Multi-source data fusion processing method for tunnel fire-fighting water system

Abstract

The invention belongs to the field of tunnel safety monitoring, and particularly relates to a multi-source data fusion processing method for a tunnel fire-fighting water system. The method comprises the following steps: acquiring observation data and control instruction data; driving the digital twin model to perform synchronous fluid dynamic simulation, and outputting simulation state data of topological positions corresponding to sensors in the physical pipe network; comparing the observation data acquired by each sensor with the corresponding simulation state data point by point, determining a real-time deviation vector, and matching the real-time deviation vector with a plurality of standard symptom vectors pre-stored in the fault knowledge graph; taking the matched symptom node as a starting point, performing reverse traversal along a directed edge of the fault knowledge graph, and positioning at least one candidate fault node; and respectively aggregating each directed edge weight on the reverse traversal path of each candidate fault node, determining a comprehensive confidence score of each candidate fault node, and determining a potential fault type according to the comprehensive confidence score. According to the invention, the sensitivity and accuracy of fault diagnosis are obviously improved.

Description

Technical Field

[0001] This invention belongs to the technical field of tunnel safety monitoring, and specifically relates to a method for multi-source data fusion processing of tunnel fire water systems. Background Technology

[0002] As a crucial component of modern transportation networks, the reliability and stability of the internal fire-fighting water system in tunnels are essential for ensuring tunnel safety and effectively responding to sudden fire accidents. Tunnel fire-fighting water systems typically have complex pipe networks, long pipelines, and numerous devices. They are also exposed to high humidity and enclosed environments for extended periods, making them prone to various malfunctions such as pipe leaks, blockages, valve jamming, pump performance degradation, and sensor drift.

[0003] Traditional monitoring methods for tunnel fire water systems often rely on periodic manual inspections or threshold alarms based on single physical quantities (such as pressure or flow rate). Manual inspections are not only inefficient and costly, but also struggle to detect hidden, progressive early-stage faults in real time. While simple threshold alarm mechanisms can provide real-time alerts, they often result in high false alarm and false negative rates because they do not consider the dynamic correlations and multivariate coupling effects between different parts of the system. Furthermore, they cannot accurately diagnose the fault type and specific location, making it difficult to effectively guide maintenance personnel in rapid repairs and impacting the efficiency and accuracy of emergency response.

[0004] Currently, related technologies attempt to utilize historical data for fault diagnosis. However, these methods often require a large number of labeled fault samples for model training, while fault data in real-world scenarios is sparse and difficult to obtain. Furthermore, purely data-driven models lack a deep understanding of the system's physical mechanisms, resulting in poor interpretability of their diagnostic results. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, namely inaccurate fault diagnosis, poor real-time performance, and insufficient interpretability, this invention proposes a multi-source data fusion processing method for tunnel fire-fighting water systems. The method includes: The system acquires real-time observation data from sensors deployed in the physical network of the tunnel fire water system, and simultaneously acquires control command data of the tunnel fire water system. The control command data is input into a pre-built digital twin model, which drives the digital twin model to perform synchronous fluid dynamics simulation and outputs the simulation status data of the topological positions of each sensor in the physical pipeline network. The digital twin model is constructed based on the topology, pipeline parameters and fluid dynamics equations of the physical pipeline network. The observation data collected by each sensor is compared point by point with the corresponding simulation state data to determine the real-time deviation vector including the deviation value of each point. The real-time deviation vector is then matched with multiple standard symptom vectors pre-stored in the fault knowledge graph. The standard symptom vectors are used to associate symptom nodes in the fault knowledge graph. Starting from the matched symptom node, the system traverses backward along the directed edges of the fault knowledge graph to locate at least one candidate fault node. The directed edges are associated with weight values, which are pre-assigned based on the simulation results of the digital twin model on the corresponding fault by quantifying the degree of impact of the fault on the symptom. The weights of directed edges on the reverse traversal path of each candidate fault node are aggregated to determine the comprehensive confidence score of each candidate fault node, and the potential fault type is determined based on the comprehensive confidence score.

[0006] In some preferred embodiments, the observation data includes a pressure data sequence, a flow rate data sequence, and a water tank level data sequence; the control command data includes a valve opening data sequence and a pump status data sequence; and after acquiring the observation data and control command data, the method further includes: Based on the precision clock signals embedded in each data source, and using a preset control cycle as the reference time window, synchronous interpolation processing is performed on the pressure data sequence, flow data sequence, water tank level data sequence, valve opening data sequence, and water pump status data sequence to generate a status control dataset and a status observation dataset with a unified time sequence. Outlier removal and smoothing noise reduction processing are then performed on the datasets.

[0007] In some preferred embodiments, prior to determining the real-time deviation vector, the method further includes: The state observation dataset is subjected to a standardization transformation, which employs either the min-max normalization method or the Z-score normalization method. When the min-max normalization method is used, the global maximum and global minimum values ​​of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is then subtracted from the global minimum value and divided by the difference between the global maximum and global minimum values, so that the range of each physical quantity data sequence is mapped to the [0,1] interval. When using the Z-score standardization method, the arithmetic mean and standard deviation of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is subtracted from the arithmetic mean and then divided by the standard deviation, so that each physical quantity data sequence is converted into standard normal distribution data.

[0008] In some preferred embodiments, each of the standard symptom vectors is uniquely associated with a symptom node, and the matching of the real-time deviation vector with the standard symptom vector includes: Determine the similarity metric between the real-time deviation vector and each standard symptom vector in the spectrum; All calculated similarity metrics are compared with a preset matching threshold. The symptom nodes associated with standard symptom vectors whose similarity metrics are greater than the matching threshold, or the symptom nodes associated with the top N standard symptom vectors in terms of similarity metrics, are determined as the set of symptom nodes that are successfully matched, where N is an integer greater than 0.

[0009] In some preferred embodiments, the process of constructing the fault knowledge graph includes: Based on the equipment list and historical maintenance records of the tunnel fire water system, a set of fault nodes is defined, which includes fault types such as pipe leakage, pipe blockage, valve jamming, water pump performance degradation, and sensor drift. Based on the monitoring point layout of the digital twin model, a set of symptom nodes is defined, and each node in the set corresponds to a pressure, flow rate or liquid level deviation pattern. Based on the experience of experts in the pre-defined domain or historical failure cases, a directed edge with causal orientation is constructed between the failure node and the symptom node. By injecting the specific fault mode and severity represented by the fault node into the digital twin model, the response deviation data of all monitoring points in the network are obtained after running the simulation. The steady-state change amplitude, dynamic response speed or probability distribution change of the monitoring point deviation of each symptom node under the fault mode are analyzed. A normalized value is calculated using a preset sensitivity analysis formula, and the value is used as the weight value attached to the directed edge connecting the fault node and the corresponding symptom node.

[0010] In some preferred embodiments, the step of aggregating the weights of directed edges on the reverse traversal paths of each candidate fault node to determine the comprehensive confidence score of each candidate fault node includes: For a candidate fault node that has been located, a reverse traversal path is determined in the fault knowledge graph, starting from all the successfully matched symptom nodes and finally reaching the candidate fault node via directed edges. The weights of all directed edges on each reverse traversal path are multiplied sequentially to obtain the path support of that path for the faulty node. If a candidate faulty node has multiple reverse traversal paths, the maximum value of the support of all reverse traversal paths, or the arithmetic sum of the support of all paths, is used as the comprehensive confidence score of the candidate faulty node.

[0011] In some preferred embodiments, determining the potential fault type based on the comprehensive confidence score includes: Identify the candidate fault node with the highest overall confidence score, and directly determine the fault type it represents as the potential fault type; and / or, By combining the historical prior probability of the fault occurrence, the comprehensive confidence score is corrected by Bayesian posterior probability, and the fault type with the highest posterior probability is selected as the determination result.

[0012] In some preferred embodiments, after determining the potential fault type, the method further includes: When the potential fault type is pipeline leakage or blockage, determine the target dimension in the real-time deviation vector whose numerical deviation is higher than a preset threshold. Based on the actual geographical location and topological connection relationship of the sensor corresponding to the target dimension in the physical pipeline network, the physical pipeline segment range where a fault may occur is determined.

[0013] In some preferred embodiments, the method further includes the output of diagnostic results and the triggering of response actions: The potential fault types and their corresponding comprehensive confidence scores are encapsulated into a structured diagnostic report, output to a preset human-machine interface for visualization, and simultaneously stored in a historical database; Based on the potential fault types and the preset fault-response rule base, preliminary control suggestion instructions are generated and issued to suppress the impact of the fault.

[0014] The beneficial effects of this invention are: Compared to existing technologies, in one aspect, the embodiments of the present invention construct a digital twin model to compare real-time observation data with simulation data based on physical mechanisms, thereby capturing subtle deviations that are difficult to detect by traditional threshold alarm methods, significantly improving the sensitivity and accuracy of fault diagnosis, and enabling effective early warning of potential faults.

[0015] On the other hand, embodiments of the present invention introduce a fault knowledge graph to express the causal relationship between equipment faults and monitored symptoms in a structured form. Through reverse traversal and weight aggregation in the fault knowledge graph, the reasoning path for fault types can be determined, making the diagnostic process more transparent and improving reliability. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating a multi-source data fusion processing method for a tunnel fire-fighting water system proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the framework of a multi-source data fusion processing system for a tunnel fire-fighting water system proposed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer system proposed in an embodiment of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] The method of this invention can be applied to the intelligent operation and maintenance of fire water systems in large infrastructure such as tunnels and underground utility tunnels.

[0020] For example, a 5-kilometer-long twin-tube highway tunnel has a fire water system that includes a water storage tank, two main fire pumps (one for use and one for standby), a main pipeline laid along the tunnel, several zone valves, and fire hydrants installed every 50 meters. Ten pressure sensors and two flow sensors are installed at the main pump outlet and key nodes of the main pipeline (such as branches and diameter changes), and one level sensor is installed in the water storage tank.

[0021] Please refer to Figure 1 The first embodiment of the present invention provides a method for multi-source data fusion processing of a tunnel fire water system, comprising: Step S10: Obtain observation data collected in real time by sensors deployed in the physical pipeline network of the tunnel fire water system, and at the same time obtain control command data of the tunnel fire water system. Step S20: Input the control command data into the pre-built digital twin model, drive the digital twin model to perform synchronous fluid dynamics simulation, and output the simulation status data of the topological positions of each sensor in the physical pipeline network. The digital twin model is constructed based on the topology, pipeline parameters and fluid dynamics equations of the physical pipeline network. Step S30: The observation data collected by each sensor is compared point by point with the corresponding simulation state data to determine the real-time deviation vector including the deviation value of each point, and the real-time deviation vector is matched with multiple standard symptom vectors pre-stored in the fault knowledge graph. The standard symptom vectors are used to associate symptom nodes in the fault knowledge graph. Step S40: Starting from the matched symptom node, traverse backward along the directed edges of the fault knowledge graph to locate at least one candidate fault node. The directed edges are associated with weight values, which are pre-assigned based on the simulation results of the digital twin model on the corresponding fault by quantifying the degree of influence of the fault on the symptom. Step S50: Aggregate the weights of the directed edges on the reverse traversal path of each candidate fault node, determine the comprehensive confidence score of each candidate fault node, and determine the potential fault type based on the comprehensive confidence score.

[0022] In a typical tunnel fire water system, the physical pipeline network extends for several kilometers along the tunnel, and various monitoring and control devices are densely distributed on it. Step S10 of this embodiment aims to simultaneously collect two types of data that are different in nature but closely related.

[0023] The first type of data reflects the actual operating status of the system. This data is collected in real time by various types of sensors directly installed in the physical pipeline network. For example, pressure sensors installed at the pump outlet, main pipe, and the starting point of each zone continuously report pressure readings; flow meters installed in key pipe sections measure the instantaneous flow rate; and level sensors in the fire-fighting water storage tank monitor the water level. These data constitute a direct measurement of the physical reality of the system.

[0024] Secondly, there is control command data reflecting the system's operational intent. This data typically originates from the tunnel's central monitoring system (such as a SCADA system) and includes operational instructions issued to field actuators. For example, instructions to start / stop or adjust the speed of the main fire pump, as well as instructions to open, close, or adjust the opening degree of various electric valves in the pipeline network. By simultaneously acquiring these two types of data, this method can comprehensively grasp complete information on "what state the system should be in" (determined by control commands) and "what state it is actually in" (reflected by observation data).

[0025] Based on this, this embodiment uses a pre-built digital twin model for synchronous simulation to generate an idealized reference benchmark, which serves as a high-fidelity virtual copy of the tunnel fire-fighting hydrophysical system.

[0026] Specifically, the construction of the digital twin model relies on the system's engineering design drawings and equipment parameters to accurately reproduce the topology of the physical pipeline network, namely the connection relationships and spatial locations of all components such as pipes, valves, pumps, and fire hydrants. Simultaneously, the model includes detailed pipeline parameters, such as the length, diameter, and wall roughness corresponding to the material of each pipe segment. The core driving force of the model is the recognized fundamental equations of fluid mechanics, such as the energy conservation equation and the continuity equation, which constitute the physical basis of hydraulic calculations.

[0027] During the execution of this embodiment, the control command data acquired in step S10 is input into the digital twin model in real time. The model then performs rapid fluid dynamics simulation calculations based on these commands (e.g., simulating the start-up of a water pump or the opening of a valve) and its internal physical laws. The simulation output is a set of simulation state data that perfectly corresponds to the physical sensor layout; that is, under the current control conditions, without any faults or unknown disturbances, the theoretical values ​​of physical quantities such as pressure and flow rate at each monitoring point.

[0028] Specifically, the actual observation data collected by the sensors is compared point-by-point with the simulated state data output by the digital twin model at the same time, according to the monitoring points. For example, the pressure deviation at a point is obtained by subtracting its simulated reading in the model from the actual reading of pressure sensor No. 1.

[0029] Repeating this operation for all monitoring points constructs a multi-dimensional real-time deviation vector, where each dimension represents the degree of deviation at a monitoring point. This embodiment matches the real-time deviation vector with multiple pre-stored standard symptom vectors in a fault knowledge graph.

[0030] Specifically, the fault knowledge graph is a structured database containing expert knowledge and historical data. Each standard symptom vector is a digital template of a typical anomaly pattern (for example, a vector might represent a situation where upstream pressure is normal but downstream pressure is generally low), and each standard symptom vector is uniquely associated with a symptom node in the graph. The internal structure of the fault knowledge graph is a network composed of fault nodes representing the root cause (such as pipe leaks or pump failures) and symptom nodes representing external manifestations, connected by directed edges.

[0031] By calculating the similarity between the real-time deviation vector and each standard symptom vector, one or more of the most prominent symptoms currently being observed can be identified.

[0032] Based on the matched symptoms, reverse causal tracing is performed in the fault knowledge graph to locate possible fault sources. The directed edges in the fault knowledge graph point from cause to effect, that is, from fault nodes to symptom nodes that they may trigger. Furthermore, each directed edge is accompanied by a weight value, which is a quantitative result obtained from numerous fault simulations using a digital twin model. For example, by simulating a pipeline leak at a specific location in the model, the impact on readings at all monitoring points across the network is analyzed, thereby accurately calculating the correlation strength (i.e., weight) between the fault and each related symptom.

[0033] Based on this, starting from the successfully matched symptom node, the system traverses along the opposite direction of the directed edges to locate all candidate fault nodes that may cause these symptoms.

[0034] Furthermore, this embodiment performs confidence assessments on all candidate faulty nodes to determine the most likely potential fault type. For each located candidate faulty node, all paths tracing back to it from the discovered symptom nodes are determined.

[0035] Specifically, this embodiment uses a preset aggregation algorithm to comprehensively consider the weight values ​​of all directed edges on these paths. For example, the weight of a path can reflect the overall credibility of the causal chain it represents, and if a candidate fault node can be pointed to by multiple high-weight paths, it is more likely to be the true cause of the fault. In this way, a comprehensive confidence score is calculated for each candidate fault node.

[0036] Finally, based on the overall confidence scores, the candidate fault nodes with the highest scores are ranked or compared, and the fault type represented by the candidate fault node with the highest score is determined as the final conclusion of this diagnosis, i.e., the potential fault type, thus completing the entire process from data collection to intelligent diagnosis.

[0037] In a preferred embodiment, the observation data includes a pressure data sequence, a flow rate data sequence, and a water tank level data sequence; the control command data includes a valve opening data sequence and a pump status data sequence; and after acquiring the observation data and control command data, the method further includes: Based on the precision clock signals embedded in each data source, and using a preset control cycle as the reference time window, synchronous interpolation processing is performed on the pressure data sequence, flow data sequence, water tank level data sequence, valve opening data sequence, and water pump status data sequence to generate a status control dataset and a status observation dataset with a unified time sequence. Outlier removal and smoothing noise reduction processing are then performed on the datasets.

[0038] This embodiment is a further detailed explanation of the data acquisition and preprocessing steps in the foregoing embodiments. In the actual operating environment of tunnel fire water systems, data sources are diverse and data quality varies. Therefore, in order to ensure the accuracy of subsequent model simulation and deviation analysis, preprocessing is necessary.

[0039] In this embodiment, the observation data specifically includes the following types of time series data: pressure data series, derived from pressure sensors installed at key nodes in the pipeline network (such as pump outlets, main pipeline branches, elevation points, etc.); flow data series, derived from flow meters installed at the main pump house outlet or on the main pipeline; and water tank level data series, derived from level gauges in fire-fighting water storage tanks. These data collectively constitute a real-time description of the system's physical state.

[0040] Meanwhile, the control command data specifically includes: valve opening data sequence, which records the percentage of opening commands issued by the monitoring center to each remote control valve (such as zone valve and pressure relief valve) in the pipeline network; and pump status data sequence, which records the start and stop commands for the main fire pump and standby pump or the operating frequency commands for the variable frequency pump.

[0041] After acquiring the aforementioned raw observation data and control command data, time synchronization and alignment processes are performed on them: Since different sensors and controllers may use different communication protocols and sampling periods, the timestamps of their uploaded data may deviate to some extent. Therefore, in order to establish a unified time reference, this embodiment sets a preset control period as a reference time window (for example, it can be set to 5 seconds or 10 seconds), which should match the main time scale of system control and response.

[0042] Then, synchronous interpolation is performed on all acquired data sequences, including pressure, flow rate, liquid level, valve opening, and pump status. For example, the start time of each time window is used as a unified time point. If a data sequence does not have an original data point at that time, the value at that time can be estimated using the two nearest data points before and after it through linear interpolation.

[0043] Accordingly, all discrete and asynchronous data sequences are transformed into regular datasets aligned on a unified time axis: a state control dataset containing synchronized valve opening data and pump status data; and a state observation dataset containing synchronized pressure, flow, and level data.

[0044] Furthermore, outlier removal and smoothing / denoising processes are performed.

[0045] During data acquisition and transmission, significant outliers may occur due to electromagnetic interference, communication packet loss, or momentary sensor malfunctions. These outliers can severely interfere with subsequent analysis. The outlier removal process in this embodiment can employ the statistical 3-sigma principle, which calculates the mean (μ) and standard deviation (σ) of each data sequence within a historical window. Any data point outside the range [μ-3σ, μ+3σ] is considered an outlier and is either removed or replaced (e.g., replaced with the mean of its nearest neighbor).

[0046] Even after removing obvious outliers, the data sequence may still contain high-frequency fluctuations caused by sensor noise or minute fluid pulsations. To extract the main trend reflecting changes in the system's macroscopic state, this implementation further smooths and denoises the dataset. Common methods include applying a moving average filter, which replaces the true value of a given point with the average of N neighboring points, effectively filtering out random noise. Another more advanced method is to use a Kalman filter, which can make an optimal estimate of the system's true state based on the system's dynamic model and noisy measurement data, making it particularly suitable for processing data from dynamic systems.

[0047] After a series of refined processing steps, including time synchronization, outlier removal, and smoothing and denoising, the original multi-source heterogeneous data is transformed into high-quality state control datasets and state observation datasets with unified time sequence. This provides a more reliable data foundation for subsequent accurate simulation of the digital twin model and reliable deviation comparison analysis.

[0048] As a preferred embodiment, before determining the real-time deviation vector, the method further includes: The state observation dataset is subjected to a standardization transformation, which employs either the min-max normalization method or the Z-score normalization method. When the min-max normalization method is used, the global maximum and global minimum values ​​of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is then subtracted from the global minimum value and divided by the difference between the global maximum and global minimum values, so that the range of each physical quantity data sequence is mapped to the [0,1] interval. When using the Z-score standardization method, the arithmetic mean and standard deviation of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is subtracted from the arithmetic mean and then divided by the standard deviation, so that each physical quantity data sequence is converted into standard normal distribution data.

[0049] This implementation addresses the problems of inconsistent dimensions and significant differences in orders of magnitude in multidimensional observation data. For example, in a state observation dataset, pressure data is typically measured in megapascals (MPa), with values ​​ranging from 0.1 to 1.6; flow rate data is measured in cubic meters per second (m³ / s), with values ​​ranging from 0.05 to 0.5; and pool levels are measured in meters (m), with values ​​ranging from 1 to 5. If these raw-scale values ​​are used directly for subsequent bias calculations and pattern matching, variables with larger value ranges (such as level or pressure after unit conversion) will dominate the calculations, thus masking the true system changes reflected by variables with smaller value ranges (such as flow rate), leading to biased diagnostic results.

[0050] To eliminate the adverse effects of differences in units and scale, this method performs a standardization transformation on the state observation dataset before using it for bias analysis, converting it into dimensionless, comparable data. This implementation provides two preferred standardization methods: Min-Max Normalization: The raw data for each dimension is linearly mapped to the interval [0, 1]. For any physical quantity in the state observation dataset (such as the pressure at point 1 of the pipeline network), its historical minimum value Xmin and historical maximum value Xmax under normal operating conditions are determined (these can be obtained based on the equipment's design parameters, safety thresholds, or statistical analysis of long-term stable operating data of the system). For any observed value Xt at time t, its standardized value Xnorm can be calculated using the following formula: ; By performing this operation on every data sequence in the observation dataset (all pressure, flow, and level sequences), a standardized state observation dataset is obtained in which all values ​​are distributed between [0, 1].

[0051] Z-score standardization: This method transforms the raw data for each dimension into a standard normal distribution with a mean of 0 and a standard deviation of 1. For each physical quantity in the dataset, historical data under a large number of normal operating conditions are collected, and its mean μ and standard deviation σ are calculated.

[0052] For any observation value Xt at time t, its Z-score standardized value Xz can be calculated using the following formula: ; After Z-score standardization, the values ​​in the standardized state observation dataset represent the degree to which the current observation deviates from its historical mean, and the unit is standard deviation.

[0053] In a preferred embodiment, each of the standard symptom vectors is uniquely associated with a symptom node, and the matching of the real-time deviation vector with the standard symptom vector includes: Determine the similarity metric between the real-time deviation vector and each standard symptom vector in the spectrum; All calculated similarity metrics are compared with a preset matching threshold. The symptom nodes associated with standard symptom vectors whose similarity metrics are greater than the matching threshold, or the symptom nodes associated with the top N standard symptom vectors in terms of similarity metrics, are determined as the set of symptom nodes that are successfully matched, where N is an integer greater than 0.

[0054] In this embodiment, a real-time deviation vector is constructed based on standardized observation data and simulation data at the same timestamp t. The dimension n of this vector is equal to the total number of monitoring points (including all pressure, flow, liquid level, etc. monitoring points), and each element d in this vector... i All are standardized observations o from the corresponding monitoring point i i The result obtained by subtracting the standardized simulation value at that point: ; The vector This is a digital fingerprint of the system's current health status. Under ideal, fault-free conditions, this vector should be close to an all-zero vector. When an anomaly occurs in the system, certain elements in this vector will exhibit specific positive and negative values ​​and magnitudes, and its overall pattern characterizes the external signs of the anomaly.

[0055] Multiple standard symptom vectors are stored in the pre-built fault knowledge graph. Each of the standard symptom vectors These are all idealized and standardized vector representations of typical anomalies (such as generally low pressure at the end of a certain zone or abnormally high flow rate at the main pump outlet) in an n-dimensional monitoring space.

[0056] In order to identify the current real-time deviation vector To determine which known symptom corresponds to which symptom, this implementation method traverses all standard symptom vectors in the knowledge graph and calculates them one by one. With each The similarity.

[0057] This embodiment provides two preferred similarity calculation methods: Cosine Similarity: The similarity measurement is the similarity of two vectors in direction, with a value range of [-1, 1]. The closer the value is to 1, the more consistent the directions of the two vectors are, meaning the more similar the deviation patterns they represent. The calculation formula is as follows: ; calculate With all The cosine similarity is used to select the standard symptom vector with the largest similarity value (which usually needs to exceed a preset threshold, such as 0.9) as the matching result.

[0058] Euclidean distance: Euclidean distance measures the geometric distance between two vectors in n-dimensional space; the smaller the distance, the closer the two vectors are. To convert this into a similarity concept (the larger the value, the more similar), we can take its reciprocal or use a Gaussian kernel-based similarity function. When directly comparing using Euclidean distance, the goal is to find the match with the smallest distance. The calculation formula is as follows: ; calculate With all The Euclidean distance is used to select the N standard symptom vectors with the smallest distance values ​​as the matching results.

[0059] Through the above process, the most critical symptoms exhibited by the current system are identified from the raw deviation data. For example, the matching result might be symptom S3: upstream pressure is normal, while downstream zone pressure collectively decreases. The identified symptom nodes will serve as the starting point for the next step of causal tracing in the fault knowledge graph.

[0060] Specifically, the construction process of the fault knowledge graph includes: Based on the equipment list and historical maintenance records of the tunnel fire water system, a set of fault nodes is defined, including fault types such as pipe leakage, pipe blockage, valve jamming, pump performance degradation, and sensor drift. According to the monitoring point layout of the digital twin model, a set of symptom nodes is defined, where each node corresponds to a pressure, flow rate, or level deviation mode. Based on preset domain expert experience or historical fault cases, causal directed edges are constructed between fault nodes and symptom nodes. By injecting the fault mode and severity represented by the fault node into the digital twin model, and after running the simulation, response deviation data of all monitoring points in the network are obtained. The steady-state change amplitude, dynamic response speed, or probability distribution change of the monitoring point deviation corresponding to each symptom node under the fault mode is determined. A normalized value is calculated using a preset sensitivity analysis formula, and this value is used as the weight value attached to the directed edge connecting the fault node and the corresponding symptom node.

[0061] More specifically, the step of aggregating the weights of directed edges on the reverse traversal paths of each candidate fault node to determine the comprehensive confidence score of each candidate fault node includes: For a candidate fault node that has been located, a reverse traversal path is determined in the fault knowledge graph that starts from all successfully matched symptom nodes, passes through directed edges, and finally reaches the candidate fault node. The weight values ​​attached to all directed edges on each reverse traversal path are multiplied sequentially to obtain the path support of the path to the fault node. If a candidate fault node has multiple reverse traversal paths, the maximum value of the support of all reverse traversal paths, or the arithmetic sum of the support of all paths, is used as the comprehensive confidence score of the candidate fault node.

[0062] The fault knowledge graph consists of a directed graph, nodes, and weighted edges.

[0063] In this context, symptom nodes represent macroscopic anomalies in the system that can be observed by the sensor network. Each symptom node corresponds one-to-one with a standard symptom vector mentioned in Example 4. For example, low pressure in the main pipeline while the water level in the tank drops at a normal rate, a decrease in flow rate instead of an increase after the water pump is started, and generally lower-than-normal readings of all end pressure sensors in a fire compartment can all serve as specific symptom nodes.

[0064] Fault node: This represents the root physical cause of the abnormality, i.e., the final diagnosed result. For example, severe wear of the main pump impeller, serious leakage in a certain part of the pipeline network, jamming of the zone valve V-101 that cannot be fully opened, and signal drift of the pressure sensor P-05 can all be considered specific fault nodes.

[0065] In a fault knowledge graph, a directed edge points from a fault node to a symptom node, indicating that the fault is one of the causes of the symptom. For example, a directed edge might point from a fault node (a serious leak in the pipeline) to a symptom node (low pressure in the main pipe and an abnormally rapid rate of water level drop in the pool). Each directed edge is assigned a weight value within the range [0,1] to describe the strength or determinism of the causal relationship. The weight assignment process includes: By using a digital twin model, various single or compound faults are actively injected, simulations are run, and the resulting system state deviations are observed. Through a large number of simulation experiments, the strong correlation between specific faults and specific symptom patterns is statistically determined, which serves as the data basis for weight assignment.

[0066] Specifically, the preset sensitivity analysis formula is as follows: ; in, It's a malfunction. with signs Edge weights between them yes Caused by, and The average steady-state deviation of the relevant monitoring points It is the maximum deviation of the monitoring point under any possible failure condition. It is the dynamic response time. This is the reference response time (preset, such as 0.2s). and These are weighting coefficients (which sum to 1) used to adjust the importance of the amplitude and velocity features.

[0067] Calculated using this formula It is a value between [0, 1].

[0068] It's easy to understand that in complex fault scenarios, a root cause (such as a severe leak in the main pipeline) often triggers multiple observable symptoms simultaneously (such as low pressure in the main pipeline and an abnormal rate of drop in the water level in the tank). In this case, the diagnostic system will successfully match multiple activated symptom nodes.

[0069] This implementation method integrates and quantifies evidence from multiple symptoms through the following steps: After the matching process described in the above embodiments, the set of activated symptom nodes is determined, for example, {S1, S2}. Starting from each symptom node (S1 and S2) in the set, a reverse tracing is performed in the fault knowledge graph to find all candidate fault nodes pointing to them, and finally a list of candidate fault nodes is obtained, as well as all reverse traversal paths corresponding to each candidate fault node.

[0070] For example: There are two reverse paths for candidate fault F1: one from S1 to F1, and the other from S2 to F1. Candidate fault F2 has only one reverse path: from S1 to F2; Candidate fault F3 has only one reverse path: from S2 to F3.

[0071] For each determined reverse traversal path, calculate its path support, which is obtained by sequentially multiplying the weights of all directed edges along the path.

[0072] It's easy to understand that a causal relationship is a direct directed edge from the fault node to the symptom node, so a reverse traversal path contains only one edge. In this case, the result of the "multiplication operation" is the weight of that edge itself. Therefore, the path support is equal to the weight of the causal relationship edge.

[0073] The support of the path S1→F1 is: P(F1)=W(F1,S1); The support of the path S2→F1 is: P(F1)=W(F1,S2); The support of the path S1→F2 is: P(F2)=W(F2,S1)

[0074] If a candidate fault node has multiple reverse traversal paths pointing to it (i.e., it can explain multiple occurrences of symptoms), then the support of these paths needs to be aggregated to form the final comprehensive confidence score of the fault node. In this implementation, the maximum value among all path support is taken as the final comprehensive confidence score, that is, the strongest "causal evidence chain" of the fault is taken as the representative of its confidence.

[0075] Furthermore, determining the potential fault type based on the comprehensive confidence score includes: The candidate fault node with the highest overall confidence score is identified, and the fault type it represents is directly identified as the potential fault type; and / or, the overall confidence score is corrected using Bayesian posterior probability by combining the historical prior probability of the fault occurrence, and the fault type with the highest posterior probability is selected as the determination result.

[0076] In this embodiment, the candidate fault node with the highest comprehensive confidence score is selected, and / or a Bayesian posterior probability correction method combining prior probability is adopted: Taking the Bayesian posterior probability correction method as an example, a Bayesian prior probability is assigned to each fault node F_k in the fault knowledge graph, forming an example prior probability library. For instance, this probability value can be obtained by analyzing maintenance records from the past few years and statistically analyzing the frequency of occurrence of various faults.

[0077] Bayesian posterior probability correction is performed based on the example prior probability library. The comprehensive confidence score of each candidate fault node is multiplied by its prior probability to obtain an unnormalized posterior probability score. The fault type with the highest posterior probability score is selected as the final determination result.

[0078] Furthermore, after determining the potential fault type, the method further includes: When the potential fault type is pipeline leakage or blockage, the target dimension in the real-time deviation vector with a numerical deviation higher than a preset threshold is determined; based on the actual geographical location and topological connection relationship of the sensor corresponding to the target dimension in the physical pipeline network, the physical pipeline segment range where the fault may occur is determined.

[0079] In this embodiment, each element in the real-time deviation vector is traversed. If the absolute value of the element (representing the deviation value of the i-th sensor) is higher than the preset alarm threshold, then the dimension is marked as the target dimension.

[0080] Based on this, by utilizing the physical geographical location information and topological connection relationship of the pipeline network embedded in the digital twin model (i.e., the upstream and downstream connection relationship between each sensor, valve, and pipe), the physical pipe segment where the fault is located can be identified. For example, when the potential fault type is a pipe leak, since the leak point will cause a general drop in pressure in all pipe segments downstream, the fault can be located along the direction of water flow in the pipeline network topology.

[0081] The goal of the location is to find a point of sharp pressure drop. Specifically, it is to identify the topologically adjacent sensors (sensor A is upstream of sensor B): if the pressure deviation of sensor A is within the normal range, while the pressure deviation of sensor B is significantly negative (i.e., it is a critical deviation sensor), then the physical pipe segment where the leak occurs is most likely the segment connecting sensor A and sensor B.

[0082] More specifically, the method further includes: The potential fault types and their corresponding comprehensive confidence scores are encapsulated into a structured diagnostic report, output to a preset human-machine interface for visualization, and simultaneously stored in a historical database; Based on the potential fault types and the preset fault-response rule base, preliminary control suggestion instructions are generated and issued to suppress the impact of the fault.

[0083] In this embodiment, all relevant diagnostic information is aggregated into a structured data object (e.g., JSON or XML format) to form a diagnostic report.

[0084] For example, the diagnostic report includes at least: Event ID: A unique identifier for this diagnostic event; Diagnosis Time: The precise timestamp of the generated report; Potential Fault Type: Such as a serious leak in the main pipeline; Overall Confidence Score: Such as 1.85 or a posterior probability score of 0.925; Fault Location Information: Such as the pipe segment located between sensors P2 and P3; Chain of Evidence: A list of activated symptom nodes: Such as [S1: Low pressure in the main pipeline, S2: Abnormal rate of drop in water level in the pool]; List of key deviation sensors: Such as [P3, P4].

[0085] While the diagnostic report is being pushed to the human-machine interface, it will be written to a historical database (such as a relational database MySQL or a time-series database InfluxDB).

[0086] This implementation method automatically generates handling suggestions based on a preset fault-response rule base. For example, this fault-response rule base stores the mapping relationship between different fault types and recommended response measures in the form of key-value pairs. This fault-response rule base can be formulated based on expert experience, and this implementation method does not limit its specific form. The diagnosed potential fault type is used as the query key to match the corresponding response measures from the rule base.

[0087] Please refer to Figure 2 The second embodiment of the present invention proposes a multi-source data fusion processing system for a tunnel fire-fighting water system, comprising: The data acquisition module 210 is used to acquire observation data collected in real time by sensors deployed in the physical pipe network of the tunnel fire water system, and at the same time acquire the control command data of the tunnel fire water system. The simulation module 220 is used to input the control command data into a pre-built digital twin model, drive the digital twin model to perform synchronous fluid dynamics simulation, and output the simulation status data of the topological positions of each sensor in the physical pipeline network. The digital twin model is constructed based on the topology, pipeline parameters and fluid dynamics equations of the physical pipeline network. The vector matching module 230 is used to compare the observation data collected by each sensor with the corresponding simulation state data point by point to determine the real-time deviation vector including the deviation value of each point, and to match the real-time deviation vector with multiple standard symptom vectors pre-stored in the fault knowledge graph. The standard symptom vectors are used to associate symptom nodes in the fault knowledge graph. The node matching module 240 is used to traverse the directed edges of the fault knowledge graph in reverse, starting from the matched symptom node, to locate at least one candidate fault node. The directed edges are associated with weight values, which are pre-assigned based on the simulation results of the digital twin model on the corresponding fault by quantifying the degree of influence of the fault on the symptom. The fault output module 250 is used to aggregate the weights of the directed edges on the reverse traversal path of each candidate fault node, determine the comprehensive confidence score of each candidate fault node, and determine the potential fault type based on the comprehensive confidence score.

[0088] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process and related descriptions of the system described above can be found in the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0089] The following is for reference. Figure 3 It shows a schematic diagram of the structure of a computer system suitable for implementing the system and method embodiments of the present invention. Figure 3 The server shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0090] like Figure 3 As shown, the computer system includes a Central Processing Unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in Read Only Memory (ROM) 302 or programs loaded from storage section 308 into Random Access Memory (RAM) 303. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An Input / Output (I / O) interface 305 is also connected to the bus 304.

[0091] The following components are connected to the input / output interface 305: an input section 306 including a keyboard, mouse, etc.; an output section 307 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 308 including a hard disk, etc.; and a communication section 309 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 309 performs communication processing via a network such as the Internet. A drive 310 is also connected to the input / output interface 305 as needed. A removable medium 311, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 310 as needed so that computer programs read from it can be installed into the storage section 308 as needed.

[0092] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit 301, it performs the functions defined in the methods of the present invention. It should be noted that the computer-readable medium described above in the present invention can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0093] More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, optical fiber, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can transmit, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0094] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including local area networks (LANs) or wide area networks (WANs), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0095] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0096] The terms “first”, “second”, etc., are used to distinguish similar objects, not to describe or indicate a specific order or sequence.

[0097] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus / device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent in such process, method, article, or apparatus / device.

[0098] The technical solution of the present invention has now been described in conjunction with the preferred embodiments shown in the accompanying drawings.

[0099] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for multi-source data fusion processing of a tunnel fire-fighting water system, characterized in that, include: The system acquires real-time observation data from sensors deployed in the physical network of the tunnel fire water system, and simultaneously acquires control command data of the tunnel fire water system. The control command data is input into a pre-built digital twin model, which drives the digital twin model to perform synchronous fluid dynamics simulation and outputs the simulation status data of the topological positions of each sensor in the physical pipeline network. The digital twin model is constructed based on the topology, pipeline parameters and fluid dynamics equations of the physical pipeline network. The observation data collected by each sensor is compared point by point with the corresponding simulation state data to determine the real-time deviation vector including the deviation value of each point. The real-time deviation vector is then matched with multiple standard symptom vectors pre-stored in the fault knowledge graph. The standard symptom vectors are used to associate symptom nodes in the fault knowledge graph. Starting from the matched symptom node, the system traverses backward along the directed edges of the fault knowledge graph to locate at least one candidate fault node. The directed edges are associated with weight values, which are pre-assigned based on the simulation results of the digital twin model on the corresponding fault by quantifying the degree of impact of the fault on the symptom. The weights of directed edges on the reverse traversal path of each candidate fault node are aggregated to determine the comprehensive confidence score of each candidate fault node, and the potential fault type is determined based on the comprehensive confidence score.

2. The method according to claim 1, characterized in that, The observation data includes pressure data sequences, flow rate data sequences, and water tank level data sequences; the control command data includes valve opening data sequences and pump status data sequences; after acquiring the observation data and control command data, the method further includes: Based on the precision clock signals embedded in each data source, and using a preset control cycle as the reference time window, synchronous interpolation processing is performed on the pressure data sequence, flow data sequence, water tank level data sequence, valve opening data sequence, and water pump status data sequence to generate a status control dataset and a status observation dataset with a unified time sequence. Outlier removal and smoothing noise reduction processing are then performed on the datasets.

3. The method according to claim 2, characterized in that, Prior to determining the real-time deviation vector, the method further includes: The state observation dataset is subjected to a standardization transformation, which employs either the min-max normalization method or the Z-score normalization method. When the min-max normalization method is used, the global maximum and global minimum values ​​of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is then subtracted from the global minimum value and divided by the difference between the global maximum and global minimum values, so that the range of each physical quantity data sequence is mapped to the [0,1] interval. When using the Z-score standardization method, the arithmetic mean and standard deviation of each type of physical quantity data sequence in the state observation dataset are determined within a preset historical time interval. The value of each data point in each physical quantity data sequence is subtracted from the arithmetic mean and then divided by the standard deviation, so that each physical quantity data sequence is converted into standard normal distribution data.

4. The method according to claim 1, characterized in that, Each of the aforementioned standard symptom vectors is uniquely associated with a symptom node, and the matching of the real-time deviation vector with the standard symptom vector includes: Determine the similarity metric between the real-time deviation vector and each standard symptom vector in the spectrum; All calculated similarity metrics are compared with a preset matching threshold. The symptom nodes associated with standard symptom vectors whose similarity metrics are greater than the matching threshold, or the symptom nodes associated with the top N standard symptom vectors in terms of similarity metrics, are determined as the set of symptom nodes that are successfully matched, where N is an integer greater than 0.

5. The method according to claim 1, characterized in that, The process of constructing the fault knowledge graph includes: Based on the equipment list and historical maintenance records of the tunnel fire water system, a set of fault nodes is defined; Based on the monitoring point layout of the digital twin model, a set of symptom nodes is defined; Based on pre-defined domain expert experience or historical failure cases, directed edges with causal orientation are constructed between failure nodes and symptom nodes. By injecting the fault mode and severity represented by the fault node into the digital twin model, and obtaining the response deviation data of all monitoring points in the network after running the simulation, the steady-state change amplitude, dynamic response speed or probability distribution change of the monitoring point deviation of each symptom node under the fault mode is determined. The normalized value is calculated using the preset sensitivity analysis formula, and the value is used as the weight value attached to the directed edge connecting the fault node and the corresponding symptom node.

6. The method according to claim 1, characterized in that, The process of aggregating the weights of directed edges on the reverse traversal paths of each candidate fault node to determine the comprehensive confidence score of each candidate fault node includes: For a candidate fault node that has been located, a reverse traversal path is determined in the fault knowledge graph, starting from all the successfully matched symptom nodes and finally reaching the candidate fault node via directed edges. The path support of the faulty node is obtained by sequentially multiplying the weights of all directed edges on each reverse traversal path. If a candidate faulty node has multiple reverse traversal paths, the maximum value among the support scores of all reverse traversal paths is used as the overall confidence score of the candidate faulty node.

7. The method according to claim 6, characterized in that, The determination of potential fault types based on the comprehensive confidence score includes: Identify the candidate fault node with the highest overall confidence score, and directly determine the fault type it represents as the potential fault type; and / or, By combining the historical prior probability of the fault occurrence, the comprehensive confidence score is corrected by Bayesian posterior probability, and the fault type with the highest posterior probability is selected as the determination result.

8. The method according to claim 1, characterized in that, After determining the potential fault type, the method further includes: When the potential fault type is pipeline leakage or blockage, determine the target dimension in the real-time deviation vector whose numerical deviation is higher than a preset threshold. Based on the actual geographical location and topological connection relationship of the sensor corresponding to the target dimension in the physical pipeline network, the physical pipeline segment range where a fault may occur is determined.

9. The method according to claim 1, characterized in that, The method further includes: The potential fault types and their corresponding comprehensive confidence scores are encapsulated into a structured diagnostic report, output to a preset human-machine interface for visualization, and simultaneously stored in a historical database; Based on the potential fault types and the preset fault-response rule base, preliminary control suggestion instructions are generated and issued to suppress the impact of the fault.

Images