. The derivative is approximated by difference, the same as in S6.

[0024] Step 8: Congestion sub-index and topology enhancement; congestion risk combined with pressure differential and flow attenuation. ; Among them, weight and ; The robust normalized result of F at time index t and spatial location index x, pressure difference. ,in For upstream pressure This is due to downstream pressure; Between 0.5 and 0.8, Parameters were tuned within the range of 0.2–0.5. To reflect path dependence and upstream chain effects, topology-weighted path characteristics were calculated. in, As Additional features, For location index The set of reachable paths For path Average pressure difference on the edge, path weight Normalized to pipe diameter (smaller diameter, higher weight), material (older materials have a 10-30% increased weight), and historical failure rate (linearly weighted by the frequency of events in the last two years). Topology enhancement with And positive part operators are added: , Step 9: Weighted fusion and quantile threshold determination to determine the overall risk index. initial The threshold uses an adaptive rolling quantile: ,in , This is the sample quantile operator. To avoid critical jitter, hysteresis is set: the trigger condition is... The reset condition is And set the hold time. Alerts are issued only afterward. All alert events are entered into the database, including... Values, sub-index decompositions, and context fragments.

[0025] Step 10: Online weight learning; once the event is verified on-site and a label is obtained. Then, press Online updates, including The weight at time t+1, Let be the weight at time t. Standard probability simplex The projection operator on the loss function is: Sigmoid function Learning rate .

[0026] Step 11: Leakage cross-correlation location, when the alarm type is leakage or occupy When the contribution exceeds 50% and holds true for three consecutive windows, the localization process is initiated. The acoustic signal is first passed through a 1–5 kHz bandpass filter and pre-whitened, and then the cross-correlation is calculated. Obtain the time delay difference , For time index Spatial location index Robust normalization results at point A For time index Spatial location index Robust normalization result at point A. Distance estimate from the leak point to the i-th sensor. ,in For sensor spacing, Equivalent sound velocity; preferably in an aqueous medium. =1480 m / s (calibrated within 1350–1550 m / s). Multiple sensor estimation results are intersected geometrically using weighted least squares, with weights set according to correlation peak value and signal-to-noise ratio; if If the boundary is exceeded, the boundary will be truncated at the nearest endpoint. The location results are linked with GIS to provide priority well locations for investigation and recommended valve operation guidelines.

[0027] Step 12: Sensor drift self-calibration, RFA maintains a steady-state baseline for each monitoring point. With scale During the 24-hour long window If the drift lasts for more than 2 hours, it is considered a slow drift, and the adjustment factor is... After being determined, a mild update is applied: the new normalization reference is based on the coefficients. Perform exponential sliding updates to ensure that no mutations are introduced and monitoring is not interrupted.

[0028] Step 13, Active sampling gain arrangement, when a certain pipe section's Reaching the 0.9 quantile of its historical distribution and When there is a significant increase in the last 10 minutes (linear fitting slope > 0.01), the sampling rates of the P and F channels associated with this segment are multiplied by the gain coefficient. For non-hotspot segments, bandwidth is reduced by the same amount across the entire network to maintain bandwidth conservation; the gain duration is 30 minutes, during which if R is continuously lower than... If the time limit is 10 minutes, the process is terminated early. The gain process, together with the adaptive sampling of S2, works to ensure that resources converge towards high-risk paths.

[0029] Step 14: Closed-loop verification and operation and maintenance linkage; all early warning events are entered into the work order system, according to... value, Based on geographical accessibility, a "priority response list" is generated. Post-response feedback (leakage / non-leakage, blockage cleared, discovery during pitting / excavation) is written back to the HDB; the RFA updates the tag y accordingly and triggers a weight update in S10. The system provides a post-event retrospective and interpretability view, displaying the time series of R, tri-exponential decomposition, topology enhancement contribution, and key original fragments involved in alarms, supporting professional review and model accountability.

[0030] Through the above 14 consecutive steps, the system can achieve a highly sensitive, low-false-alarm, and locatable full-process investigation of underground pipeline leakage, collapse, and blockage risks while ensuring privacy and computational efficiency.

[0031] As used herein, the term "preferred" is meant as an example, illustration, or illustration. Any aspect or design described herein as "preferred" need not be construed as being more advantageous than other aspects or designs. Rather, the use of the term "preferred" is intended to present the concept in a specific manner. As used in this application, the term "or" is intended to mean an inclusive "or" rather than an exclusionary "or." That is, unless otherwise specified or clear from the context, "X uses A or B" naturally includes either of the permutations. That is, if X uses A; X uses B; or X uses both A and B, then "X uses A or B" is satisfied in any of the foregoing examples.

[0032] Furthermore, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art based on a reading and understanding of this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the appended claims. In particular, with respect to the various functions performed by the aforementioned components (e.g., elements, etc.), the terminology used to describe such components is intended to correspond to any component (unless otherwise indicated) that performs the specified function of said component (e.g., is functionally equivalent to it), even if structurally not equivalent to the disclosed structure performing the functions in the exemplary implementations of this disclosure shown herein. Moreover, although specific features of this disclosure have been disclosed with respect to only one of several implementations, such features may be combined with one or more features of other implementations that may be desirable and advantageous for a given or particular application. Furthermore, with regard to the use of the terms “comprising,” “having,” “containing,” or variations thereof in the Detailed Description or claims, such terms are intended to be included in a manner similar to the term “including.”

[0033] The functional units in this invention embodiment can be integrated into a processing module, or each unit can exist physically separately, or multiple units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. The aforementioned devices or systems can execute the storage methods in the corresponding method embodiments.

[0034] In summary, the above embodiments are one implementation of the present invention, but the implementation of the present invention is not limited to the embodiments described above. Any changes, modifications, substitutions, combinations, or simplifications made that deviate from the spirit and principle of the present invention should be considered equivalent substitutions and are included within the protection scope of the present invention.

Claims

1. A multi-source data fusion system for underground pipeline network risk assessment, characterized in that, The system includes: The IoT sensor node unit S is deployed at multiple monitoring points in the underground pipeline network to collect real-time operational status data, including pressure values. Flow value Acoustic signal amplitude Settlement amount is Pipe wall strain and soil moisture ; The data transmission module N is used to send the data collected by S to the backend via a communication link; The historical database module HDB is used to store historical operating data of the pipeline network, accident records, maintenance logs, and geological and structural information. The Risk Fusion Analysis (RFA) module is used to fuse the data from S and HDB to calculate the comprehensive risk index. , ； Among them, the spatial location index is The time index is The leakage risk sub-index is: The collapse risk sub-index is The congestion risk sub-index is The non-negative fusion weight is ,satisfy and ;when Warnings are triggered at certain times, including risk thresholds. These are configurable parameters; The S uses an adaptive sampling rate Data collection is performed. ； Among them, the baseline sampling rate is The adjustment coefficient is Sliding window variance operator In length Calculation within a time window, spatial location index Time Index The amount detected at the location is expressed as , , For in position The detected value over the entire time period ; HDB also stores pipeline topology maps. , where the node set Indicates the location of monitoring points and pipe fittings, edge set This indicates pipe segment connection; the RFA calculates the topological weighted path characteristics of the corresponding spatial location x based on the pipe network topology map G=(V,E). , ； Among them, path set For location index Reachable paths, path weights The average differential pressure along the path is weighted by pipe diameter, material, and historical failure rate. In the path Take the average value on the edges; and As Additional features; The RFA is used to input time series. Application of exponential sliding smoothing ； in, For time index t The exponential moving average, For time index time Exponential moving average, smoothing coefficient , The smoothed detection quantity is the current sampling interval or an integer multiple thereof. ; The RFA is for the The calculation is as follows: (i) Leakage risk sub-index ； Among them, the positive part operator Weight and ; For time index Spatial location index Place of The robust normalization results For short-term acoustic energy; ； Where m is the window length; The total window length is and the sampling interval is . , For time index Spatial location index Robust normalization result for pressure value A; (ii) Collapse Risk Sub-index ； Among them, weight and ; For time index Spatial location index Robust normalization results for ε For time index Spatial location index Robust normalization results for Sa For time index Spatial location index The exponential moving average of H; Humidity normalization , value with standard deviation This data was obtained from historical statistics of the HDB. (iii) Congestion risk sub-index ； Among them, weight and ; For time index Spatial location index The robust normalization result for F, pressure difference ,in For upstream pressure This is due to downstream pressure; The RFA includes a leak localization module for using acoustic cross-correlation functions. Estimated delay difference , ； The sensor location is and The discrete time series is , The length of the discrete time series; For time index Spatial location index Robust normalization results for point A For time index Spatial location index The robust normalization result at point A; And based on Calculate the leak point to the 1st Sensor distance estimation Among them, the sensor spacing The speed of sound is obtained from the engineering layout. It is obtained from the calibration of the medium and temperature inside the pipe.

2. The system according to claim 1, characterized in that, The N includes an edge gateway configured with a timestamp buffer queue. With maximum tolerable jitter This is used to ensure that the reordering of multi-source data arriving at the backend satisfies [the required parameters]. Timing alignment constraints, , Indicates time; Improve the synchronicity of cross-source fusion.

3. The system according to claim 1, characterized in that, The RFA is a smoothed time series of any input before fusion. Implement robust normalization ； in, The median, The median absolute deviation is calculated from historical samples of the HDB. For time index Spatial location index Place The robust normalization results For time index Spatial location index Place The exponential moving average.

4. The system according to claim 1, characterized in that, The Iterate through online update rules with simplex projection: ； in, The weight at time t+1, Let be the weight at time t. Weight vector Learning rate ; loss function Among them, event tags The Sigmoid function is obtained from the verification of known alarms. ; Standard probability simplex Projection operator on.

5. The system according to claim 1, characterized in that, The risk threshold Adaptive setting of quantiles ； Among them, quantile level Given by the operation and maintenance strategy, For HDB history The sample set, The sample quantile operator is as follows: ； in, Indicator functions, For the infimum operator, This is for scanning candidate quantile values.

Images