A pipeline coordination supervision and abnormality identification system based on multi-source data large model fusion

Abstract

This invention discloses a multi-source data large model fusion system for collaborative supervision and anomaly identification of utility tunnels, specifically relating to the field of intelligent operation and maintenance technology for utility tunnels. It includes a multimodal spatiotemporal unified coding module, a dynamic baseline self-learning module, a multidimensional anomaly joint judgment module, and a control coordination module. The multimodal spatiotemporal unified coding module constructs an environmental fingerprint feature matrix using a trend vectorization method; the dynamic baseline self-learning module constructs a differentiated environmental fingerprint baseline with embedded ventilation performance evaluation coefficients; the multidimensional anomaly joint judgment module performs triple verification of temporal drift detection, spatial correlation analysis, and visual-data cross-validation, outputting anomaly judgment results with confidence; the control coordination module integrates a knowledge graph with a vertical large model of the utility tunnel, automatically generating semantic reports on anomaly causes and cross-professional equipment linkage control commands. This invention can solve the problems of missed reporting, high false alarm rate, and low coordination efficiency of slowly changing anomalies in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for utility tunnels, and more specifically, to a collaborative monitoring and anomaly identification system for utility tunnels that integrates multi-source data and large-scale models. Background Technology

[0002] Underground utility tunnels serve as the "underground lifeline" for urban operations, housing a variety of urban pipelines, including electricity, communications, broadcasting and television, water supply, drainage, gas, and heating. Because these tunnels are located underground, are long and narrow, and relatively enclosed, any localized hazard can easily escalate into a major safety accident.

[0003] Traditional utility tunnel anomaly monitoring systems often rely on single sensor threshold alarms. In recent years, technologies such as patent CN118966581B have begun to incorporate artificial intelligence algorithms. These algorithms analyze environmental factors to eliminate sensor data errors and correct sensor anomalies based on the "ventilation performance evaluation" of each monitoring section, thereby improving the accuracy of anomaly detection. Furthermore, patent CN117212661A proposes a CIM-based underground integrated utility tunnel management platform. This platform utilizes inspection equipment running along overhead rails and a lifting CCD camera for high-level visual inspection, expanding the visual detection range and reducing detection gaps in physical space.

[0004] However, in actual high-risk and complex underground utility tunnel operation and maintenance scenarios, the aforementioned existing technologies still have certain limitations when facing in-depth collaborative supervision and complex hazard identification:

[0005] On the one hand, while the existing technology CN117212661A integrates vision, CIM, and traditional sensors, it is essentially still an "island-style" or "simple splicing" monitoring system where each data source operates independently. The numerical streams from IoT sensors and the pixel streams from inspection cameras lack a unified semantic representation at the underlying level, preventing the system from uncovering the nonlinear, deep collaborative information between "multi-dimensional temporal fluctuations, spatial physical correlations, and visual appearance representations." When faced with complex situations spanning multiple disciplines and regions, the existing technology cannot depict the overall evolution trend of the utility tunnel and struggles to support white-box logical reasoning and collaborative response.

[0006] On the other hand, existing technologies such as CN118966581B rely heavily on established mathematical correction models or specific rules for anomaly detection. However, in the actual environment of utility tunnels, many high-risk situations typically exhibit complex characteristics of "multiple measurement points, intertwined multiple physical fields, and gradual temporal variation." Single-point or piecewise correction models are severely inadequate in identifying such nonlinear characteristics, making them highly susceptible to missed alarms or high-frequency false alarms due to complex environmental noise. Furthermore, existing systems lack contextual understanding capabilities at the large-scale model level and cannot organically combine historical maintenance data and external heterogeneous environments to perform advanced attribution analysis of anomalies, resulting in a lack of intelligent decision support for the formulation of preventative maintenance strategies. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a multi-source data large model fusion-based collaborative supervision and anomaly identification system for utility tunnels, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-source data large model fusion-based system for collaborative monitoring and anomaly identification of utility tunnels, specifically comprising:

[0009] The multimodal spatiotemporal unified coding module is used to align the continuous time-series IoT sensor data and monitoring image sequences generated by each monitoring section in the target pipe gallery in terms of time and space dimensions through a sliding window. It also uses a trend vectorization method to calculate the multidimensional change feature vector of each measuring point data within the sliding time window, constructs an environmental fingerprint feature matrix with spatiotemporal correlation semantics, and maps the environmental fingerprint feature matrix into a structured text summary.

[0010] The dynamic baseline self-learning module is used to continuously learn the multi-dimensional environmental parameter characteristics of different monitoring sections based on historical normal operation data, and construct a dynamically updated normal operation environment fingerprint baseline. The environmental fingerprint baseline is embedded with ventilation performance evaluation coefficients for the differences in physical structure of different monitoring sections, which are used to perform baseline differentiation correction for gas diffusion and temperature and humidity field characteristics under different ventilation conditions.

[0011] The multidimensional anomaly joint determination module is used to compare the real-time generated environmental fingerprint feature matrix with the corresponding environmental fingerprint baseline, perform triple verification including time-domain drift detection, spatial correlation analysis and visual-data cross-validation, and combine the ventilation performance evaluation coefficient for comprehensive evaluation, and output the anomaly joint determination result with confidence score.

[0012] The control and coordination module is used to integrate the joint anomaly determination results output by the multi-dimensional anomaly joint determination module, the relevant structured text summary, and the historical operation and maintenance knowledge graph of the utility tunnel into a multimodal semantic context prompt word, which is then input into a preset vertical large model of the utility tunnel. The vertical large model of the utility tunnel performs advanced attribution analysis based on contextual understanding, and outputs a white-box semantic report on the causes of abnormal events and cross-disciplinary equipment collaborative linkage control instructions.

[0013] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system of the present invention, it includes the multimodal spatiotemporal unified coding module, which aligns the continuous time-series IoT sensor data and monitoring image sequences generated by each monitoring section within the target pipe gallery in terms of time and space dimensions through a sliding window, specifically including the following steps:

[0014] Time dimension alignment: Set a global sliding time window with a length of [value missing]. The sliding step size is Based on the current system clock, a length of [value] is extracted forward. The time segment is defined; for IoT sensors with sampling frequencies higher than the preset standard frequency, mean smoothing is used for frequency reduction and alignment; for IoT sensors with sampling frequencies lower than the preset standard frequency, linear interpolation algorithm is used for interpolation completion to ensure that all sensors are within the time window. It contains equidistant time series with the same dimensions, and synchronously extracts time windows. The corresponding frames of the internal monitoring image sequence;

[0015] Spatial Dimension Alignment: Based on the building information model of the utility tunnel, a unified three-dimensional spatial coordinate system is established, and the physical installation positions of fixed sensors and dynamic / static monitoring cameras in each monitoring section are pre-mapped into three-dimensional spatial coordinates. Based on the axial distance of the monitoring section, all data sources within the same section are labeled with a unified spatial region tag. This enables the binding of the "spatial domain".

[0016] Preferably, as a preferred embodiment of the multi-source data large model fusion system for collaborative monitoring and anomaly identification of utility tunnels according to the present invention, it includes the multimodal spatiotemporal unified coding module, which uses a trend vectorization method to capture the dynamic evolution characteristics of the data. Specifically, it includes: calculating the first derivative of the discrete time-series data of each sensor point within the current sliding time window to characterize the parameter change rate, the second derivative to characterize the change acceleration, and the discrete integral to characterize the cumulative effect, and then performing dimensionless processing on the above three calculation results and splicing them into the three-dimensional trend feature components of the current sensor point; the three-dimensional trend feature components are used to quantify the dynamic evolution of environmental parameters within the time window and the energy accumulation degree of slowly changing anomalies.

[0017] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative monitoring and anomaly identification system of the present invention, it includes the multimodal spatiotemporal unified coding module, which, after extracting the trend feature components of a single measuring point, groups the same monitoring section... The trend feature components of all measuring points within the tunnel are organized and arranged in a matrix according to spatial topological order, forming the overall environmental fingerprint feature matrix of the tunnel within the current time window, specifically including:

[0018] Horizontal stitching of IoT features: When M IoT measurement points are deployed within a corresponding segment, the trend feature components of these M measurement points are stitched together. By horizontally concatenating the segments according to their spatial topological order, a temporal feature matrix is ​​constructed. ;in, Represents the trend characteristic components of the Mth IoT measurement point within the segment; matrix The dimension is ;

[0019] Multimodal feature fusion: extraction time window Apparent change feature vector of internal monitoring image sequence and will Introduced as an additional column of the matrix, along with the time-series feature matrix. To merge;

[0020] Adaptive weight mapping: In the fusion process, an adaptive dynamic weight mapping function driven by visual semantics is introduced: When the semantic recognition result of the monitoring image sequence outputs an abnormal appearance, the dynamic weight mapping function automatically increases the feature component weight coefficient of the IoT measurement point at the corresponding three-dimensional coordinate point in space, so as to realize cross-modal semantic cross-enhancement between machine vision features and continuous time-series numerical features.

[0021] Matrix generation: within the current time window Within this matrix, a two-dimensional high-dimensional matrix is ​​formed, integrating temporal fluctuations, spatial topology, and visual appearance features, defined as the environmental fingerprint feature matrix. It can uniquely and comprehensively depict the multi-physics coupling situation of the utility tunnel section in the current time period.

[0022] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system of the present invention, it includes the dynamic baseline self-learning module, used to continuously learn the multi-dimensional environmental parameter characteristics of different monitoring sections based on historical normal operation data, and construct a dynamically updated normal operation environment fingerprint baseline; wherein, the environmental fingerprint baseline embeds ventilation performance evaluation coefficients for the differences in physical structure of different monitoring sections, used to perform baseline differentiation correction for gas diffusion and temperature and humidity field characteristics under different ventilation conditions; specifically including the following:

[0023] The dynamic baseline self-learning module constructs a dynamically updated normal operating environment fingerprint baseline, specifically including the following steps:

[0024] Collect environmental fingerprint feature matrices of all target utility tunnels that are confirmed to be in normal operating condition within a preset historical period. Unsupervised clustering algorithm is used to perform time-based and seasonal clustering on the collected environmental fingerprint feature matrix to obtain the normal state clusters of the utility tunnel under different spatiotemporal scenarios.

[0025] For each normal state cluster, an autoencoder neural network is used for reconstruction learning. By compressing and reconstructing the fingerprint of the normal environment, the probability distribution boundaries of the first derivative, second derivative, and cumulative integral value under normal fluctuation states are extracted, and the baseline envelope of the normal operating state is established. The baseline envelope defines the multidimensional permissible fluctuation range of each monitoring parameter under normal conditions.

[0026] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative monitoring and anomaly identification system of the present invention, it includes the dynamic baseline self-learning module, which quantifies the ventilation differences of each monitoring section by constructing a simplified fluid dynamics model of the pipe gallery's physical structure, specifically including:

[0027] For each monitoring section Extract its corresponding physical structural parameters, including: the rated air volume of the mechanical exhaust fan in the section. Axial distance of the nearest natural air inlet / make-up air inlet The cross-sectional area A of the cabin and the air resistance coefficient of the section. ;

[0028] A physical formula was constructed to characterize the air convection and dilution capacity of a monitoring section, and the ventilation performance evaluation coefficient of the monitoring section was calculated. ;in, The normalized adjustment coefficient is used to make the ventilation performance evaluation coefficient... The range of values ​​is normalized to ; The closer the value is to 1.5, the closer the section is to the ventilation opening, the higher the air velocity, and the easier it is for heat and gas to dissipate; The closer the value is to 0.5, the more likely that the section is in a long-distance ventilation dead zone;

[0029] The calculated ventilation performance evaluation coefficient As a correction gateway, it is embedded into the corresponding environmental fingerprint baseline to perform nonlinear correction on the judgment boundaries of different segments and establish adaptive judgment boundaries; the specific correction mechanism is as follows:

[0030] Correction for cumulative gas concentration limits: for ventilation performance evaluation coefficients For poorly ventilated sections that tend to the lower limit, the cumulative integral boundary of gas concentration in their environmental fingerprint is widened proportionally to tolerate trace amounts of gas accumulated due to poor air circulation and to intercept background false alarms.

[0031] Temperature rise rate limit correction: for ventilation performance evaluation coefficient For the section adjacent to the wind turbine that is approaching the upper limit, the temperature first and second derivative boundaries in the environmental fingerprint of this section are tightened to sensitively detect abnormal temperature rises.

[0032] Differential baseline matrix generation: Through the above mechanism, tensor correction operators are used. The unified baseline envelope The specific ventilation performance evaluation coefficient of the section Perform fusion calculations to generate a differentiated environmental fingerprint baseline matrix with specific segment physical structure characteristics. ,in, Spatial identifiers for utility tunnel sections are used to eliminate the interference of heterogeneous spatial environments on the accuracy of anomaly identification.

[0033] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system of the present invention, it includes the multi-dimensional anomaly joint judgment module, which is used to compare the real-time generated environmental fingerprint feature matrix with the corresponding environmental fingerprint baseline, perform triple verification including temporal drift detection, spatial correlation analysis, and visual-data cross-validation, and combine it with the ventilation performance evaluation coefficient for comprehensive evaluation, and output the anomaly joint judgment result with confidence score, specifically including the following:

[0034] Set the current time window Real-time generated environmental fingerprint feature matrix Input the differential environment fingerprint baseline matrix and perform the following triple verification synchronously in parallel or serially:

[0035] The first layer of verification: temporal drift detection, used to capture slow-changing potential hazards caused by minor leaks or localized dampness. Specifically, it calculates the topological distance between the IoT temporal feature portion of the real-time environmental fingerprint feature matrix and the corresponding baseline envelope. If, across multiple consecutive sliding time windows, the topological distance continuously exceeds a preset safety confidence interval, and the first derivative of the environmental parameters exhibits a long-period nonlinear monotonic deviation trend, it is determined that the system has a slow-changing anomaly in the temporal domain, and a temporal anomaly evaluation value is output. ;

[0036] The second layer of verification: spatial correlation analysis, used to capture spatially distributed anomalies coupled across multiple monitoring points; based on the spatial topology of the utility tunnel, the current monitoring section is extracted. and its upstream section and downstream section The real-time environmental fingerprint feature matrix is ​​used to calculate the correlation feature matrix across the spatial span; the determination logic is: current segment The first derivative of temperature increases positively and simultaneously triggers the downstream section. The cumulative integral value of humidity increases in the same direction, identifying a multi-physics spatial collaborative change pattern, determining that a multi-point spatial composite anomaly has occurred, and outputting a spatial anomaly evaluation value. ;

[0037] The third layer of verification: visual-data cross-validation, used to combat environmental noise and intercept potential false alarms; the specific process is as follows: extracting the current time window. Feature vectors of apparent changes in monitoring images that are in the same spatiotemporal alignment domain as IoT data The cross-validation logic is as follows: when the temporal drift detection of the first layer of validation or the spatial correlation analysis of the second layer of validation indicates the presence of abnormal temperature, smoke, or gas, and the image appearance change feature vector, after semantic recognition and parsing, shows that there is no personnel intrusion, no open flame appearance, and no smoke particle diffusion pattern in the corresponding target physical area, it is determined that the current data fluctuation is likely due to environmental background noise or single-point sensor drift, triggering the visual weighting operator. ;

[0038] After completing the above triple verification, the time-domain anomaly evaluation value is calculated using the anomaly comprehensive scoring function. The spatial anomaly evaluation value and the aforementioned visual weighting operator By coupling the parameters, a comprehensive score is calculated, as shown in the following formula: ;in, and These are the preset first and second weight coefficients, respectively; For comprehensive scoring; when the false alarm interception logic of the vision-data cross-validation is triggered, the vision weighting operator... Assign a preset low value to the comprehensive score. It directly reduces the noise level to below the preset safety warning line, thereby enabling intelligent interception of false alarms about environmental noise.

[0039] Utilizing the ventilation performance evaluation coefficient of the corresponding monitoring section Overall score The final correction is performed, and the joint anomaly determination result with confidence score is output. The specific formula is as follows: ;in, This is a mapping function that is inversely proportional to ventilation performance.

[0040] Preferably, as a preferred embodiment of the multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system of the present invention, it includes the control and coordination module, which fuses and constructs multimodal semantic contextual prompts and inputs them into a preset vertical large model of the pipe gallery, specifically including the following:

[0041] The system automatically retrieves heterogeneous information from the currently triggered alarm section and splices it according to a preset semantic template. The heterogeneous information includes quantitative perception features, spatiotemporal semantic summaries, and historical mechanism backgrounds. The quantitative perception features include the final confidence score, temporal anomaly evaluation value, spatial anomaly evaluation value, and corresponding ventilation performance evaluation coefficients. The spatiotemporal semantic summaries include structured text summaries of environmental feature trends and visual feature appearances. The historical mechanism backgrounds include the types, years of operation, maintenance records, and high-incidence fault modes of various professional pipelines within the section retrieved from the historical operation and maintenance knowledge graph of the utility tunnel based on the spatial area label of the alarm section.

[0042] The spliced ​​heterogeneous information elements are embedded into the industrial role setting and reasoning constraint framework to generate a text complex consisting of a role definition segment, a spatiotemporal awareness data segment, a knowledge graph background segment, and a reasoning format constraint segment, which serves as a multimodal semantic context prompt for the vertical large model of the pipe gallery.

[0043] The control coordination module also includes a hard-coded physical mechanism verification gateway for constructing a dual-loop security control mechanism, specifically including:

[0044] Instruction parsing: Using a fixed lightweight named entity recognition operator, the action subject, target operating equipment, and quantized control amplitude in the text linkage control instructions generated by the vertical large model of the pipe gallery are extracted and mapped to the physical address of the register of the underlying field programmable logic controller (PLC) and the written value, and converted into a standard control code string; The lightweight named entity recognition operator constructs a triplet of operation verb-equipment object-quantization threshold, compares it with a preset industrial equipment mapping table, and absolutely maps the text instruction to the physical address of the corresponding PLC register;

[0045] Full-scale closing verification: Before the standard control code string is sent to the PLC controller on site, a full-scale closing verification is performed through a physical mechanism verification gateway; the physical mechanism verification gateway has hard-coded absolute regulations for fire safety in pipe corridors, equipment interlocking logic, and hard constraints on physical boundaries, and its text review rules include:

[0046] Taboo rule verification: Verify whether there are contradictory instructions, including verifying whether the smoke exhaust valve and fire damper are simultaneously closed before the fire is confirmed to be extinguished;

[0047] Boundary hard constraint verification: Verify whether the generated equipment adjustment exceeds the physical load limit of the corresponding equipment;

[0048] Different branch controls are executed based on the verification results, specifically:

[0049] Branch A: Indicates that the verification is successful. When the control code string fully complies with the hard-coded safety procedures, the physical mechanism verification gateway performs the closing and release, and officially sends the standard control code string to the corresponding professional PLC controller, driving the hardware device to perform cross-professional collaborative blocking.

[0050] Branch B: Represents verification interception and reflection. When the standard control code string violates any safety procedure, the physical mechanism verification gateway triggers the interception and blocking mechanism, and converts the specific reason for the violation and the violated physical rule into reverse correction text feedback, which is then re-inputted into the vertical large model of the pipe gallery as a follow-up prompt. This forces the vertical large model of the pipe gallery to perform secondary instruction reasoning and regenerate the linkage control instruction until it passes the verification of the physical mechanism verification gateway.

[0051] Branch B incorporates a dynamic step-count blocking and physical safety fallback switching mechanism: the physical mechanism verification gateway monitors in real time the number of reverse correction and inquiry rounds for the same abnormal event; when the vertical large model of the utility tunnel fails to pass the full-scale closing verification after experiencing the preset maximum number of reflection rounds, the physical mechanism verification network closes the closed-loop control authority of the vertical large model of the utility tunnel, instantly activates the cross-professional safety fallback mechanism control strategy hard-coded in the field PLC, and sends a manual takeover request to the central operation and maintenance terminal; the maximum number of reflection rounds is preferably 3 rounds.

[0052] On the other hand, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements the functional modules of a multi-source data large model fusion pipeline collaborative supervision and anomaly identification system as described above.

[0053] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements a multi-source data large model fusion system for collaborative supervision and anomaly identification of utility tunnels as described above.

[0054] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0055] By extracting three-dimensional features of change rate, acceleration, and cumulative effect using trend vectorization, and replacing the traditional single-point absolute value threshold judgment, it can identify potential leaks and structural micro-deformation hazards in their early stages, enabling early warning. Differentiated environmental fingerprint baselines are used to specifically correct sections with different ventilation conditions, and combined with a triple verification mechanism, it can effectively filter false alarms caused by environmental noise and normal sensor drift. By integrating anomaly judgment results, environmental fingerprint text summaries, and operation and maintenance knowledge graphs through a large vertical model of the utility tunnel, it automatically generates interpretable cause reports and cross-disciplinary equipment linkage control commands, eliminating the need for manual analysis and significantly shortening emergency response time. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0057] Figure 1 This is a flowchart of a method for a collaborative monitoring and anomaly identification system for utility tunnels based on the fusion of multi-source data and large models, according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0059] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0060] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0061] Example 1

[0062] This embodiment provides, for example Figure 1 The system shown is a collaborative monitoring and anomaly identification system for utility tunnels based on multi-source data and large-scale model fusion, specifically including:

[0063] The multimodal spatiotemporal unified coding module is used to align the continuous time-series IoT sensor data and monitoring image sequences generated in each monitoring section of the target pipe gallery with time and space dimensions through a sliding window, and to calculate the multidimensional change feature vector of each measuring point data within the sliding time window using a trend vectorization method, construct an environmental fingerprint feature matrix with spatiotemporal correlation semantics, and map the environmental fingerprint feature matrix into a structured text summary.

[0064] The dynamic baseline self-learning module is used to continuously learn the multi-dimensional environmental parameter characteristics of different monitoring sections based on historical normal operation data, and construct a dynamically updated normal operation environment fingerprint baseline. The environmental fingerprint baseline is embedded with ventilation performance evaluation coefficients for the differences in physical structure of different monitoring sections, which are used to perform baseline differentiation correction for gas diffusion and temperature and humidity field characteristics under different ventilation conditions.

[0065] The multidimensional anomaly joint determination module is used to compare the real-time generated environmental fingerprint feature matrix with the corresponding environmental fingerprint baseline, perform triple verification including time-domain drift detection, spatial correlation analysis and visual-data cross-validation, and combine the ventilation performance evaluation coefficient for comprehensive evaluation, and output the anomaly joint determination result with confidence score.

[0066] The control and coordination module is used to integrate the joint anomaly determination results output by the multi-dimensional anomaly joint determination module, the relevant structured text summary, and the historical operation and maintenance knowledge graph of the utility tunnel into a multimodal semantic context prompt word, which is then input into a preset vertical large model of the utility tunnel. The vertical large model of the utility tunnel performs advanced attribution analysis based on contextual understanding, and outputs a white-box semantic report on the causes of abnormal events and cross-disciplinary equipment collaborative linkage control instructions.

[0067] In this embodiment, the multimodal spatiotemporal unified coding module is specifically described. This module aligns the continuous time-series IoT sensor data and monitoring image sequences generated by each monitoring section within the target utility tunnel using a sliding window. It then uses a trend vectorization method to calculate the multidimensional change feature vectors of the data at each measuring point within the sliding time window, constructing an environmental fingerprint feature matrix with spatiotemporal semantics. This environmental fingerprint feature matrix is ​​then mapped to a structured text summary. By reducing the dimensionality of the high-dimensional environmental fingerprint feature matrix into a sparse structured text summary, it serves as a multimodal context. This reduces the computational complexity of the attention mechanism in processing long-cycle industrial time-series data for the large vertical model of the utility tunnel and suppresses hallucination phenomena. Specifically, it includes the following:

[0068] By aligning time and space dimensions, and eliminating issues such as inconsistent sampling frequencies, asynchronous clocks, and physical installation discrepancies between different sensors and monitoring cameras, the following steps are performed:

[0069] Time dimension alignment: Set a global sliding time window with a length of [value missing]. The sliding step size is Based on the current system clock, a length of [value] is extracted forward. The time segment is defined; for IoT sensors with sampling frequencies higher than the preset standard frequency, mean smoothing is used for frequency reduction and alignment; for IoT sensors with sampling frequencies lower than the preset standard frequency, linear interpolation algorithm is used for interpolation completion to ensure that all sensors are within the time window. It contains equidistant time series with the same dimensions, and synchronously extracts time windows. The corresponding frames of the internal monitoring image sequence;

[0070] Spatial Dimension Alignment: Based on the building information model of the utility tunnel, a unified three-dimensional spatial coordinate system is established, and the physical installation positions of fixed sensors and dynamic / static monitoring cameras in each monitoring section are pre-mapped into three-dimensional spatial coordinates. Based on the axial distance of the monitoring section, all data sources within the same section are labeled with a unified spatial region tag. This achieves the binding of the "spatial domain";

[0071] After aligning the time and spatial dimensions, a trend vectorization method is used to capture the dynamic evolution characteristics of the data to combat the absolute value drift caused by seasonal fluctuations in the sensor environment. The specific calculation process is as follows:

[0072] Step 1: Represent the discrete time sequence of the measuring point sensor within the current time window as follows: ,in, This represents the measured environmental parameters of the sensor at the nth sampling time, including but not limited to physical quantities such as temperature, humidity, gas concentration, and vibration amplitude.

[0073] Step 2, First derivative calculation: Calculate the difference between adjacent sampling points to obtain the first derivative sequence. This is used to characterize the current rate of change of the environmental parameter, where, Indicates the first The difference value is calculated using the following formula: This reflects the magnitude and direction of parameter changes at the current measurement point between adjacent sampling times;

[0074] Step 3, Second Derivative Calculation: For the sequence of first derivatives By further difference, we obtain the sequence of second derivatives. This is used to characterize whether the trend of change is in an accelerating or slowing-down / stabilizing phase; and is based on the second derivative sequence. Distinguishing between abrupt and gradual hazard types: When classified as abrupt hazard, the second derivative sequence exhibits an instantaneous extremum; when classified as gradual hazard, the second derivative sequence shows a continuous and stable trend. Through the physical feature mapping of multi-order differentials, a data representation basis is formed for subsequent accurate capture of the evolution mechanism of complex hazards.

[0075] Step 4: Cumulative Integral Calculation: Calculate the discrete time series sequence In the time window Integral value below the curve This is used to characterize the cumulative effect of energy or matter within a given period, where... Indicates the sampling time interval. For the measuring point sensor at the first Environmental parameter measurements at each sampling time;

[0076] Step 5, Dimensionless Processing: The first derivative sequence obtained above is processed... Second derivative sequence The cumulative integral value S is Z-score normalized to eliminate the influence of absolute values ​​of different physical dimensions and form the trend characteristic component of the measurement point. ,in, Describe the sequence of first derivatives The Z-score standardized value represents the rate of change of the environmental parameter at the current moment; Represents the sequence of second derivatives The Z-score, after standardization, represents the acceleration characteristic of the changing trend; a positive value indicates accelerated change, and a negative value indicates deceleration. This represents the cumulative integral value after Z-score standardization, characterizing the overall cumulative effect of environmental parameters within a time window, and is used to quantify the energy of slowly changing anomalies.

[0077] After extracting the trend feature components of a single measuring point, the same monitoring section The trend characteristic components of all measuring points within the tunnel are organized and matrixed according to spatial topological order to form the overall environmental fingerprint vector of the tunnel within the current time window, specifically including:

[0078] Horizontal stitching of IoT features: When M IoT measurement points are deployed within a segment, the trend feature components of these M measurement points are stitched together. By horizontally concatenating the segments according to their spatial topological order, a temporal feature matrix is ​​constructed. ;in, Represents the trend characteristic components of the Mth IoT measurement point within the segment; matrix The dimension is ;

[0079] Multimodal feature fusion: extraction time window Apparent change feature vector of internal monitoring image sequence and will Introduced as an additional column of the matrix, along with the time-series feature matrix. To merge;

[0080] Adaptive weight mapping: In the fusion process, an adaptive dynamic weight mapping function driven by visual semantics is introduced: When the semantic recognition result of the monitoring image sequence outputs an abnormal appearance, the dynamic weight mapping function automatically increases the feature component weight coefficient of the IoT measurement point at the corresponding three-dimensional coordinate point in space, so as to realize cross-modal semantic cross-enhancement between machine vision features and continuous time-series numerical features.

[0081] Matrix generation: within the current time window Within this matrix, a two-dimensional high-dimensional matrix is ​​formed, integrating temporal fluctuations, spatial topology, and visual appearance features, defined as the environmental fingerprint feature matrix. It can uniquely and comprehensively depict the multi-physics coupling situation of the utility tunnel section in the current time period.

[0082] In this embodiment, the dynamic baseline self-learning module is specifically described. This module continuously learns the multi-dimensional environmental parameter characteristics of different monitoring sections based on historical normal operation data, constructing a dynamically updated normal operation environment fingerprint baseline. The environmental fingerprint baseline embeds ventilation performance evaluation coefficients tailored to the differences in physical structure of different monitoring sections, used for baseline-differentiated correction of gas diffusion and temperature / humidity field characteristics under different ventilation conditions. Specifically, it includes the following:

[0083] Collect environmental fingerprint feature matrices of all target utility tunnels that are confirmed to be in normal operating condition within a preset historical period. Unsupervised clustering algorithm is used to perform time-based and seasonal clustering on the collected feature matrix to obtain the normal state clusters of the utility tunnel under different spatiotemporal scenarios;

[0084] For each normal state cluster, an autoencoder neural network is used for reconstruction learning. By compressing and reconstructing the fingerprint of the normal environment, the probability distribution boundaries of the first derivative, second derivative, and cumulative integral value under normal fluctuation states are extracted, and the baseline envelope of the normal operating state is established. The baseline envelope defines the multidimensional permissible fluctuation range of each monitoring parameter under normal conditions.

[0085] By constructing a simplified fluid dynamics model of the physical structure of the utility tunnel, the ventilation differences in each monitoring section are quantified, further including:

[0086] For each monitoring section Extract its corresponding physical structural parameters, including: the rated air volume of the mechanical exhaust fan in the section. Axial distance of the nearest natural air inlet / make-up air inlet The cross-sectional area A of the cabin and the air resistance coefficient of the section. ;

[0087] A physical formula was constructed to characterize the air convection and dilution capacity of a monitoring section, and the ventilation performance evaluation coefficient of the monitoring section was calculated. ;in, The normalized adjustment coefficient is used to make the ventilation performance evaluation coefficient... The range of values ​​is normalized to ; The closer the value is to 1.5, the closer the section is to the ventilation opening, the higher the air velocity, and the easier it is for heat and gas to dissipate; The closer the value is to 0.5, the more it indicates that the section is in a long-distance ventilation dead zone with extremely poor convection;

[0088] The calculated ventilation performance evaluation coefficient As a correction gateway, it is embedded into the corresponding environmental fingerprint baseline to perform nonlinear correction on the judgment boundaries of different segments and establish adaptive judgment boundaries; the specific correction mechanism is as follows:

[0089] Gas concentration accumulation boundary correction: For poorly ventilated areas, due to weak natural air leakage and convection, the predicted gas accumulation effect under normal conditions is relatively strong. In the dynamic baseline, the cumulative integral boundary of gas concentration in the environmental fingerprint is widened proportionally to avoid false alarms caused by trace amounts of gas accumulated due to poor daily air circulation.

[0090] Temperature rise rate boundary correction: For sections adjacent to the fan, the first and second derivative temperature boundaries in the environmental fingerprint of that section are tightened; when abnormal temperature rise occurs in the section, the potential for local overheating is sensitively detected.

[0091] Differentiated baseline matrix generation: Through the above mechanism, a differentiated environmental fingerprint baseline matrix with specific segment physical structure characteristics is generated. This eliminates the interference of spatially heterogeneous environments on the accuracy of anomaly identification. The baseline envelope, Spatial identifiers representing utility tunnel sections. This is a tensor correction operator used to unify the baseline envelope. The specific ventilation performance evaluation coefficient of the section Perform fusion calculations.

[0092] In this embodiment, the multi-dimensional anomaly joint determination module needs to be specifically described. This module compares the real-time generated environmental fingerprint feature matrix with the corresponding environmental fingerprint baseline, performs triple verification including temporal drift detection, spatial correlation analysis, and visual-data cross-validation, and combines this with the ventilation performance evaluation coefficient for comprehensive evaluation, outputting anomaly joint determination results with confidence scores. Specifically, the results include the following:

[0093] Set the current time window Real-time generated environmental fingerprint feature matrix Input the differential environmental fingerprint baseline matrix and perform the following triple verification synchronously in parallel or serially:

[0094] The first layer of verification: time-domain drift detection, used to capture "time-domain slowly changing hidden dangers" caused by minute leaks and localized moisture. The specific process is as follows: calculate the topological distance between the IoT time-series feature part in the real-time environmental fingerprint feature matrix and the corresponding baseline envelope. When the topological distance continuously exceeds the preset safety confidence interval across multiple consecutive sliding time windows, and the first derivative of the environmental parameters shows a long-period nonlinear monotonic deviation trend, it is determined that the system has a slowly changing anomaly in the time domain, and a time-domain anomaly evaluation value is output. The calculation of the topological distance introduces Mahalanobis distance metric based on the physical meaning of the covariance matrix. By using the covariance matrix of historical normal operation data to decouple multidimensional environmental parameters, the original sporadic statistical correlation between various physical quantities is eliminated. When the Mahalanobis distance monotonically increases over a long period and its integral slope is continuously greater than zero, the instantaneous jumps caused by single-point random noise are accurately identified, locking in the unique temporal hidden accumulation risk of underground enclosed spaces.

[0095] The second layer of verification: spatial correlation analysis, used to capture "spatial distributed anomalies" coupled across multiple monitoring points; based on the spatial topology of the utility tunnel, the current monitoring section is extracted. and its upstream section and downstream section The real-time environmental fingerprint feature matrix is ​​used to calculate the correlation feature matrix across the spatial span; the determination logic is: current segment The first derivative of temperature increases positively and simultaneously triggers the downstream section. The cumulative integral value of humidity increases in the same direction, identifying a multi-physics spatial collaborative change pattern, determining that a multi-point spatial composite anomaly has occurred, and outputting a spatial anomaly evaluation value. The associated feature matrix is ​​established by calculating the spatial dynamic cross-correlation coefficient of the feature matrices of adjacent sections; the judgment logic includes the co-evolution of temperature and humidity, and also has an embedded convection transmission delay correction mechanism for the gas flow field in the pipe gallery. By nonlinearly mapping the abnormal release time of the upstream section with the flow velocity driven by the ventilation performance evaluation coefficient, the delay window of the downstream section feature occurrence is accurately matched, realizing the physical attribution of cross-professional and cross-regional spatial evolution trajectory.

[0096] The third layer of verification: visual-data cross-validation, used to combat environmental noise caused by sensor aging, seasonal drift, and vibrations from vehicles on the external corridor, and to intercept potential false alarms; the specific process is as follows: extracting the current time window. Feature vectors of apparent changes in monitoring images that are in the same spatiotemporal alignment domain as IoT data The cross-validation logic is as follows: when the temporal drift detection of the first layer of validation or the spatial correlation analysis of the second layer of validation indicates the presence of abnormal temperature, smoke, or gas, and the image appearance change feature vector, after semantic recognition and parsing, shows that there is no personnel intrusion, no open flame appearance, and no smoke particle diffusion pattern in the corresponding target physical area, it is determined that the current data fluctuation is likely due to environmental background noise or single-point sensor drift, triggering the visual weighting operator. ;

[0097] After completing the above triple verification, the verification results are nonlinearly fused and the ventilation performance evaluation coefficient is introduced. Gain and attenuation corrections are applied, and the final output is a confidence score with deterministic physical semantics, which further includes:

[0098] Construct an anomaly comprehensive scoring function to evaluate time-domain anomalies. Spatial anomaly evaluation value and visual weighting operators For coupling, the preferred calculation expression is as follows: ;in, and These are the first and second weight coefficients, respectively; For comprehensive scoring; when the false alarm interception logic of the vision-data cross-validation is triggered, the vision weighting operator assigns a preset low value, directly reducing the comprehensive score to below the preset safety threshold, thereby achieving intelligent interception of false alarms caused by environmental noise;

[0099] Utilizing the ventilation performance evaluation coefficient of the corresponding monitoring section Overall score The final correction is performed, and the joint anomaly determination result with confidence score is output. The specific formula is as follows: ;in, This is a mapping function that is inversely proportional to ventilation performance.

[0100] In this embodiment, the control and coordination module is specifically described. This module integrates the anomaly joint determination results output by the multi-dimensional anomaly joint determination module, the relevant structured text summary, and the historical operation and maintenance knowledge graph of the utility tunnel into a multimodal semantic contextual prompt, which is then input into a preset vertical utility tunnel model. The vertical utility tunnel model performs advanced attribution analysis based on contextual understanding, outputting a white-box semantic report on the causes of anomalies and cross-disciplinary equipment collaborative control instructions, specifically including the following:

[0101] The system automatically retrieves heterogeneous information from the currently triggered alarm segment and concatenates it according to a preset semantic template; the heterogeneous information includes:

[0102] Quantitative sensing features: the final confidence score, time-domain anomaly evaluation value, spatial anomaly evaluation value, and corresponding ventilation performance evaluation coefficient output by the multi-dimensional anomaly joint judgment module;

[0103] Spatiotemporal semantic summarization: The structured text summary generated by the multimodal spatiotemporal unified coding module includes environmental feature trends and visual feature appearances;

[0104] Historical Mechanism Background: Based on the spatial area label of the alarm section, the types, years of operation, maintenance records and high-incidence fault modes of various professional pipelines in the section are retrieved from the knowledge graph of the historical operation and maintenance of the utility tunnel.

[0105] The above-mentioned spliced ​​elements are embedded into the industrial role setting and reasoning constraint framework to generate multimodal semantic context prompts for the vertical large model of the pipe gallery; its structured form is preferably a text complex composed of "role definition segment + spatiotemporal perception data segment + knowledge graph background segment + reasoning format constraint segment";

[0106] It should be noted that the pre-set vertical large model of the utility tunnel is not a general semantic large model, but a specialized industrial large model that has been deeply supervised and fine-tuned and aligned with reinforcement learning using corpus of multi-physics coupling evolution mechanism of utility tunnels, cross-professional control strategy texts and historical accident evolution chain datasets. By embedding the fluid dynamics and thermodynamics mechanisms of underground enclosed spaces into the parameter space of the large model in the form of internalized weights, it has the basic cognitive ability to perform physical logic fidelity reasoning in the context of dense industry.

[0107] The vertical large-scale model of the utility tunnel, based on received multimodal semantic contextual prompts, utilizes its deep dynamic attention mechanism and rich vertical domain corpus knowledge to conduct white-box contextual understanding, specifically generating semantic reports on the causes of abnormal events and cross-disciplinary equipment collaborative control commands:

[0108] A hard-coded physical mechanism verification gateway is introduced to construct a dual-loop security mechanism, specifically including:

[0109] Instruction parsing: Using a fixed lightweight named entity recognition operator, the text linkage control instructions generated by the vertical large model of the pipe gallery are parsed and converted into standardized control code strings that can be directly executed by the underlying field programmable logic controller; the lightweight named entity recognition operator constructs a triplet of operation verb-equipment object-quantization threshold, compares it with a preset industrial equipment mapping table, and absolutely maps the text instructions to the physical address of the corresponding PLC register;

[0110] Full-scale closing verification: Before the control code string is officially sent to the PLC controller at the utility tunnel site, it must pass through the physical mechanism verification gateway. This gateway has hard-coded absolute regulations for utility tunnel fire safety, equipment interlocking logic, and hard physical boundary constraints. The verification gateway performs a full-scale closing verification on the instructions generated by the large model, and its preferred text review rules include:

[0111] Taboo rule verification: Verify whether there are contradictory instructions, and whether the large model instructions erroneously triggered the simultaneous closure of the smoke exhaust valve and the fire damper before the fire is confirmed to be extinguished;

[0112] Boundary hard constraint verification: Verify whether the equipment adjustment amount generated by the large model exceeds the physical load-bearing limit of the equipment;

[0113] Different branch controls are executed based on the verification results, specifically:

[0114] Branch A: Indicates that the verification is successful. When the control code string fully complies with the hard-coded safety procedures, the gateway performs the closing and release, and formally issues the control command to the corresponding professional PLC controller, driving the hardware device to perform cross-professional collaborative blocking.

[0115] Branch B: Represents verification, interception, and reflection. When the control code string violates any safety procedure, the gateway immediately triggers the interception and blocking mechanism, strictly prohibiting the issuance of erroneous commands. The gateway converts the specific reasons for the violation and the violated physical rules back into reverse correction text feedback, which is input into the vertical large model of the pipe gallery as a follow-up prompt, forcing the large model to perform secondary command reasoning. After the large model is corrected based on the feedback, it regenerates compliant linkage control commands until it passes the gateway verification.

[0116] Furthermore, to prevent the large model from generating a logical deadlock during the interaction in branch B, causing control lag, this module embeds a dynamic step-count blocking and physical safety fallback switching mechanism in branch B:

[0117] Real-time monitoring of the number of reverse correction rounds for the same abnormal event; when the large model has undergone the preset maximum number of reflection rounds, preferably 3 rounds, and the correction command it generates still cannot pass the closing verification of the physical mechanism verification gateway, the gateway will completely lock the closed-loop control authority of the large model, and instantly activate the cross-professional safety fallback mechanism control strategy hard-coded in the field PLC, and send the highest level of manual takeover strong synchronization request to the central operation and maintenance terminal, thus completely eliminating the safety threat to the operation of industrial-grade lifeline caused by the illusion and uncontrollability of the large model algorithm from the bottom layer of the engineering structure.

[0118] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0119] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the functional modules of a multi-source data large model fusion pipeline collaborative supervision and anomaly identification system as proposed in the above embodiment. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0120] Example 2

[0121] The following is another embodiment of the present invention, which provides a collaborative monitoring and anomaly identification system for utility tunnels based on the fusion of multi-source data and large model. In order to verify the beneficial effects of the present invention, a simulation experiment is conducted for scientific demonstration.

[0122] This experiment aims to verify the effectiveness of a multi-source data large-scale model fusion system for collaborative supervision and anomaly identification of underground utility tunnels. Through multimodal spatiotemporal unified coding, dynamic baseline self-learning, multi-dimensional anomaly joint judgment, and control coordination techniques, it enhances the early warning, advanced attribution, and control coordination capabilities for in-depth collaborative supervision and identification of complex hidden dangers in underground utility tunnels. The experiment uses simulated and actual collected utility tunnel operation data, including trend feature components under different anomaly states, real-time environmental fingerprint feature matrix, ventilation performance evaluation coefficients, final confidence scores after triple verification, structured text summaries, and knowledge graph backgrounds. By analyzing the system's anomaly identification, large-scale model advanced attribution, and the consistency between the security gateway verification control results and actual operation and maintenance labels, the accuracy and robustness of the system in identifying complex, multi-dimensional, slowly changing hidden dangers, intercepting false alarms from environmental noise, and intelligent cross-disciplinary linkage control are verified.

[0123] The simulation experiment steps are implemented according to the content of the multi-source data large model fusion pipe gallery collaborative supervision and anomaly identification system provided in Example 1, and the specific steps include:

[0124] By using a global sliding time window, the continuous time-series IoT sensor data and monitoring image sequences generated in each monitoring section of the target utility tunnel are aligned in time and space. The trend vectorization method is used to calculate the first derivative, second derivative, and cumulative integral value of the numerical changes of each measuring point within the sliding time window. After Z-score dimensionless transformation, single-point aging noise is eliminated through spatiotemporal consistency verification. The data is then horizontally stitched according to the spatial topology order. At the same time, adaptive dynamic weights driven by visual semantics are incorporated to construct an environmental fingerprint feature matrix with spatiotemporal semantic association. This matrix is ​​then transformed into a structured text summary through a semantic mapping operator. The acquisition and mapping frequency is set to once every 10 seconds.

[0125] The environmental fingerprint feature matrix under historical normal conditions is collected and clustered by time period and season. The probability distribution boundary of normal fluctuation is extracted by autoencoder neural network, and a baseline envelope is established. The ventilation performance evaluation coefficient of each section is calculated based on the fluid dynamics simplified model of the pipe gallery physical structure. The baseline envelope is nonlinearly corrected by tensor correction operator to generate a differentiated environmental fingerprint baseline matrix with specific section physical structure characteristics.

[0126] The real-time generated environmental fingerprint feature matrix and the corresponding differentiated environmental fingerprint baseline matrix are synchronously input into the judgment module for triple verification: capturing slowly changing hazards in the time domain through time domain drift detection; capturing distributed composite anomalies across sections through spatial correlation analysis; and intercepting false alarms of single-point environmental noise and triggering semantic weighting through visual-data cross-validation. Finally, nonlinear correction is performed by combining ventilation performance evaluation coefficients, and the joint anomaly judgment result with final confidence score is output.

[0127] The system automatically retrieves the quantitative sensing features, spatiotemporal semantic summaries, and historical mechanism background of the alarm section, integrates and encapsulates them into multimodal semantic context prompts, and inputs them into the vertical large model of the utility tunnel. The large model performs contextual understanding and outputs semantic reports on the causes of abnormal events and control instructions for equipment coordination. The generated control instructions are parsed into PLC standardized control code strings by a lightweight named entity recognition operator. Before being issued, the system executes branch A or branch B control through a physical mechanism verification gateway with hard-coded safety procedures.

[0128] The specific data from the above simulation experiment are shown in Table 1.

[0129] Table 1 Data Recording Table of Simulation Experiment of the Invention

[0130] 0-5 Cabin No. 1 - 050 meters The first derivative of the temperature is stable, and there is no gas accumulation; the image shows no abnormalities. 1.45 (adjacent to the mechanical fan, strong convection zone) Normal / Normal / Normal 12% (Low Risk) The system is operating normally and there are no obvious security risks. Branch A (allow passage) Maintain current equipment operating parameters and continue routine system monitoring. 5-10 Cabin No. 2 - 120 meters The cumulative humidity component increased slightly, with occasional pulse fluctuations observed; vibrations were visible in the footage as vehicles passed over the walkway. 0.95 (Standard ventilation zone) Normal (excluding transient jumps) / Normal / Triggered semantic deweighting (no water mist) 18% (Low Risk) An occasional data jump was detected and determined to be single-point environmental noise caused by external vibrations of the vehicle. Branch A (allow passage) Intercept false positives caused by data noise and continue rolling monitoring. 10-15 Cabin No. 3 - 300 meters The first derivative of temperature continues to deviate positively, and the cumulative integral of methane increases slightly; there is no open flame or personnel in the footage. 0.52 (Long-distance ventilation dead zone, extremely poor convection) Slowly changing time domain deviation / normal / normal (no weight reduction triggered) 82% (High risk, high gain amplification in ventilation dead zones) The fire hazard was determined to be a minor smoldering fire caused by localized insulation aging at the B cable joint in the remote sealing section of compartment 3. Due to its location in a ventilation dead zone, the fire hazard was highly likely to accumulate. Branch A (allow passage) Drive the collaboration between the drainage and electrical engineering departments: remotely disconnect the non-power cables in section 3 and switch fan 1 to smoke exhaust mode. 15-20 Cabin No. 3 - 310 meters A sudden increase in upstream temperature was detected, along with a simultaneous surge in the cumulative humidity; slight smoke particle diffusion was observed in the image. 0.55 (long-distance ventilation dead angle) Continuous temporal deviation / spatial distributed coupling (accompanied by diffusion delay) / cooperative verification passed 96% (Extremely High Risk) Hazard evolution assessment: Smoldering in compartment 3 is escalating, and heat and secondary water mist are spreading downstream to the adjacent 310-meter section with the flow field, posing a risk of fire spread. Branch B (intercept) → After reflection → Branch A (allow) The initial instruction from the large model erroneously triggered the simultaneous closure of the smoke exhaust valve and the fire damper; the gateway intercepted and provided feedback; the large model then generated a compliant instruction through secondary inference: immediately lower the fireproof roller shutters of zones 2 and 3, and activate positive pressure ventilation to block the smoke. 20-25 Cabin No. 3 - 300 meters The cumulative integral value of gas concentration begins to decrease, and the second derivative of temperature shows a negative deceleration; the smoke in the image dissipates. 0.52 (Ventilation dead angle) Recovery trend / spatial correlation removed / normal 45% (Medium risk, in the decline phase) The linkage control has taken effect, the fire compartment has been successfully isolated, the local overheat source has been controlled and cooled down, and the secondary smoke concentration has fallen below the safety threshold. Branch A (allow passage) Maintain positive pressure air supply and fireproof isolation status, and wait for maintenance personnel to manually enter the site for verification. 25-30 Cabin No. 1 - 050 meters The first derivatives of all trend features tend to zero, and the integral below the curve reverts to the baseline interval; the image shows no abnormalities. 1.45 (Strong Convection Zone) Normal / Normal / Normal 10% (Low Risk) The overall environmental field of the utility tunnel has fully returned to the interior of the differential fingerprint baseline envelope, and the alarm has been lifted. Branch A (allow passage) Reset the hardware device and restore the utility tunnel collaborative monitoring system to the normal online inspection mode.

[0131] Experimental Analysis:

[0132] By comparing the joint judgment results of model anomalies, the semantic reports of large-scale white-box causes, and the compliance of security gateway verification control results with actual operation and maintenance labels, the accuracy and robustness of the system in identifying and predicting utility tunnel anomalies are verified. Experiments show that the environmental fingerprint feature matrix generated by trend vectorization encoding can effectively capture hidden, slowly changing, and spatially distributed composite hazards. The differentiated baseline embedded with ventilation performance evaluation coefficients and the visual-data semantic weighted verification successfully intercepted false alarms of environmental noise with high disturbance, strong convection, and complex vehicle vibration on the tunnel. The introduction of the hard-coded physical mechanism verification gateway not only perfectly eliminates the control illusion of the vertical large model and achieves dual-loop reflective evolution convergence, but also ensures the industrial-grade absolute safety of cross-professional equipment collaborative linkage commands, which can effectively improve the intelligent, white-box collaborative management and control capabilities of underground complex utility tunnels throughout their entire life cycle.

[0133] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A pipeline coordination supervision and abnormality identification system for multi-source data large model fusion, characterized in that: Specifically, it includes: The multimodal spatiotemporal unified coding module is used to align the continuous time-series IoT sensor data and monitoring image sequences generated by each monitoring section in the target pipe gallery in terms of time and space dimensions through a sliding window. It also uses a trend vectorization method to calculate the multidimensional change feature vector of each measuring point data within the sliding time window, constructs an environmental fingerprint feature matrix with spatiotemporal correlation semantics, and maps the environmental fingerprint feature matrix into a structured text summary. The dynamic baseline self-learning module is used to continuously learn the multi-dimensional environmental parameter characteristics of different monitoring sections based on historical normal operation data, and construct a dynamically updated normal operation environment fingerprint baseline. The environmental fingerprint baseline is embedded with ventilation performance evaluation coefficients for the differences in physical structure of different monitoring sections, which are used to perform baseline differentiation correction for gas diffusion and temperature and humidity field characteristics under different ventilation conditions. The multidimensional anomaly joint determination module is used to compare the real-time generated environmental fingerprint feature matrix with the corresponding environmental fingerprint baseline, perform triple verification including time-domain drift detection, spatial correlation analysis and visual-data cross-validation, and combine the ventilation performance evaluation coefficient for comprehensive evaluation, and output the anomaly joint determination result with confidence score. The control and coordination module is used to integrate the joint anomaly determination results output by the multi-dimensional anomaly joint determination module, the relevant structured text summary, and the historical operation and maintenance knowledge graph of the utility tunnel into a multimodal semantic context prompt word, which is then input into a preset vertical large model of the utility tunnel. The vertical large model of the utility tunnel performs advanced attribution analysis based on contextual understanding, and outputs a white-box semantic report on the causes of abnormal events and cross-disciplinary equipment collaborative linkage control instructions.

2. The pipeline coordination monitoring and abnormality identification system of multi-source data large model fusion according to claim 1, characterized in that: The multimodal spatiotemporal unified coding module aligns the continuous time-series IoT sensor data and monitoring image sequences generated in each monitoring section within the target utility tunnel in terms of time and space dimensions through a sliding window. Specifically, this includes the following steps: Time dimension alignment: set a global sliding time window with a length of and a sliding step of , intercept a time segment with a length of forward based on the current system clock; for Internet of Things sensors with a sampling frequency higher than the preset standard frequency, use mean smoothing processing for frequency reduction alignment; for Internet of Things sensors with a sampling frequency lower than the preset standard frequency, use linear interpolation algorithm for interpolation completion, to ensure that all sensors have equal distance scattered time sequences with the same dimension within the time window , and the corresponding frames of the monitoring image sequence within the time window are intercepted synchronously; Spatial Dimension Alignment: Based on the building information model of the utility tunnel, a unified three-dimensional spatial coordinate system is established, and the physical installation positions of fixed sensors and dynamic / static monitoring cameras in each monitoring section are pre-mapped into three-dimensional spatial coordinates. Based on the axial distance of the monitoring section, all data sources within the same section are labeled with a unified spatial region tag. This enables the binding of spatial domains.

3. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 2, characterized in that: The multimodal spatiotemporal unified coding module, after aligning the time and spatial dimensions, employs a trend vectorization method to capture the dynamic evolution characteristics of the data, in order to combat the absolute value drift caused by seasonal fluctuations in the sensor environment. Specifically, this includes: calculating the first derivative of the discrete time-series data of each sensor point within the current sliding time window to characterize the parameter change rate, the second derivative to characterize the change acceleration, and the discrete integral to characterize the cumulative effect. The above three calculation results are then dimensionless and concatenated to form the three-dimensional trend feature components of the current sensor point. The three-dimensional trend feature components are used to quantify the dynamic evolution of environmental parameters within the time window and the energy accumulation degree of slowly changing anomalies.

4. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 3, characterized in that: The multimodal spatiotemporal unified coding module, after extracting the trend feature components of a single measurement point, will encode the same monitoring segment... The trend feature components of all measuring points within the tunnel are organized and arranged in a matrix according to spatial topological order, forming the overall environmental fingerprint feature matrix of the tunnel within the current time window, specifically including: Horizontal stitching of IoT features: When M IoT measurement points are deployed within a corresponding segment, the trend feature components of these M measurement points are stitched together. By horizontally concatenating the segments according to their spatial topological order, a temporal feature matrix is ​​constructed. ;in, Represents the trend characteristic components of the Mth IoT measurement point within the segment; matrix The dimension is ; Multimodal feature fusion: extraction time window Apparent change feature vector of internal monitoring image sequence and will Introduced as an additional column of the matrix, along with the time-series feature matrix. To merge; Adaptive weight mapping: In the fusion process, an adaptive dynamic weight mapping function driven by visual semantics is introduced: When the semantic recognition result of the monitoring image sequence outputs an abnormal appearance, the dynamic weight mapping function automatically increases the feature component weight coefficient of the IoT measurement point at the corresponding three-dimensional coordinate point in space, so as to realize cross-modal semantic cross-enhancement between machine vision features and continuous time-series numerical features. Matrix generation: within the current time window Within this matrix, a two-dimensional high-dimensional matrix is ​​formed, integrating temporal fluctuations, spatial topology, and visual appearance features, defined as the environmental fingerprint feature matrix. .

5. The multi-source data large model fusion-based collaborative supervision and anomaly identification system for utility tunnels according to claim 1, characterized in that: The dynamic baseline self-learning module constructs a dynamically updated normal operating environment fingerprint baseline, specifically including the following steps: Collect environmental fingerprint feature matrices of all target utility tunnels that are confirmed to be in normal operating condition within a preset historical period. Unsupervised clustering algorithm is used to perform time-based and seasonal clustering on the collected environmental fingerprint feature matrix to obtain the normal state clusters of the utility tunnel under different spatiotemporal scenarios. For each normal state cluster, an autoencoder neural network is used for reconstruction learning. By compressing and reconstructing the fingerprint of the normal environment, the probability distribution boundaries of the first derivative, second derivative, and cumulative integral value under normal fluctuation states are extracted, and the baseline envelope of the normal operating state is established. The baseline envelope defines the multidimensional permissible fluctuation range of each monitoring parameter under normal conditions.

6. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 5, characterized in that: The dynamic baseline self-learning module quantifies the ventilation differences in each monitoring section by constructing a simplified fluid dynamics model of the pipe gallery's physical structure, specifically including: For each monitoring section Extract its corresponding physical structural parameters, including: the rated air volume of the mechanical exhaust fan in the section. Axial distance of the nearest natural air inlet / make-up air inlet The cross-sectional area A of the cabin and the air resistance coefficient of the section. ; A physical formula was constructed to characterize the air convection and dilution capacity of a monitoring section, and the ventilation performance evaluation coefficient of the monitoring section was calculated. ;in, This is the normalized adjustment coefficient; The calculated ventilation performance evaluation coefficient As a correction gateway, it is embedded into the corresponding environmental fingerprint baseline to perform nonlinear correction on the judgment boundary of different segments and establish an adaptive judgment boundary.

7. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 6, characterized in that: The calculated ventilation performance evaluation coefficient As a correction gateway, it is embedded into the corresponding environmental fingerprint baseline to perform nonlinear correction on the judgment boundary of different segments and establish an adaptive judgment boundary. Specific correction mechanisms include: Correction for cumulative gas concentration limits: for ventilation performance evaluation coefficients For poorly ventilated sections that tend to the lower limit, the cumulative integral boundary of gas concentration in their environmental fingerprint is widened proportionally to tolerate trace amounts of gas accumulated due to poor air circulation and to intercept background false alarms. Temperature rise rate limit correction: for ventilation performance evaluation coefficient For the section adjacent to the wind turbine that is approaching the upper limit, the temperature first and second derivative boundaries in the environmental fingerprint of this section are tightened to sensitively detect abnormal temperature rises. Differential baseline matrix generation: Through the above mechanism, tensor correction operators are used. The unified baseline envelope The specific ventilation performance evaluation coefficient of the section Perform fusion calculations to generate a differentiated environmental fingerprint baseline matrix with specific segment physical structure characteristics. ,in, Spatial identifiers for utility tunnel sections are used to eliminate the interference of heterogeneous spatial environments on the accuracy of anomaly identification.

8. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 1, characterized in that: The multi-dimensional anomaly joint determination module specifically includes the following: Set the current time window Real-time generated environmental fingerprint feature matrix Input the differential environment fingerprint baseline matrix and perform the following triple verification synchronously in parallel or serially: The first layer of verification: time-domain drift detection, used to capture time-domain slowly varying potential hazards caused by minor leaks or localized dampness. Specifically, it calculates the topological distance between the IoT time-series feature portion of the real-time environmental fingerprint feature matrix and the corresponding baseline envelope. If, across multiple consecutive sliding time windows, the topological distance continuously exceeds a preset safety confidence interval, and the first derivative of the environmental parameters exhibits a long-period nonlinear monotonic deviation trend, it is determined that the system has a slowly varying anomaly in the time domain, and a time-domain anomaly evaluation value is output. ; The second layer of verification: spatial correlation analysis, used to capture spatially distributed anomalies coupled across multiple measurement points; Based on the spatial topology of the utility tunnel, the current monitoring section is extracted. and its upstream section and downstream section The real-time environmental fingerprint feature matrix is ​​used to calculate the correlation feature matrix across the spatial span; the determination logic is: current segment The first derivative of temperature increases positively and simultaneously triggers the downstream section. The cumulative humidity integral value increases in the same direction, identifying a multi-physics spatial collaborative change pattern, determining that a multi-point spatial composite anomaly has occurred, and outputting a spatial anomaly evaluation value. ; The third layer of verification: visual-data cross-validation, used to combat environmental noise and intercept potential false alarms; The specific process is as follows: Extract the current time window Feature vectors of apparent changes in monitoring images that are in the same spatiotemporal alignment domain as IoT data The cross-validation logic is as follows: when the temporal drift detection of the first layer of validation or the spatial correlation analysis of the second layer of validation indicates the presence of abnormal temperature, smoke, or gas, and the image appearance change feature vector, after semantic recognition and parsing, shows that there is no personnel intrusion, no open flame appearance, and no smoke particle diffusion pattern in the corresponding target physical area, it is determined that the current data fluctuation is likely due to environmental background noise or single-point sensor drift, triggering the visual weighting operator. ; After completing the above triple verification, the time-domain anomaly evaluation value is calculated using the anomaly comprehensive scoring function. The spatial anomaly evaluation value and the aforementioned visual weighting operator By coupling the parameters, a comprehensive score is calculated, as shown in the following formula: ;in, and These are the preset first and second weight coefficients, respectively; For comprehensive scoring; when the false alarm interception logic of the vision-data cross-validation is triggered, the vision weighting operator... Assigning a preset low value will affect the overall score. It directly reduces the noise level to below the preset safety warning line, thereby enabling intelligent interception of false alarms about environmental noise. Utilizing the ventilation performance evaluation coefficient of the corresponding monitoring section The comprehensive score The final correction is performed, and the joint anomaly determination result with confidence score is output. The specific formula is as follows: ;in, This is a mapping function that is inversely proportional to ventilation performance.

9. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 1, characterized in that: The control and coordination module integrates and constructs multimodal semantic contextual prompts, which are input into a preset vertical large model of the utility tunnel. Specifically, it includes the following: Automatically retrieve heterogeneous information of the currently triggered alarm segment and concatenate it according to a preset semantic template; The heterogeneous information includes quantitative sensing features, spatiotemporal semantic summaries, and historical mechanism background. The quantitative sensing features include the final confidence score, temporal anomaly evaluation value, spatial anomaly evaluation value, and corresponding ventilation performance evaluation coefficient. The spatiotemporal semantic summaries include structured text summaries of environmental feature trends and visual feature appearances. The historical mechanism background includes the types, years of operation, maintenance records, and high-incidence fault modes of various professional pipelines within the alarm section, retrieved from the historical operation and maintenance knowledge graph of the utility tunnel based on the spatial area label of the alarm section. The spliced ​​heterogeneous information elements are embedded into the industrial role setting and reasoning constraint framework to generate a text complex consisting of a role definition segment, a spatiotemporal awareness data segment, a knowledge graph background segment, and a reasoning format constraint segment, which serves as a multimodal semantic context prompt for the vertical large model of the pipe gallery.

10. The multi-source data large model fusion-based pipe gallery collaborative supervision and anomaly identification system according to claim 9, characterized in that: The control coordination module also includes a hard-coded physical mechanism verification gateway for constructing a dual-loop security control mechanism, specifically including: Instruction parsing: Using a fixed lightweight named entity recognition operator, the action subject, target operating equipment, and quantized control amplitude in the text linkage control instructions generated by the vertical large model of the pipe gallery are extracted and mapped to the physical address of the register of the underlying field programmable logic controller (PLC) and the written value, and converted into a standard control code string; The lightweight named entity recognition operator constructs a triplet of operation verb-equipment object-quantization threshold, compares it with a preset industrial equipment mapping table, and absolutely maps the text instruction to the physical address of the corresponding PLC register; Full-scale closing verification: Before the standard control code string is sent to the PLC controller on site, a full-scale closing verification is performed through a physical mechanism verification gateway; the physical mechanism verification gateway has hard-coded absolute regulations for fire safety in pipe corridors, equipment interlocking logic, and hard constraints on physical boundaries, and its text review rules include: Taboo rule verification: Verify whether there are contradictory instructions, including verifying whether the smoke exhaust valve and fire damper are simultaneously closed before the fire is confirmed to be extinguished; Boundary hard constraint verification: Verify whether the generated equipment adjustment exceeds the physical load limit of the corresponding equipment; Different branch controls are executed based on the verification results, specifically: Branch A: Indicates that the verification is successful. When the control code string fully complies with the hard-coded safety procedures, the physical mechanism verification gateway performs the closing and release, and officially sends the standard control code string to the corresponding professional PLC controller, driving the hardware device to perform cross-professional collaborative blocking. Branch B: Represents verification interception and reflection. When the standard control code string violates any safety procedure, the physical mechanism verification gateway triggers the interception and blocking mechanism, and converts the specific reason for the violation and the violated physical rule into reverse correction text feedback, which is then re-inputted into the vertical large model of the pipe gallery as a follow-up prompt. This forces the vertical large model of the pipe gallery to perform secondary instruction reasoning and regenerate the linkage control instruction until it passes the verification of the physical mechanism verification gateway. Branch B incorporates a dynamic step-count blocking and physical safety fallback switching mechanism: the physical mechanism verification gateway monitors in real time the number of reverse correction and inquiry rounds for the same abnormal event; when the vertical large model of the utility tunnel fails to pass the full-scale closing verification after experiencing the preset maximum number of reflection rounds, the physical mechanism verification network closes the closed-loop control authority of the vertical large model of the utility tunnel, instantly activates the cross-professional safety fallback mechanism control strategy hard-coded in the field PLC, and sends a manual takeover request to the central operation and maintenance terminal; the maximum number of reflection rounds is preferably 3 rounds.

Images