Pipe gallery comprehensive monitoring and early warning method

Abstract

The present application relates to underground pipe gallery safety monitoring technical field, more specifically, the present application relates to a kind of pipe gallery comprehensive monitoring early warning method, the method includes, based on the deviation of sensor data distribution in dynamic sliding time window and historical steady-state distribution and multidimensional heterogeneous data matrix, attention weight in cross-modal feature fusion process is inhibited adjustment, generates the comprehensive state representation vector of each topological constraint space unit at current time;The cosine approximation degree of comprehensive state representation vector and each node reference feature vector is calculated, and dynamic adaptive threshold is activated node, and the dynamic adaptive threshold update value of each node is obtained;Based on dynamic adaptive threshold update value, activated node set is screened out;Extract the peak risk node of diffusion intensity superthreshold, generate hierarchical cascading early warning instruction according to node type and the diffusion intensity, the accuracy of pipe gallery comprehensive monitoring early warning is improved.

Description

Technical Field

[0001] This invention relates to the field of underground utility tunnel safety monitoring technology. More specifically, this invention relates to a comprehensive monitoring and early warning method for utility tunnels. Background Technology

[0002] Urban underground utility tunnels are equipped with a variety of heterogeneous sensing devices, including temperature and humidity sensors, hazardous gas concentration detectors, cable temperature monitoring terminals, and electromechanical equipment status acquisition devices. These devices have significantly different initial sampling frequencies, and their sampling times are independent of each other. Furthermore, the reporting formats for installation location information are inconsistent, resulting in highly fragmented data across multiple sources in terms of both time and spatial distribution. With the expansion of utility tunnel construction and the increasing demands for operation and maintenance, higher requirements are being placed on the real-time performance and reliability of early identification of abnormal conditions within the tunnels and cross-regional collaborative early warning systems.

[0003] Based on this, existing monitoring technologies generally employ direct time-series splicing or simple linear interpolation alignment when processing the aforementioned multimodal data, and then independently determine a single physical quantity based on a fixed threshold. For example, some systems forcibly align time-series data such as temperature and gas concentration by timestamps and then input them into a shallow neural network for state classification, or trigger an alarm based on whether the value of a single sensor exceeds a preset threshold.

[0004] However, these methods fail to effectively suppress noise and pseudo-anomalies introduced by asynchronous sampling. The fused state characteristics often exhibit irregular spike oscillations, leading to frequent false alarms and missed alarms. Regarding risk reasoning and linkage control, most existing systems are based on single-point thresholds or isolated logic judgments, lacking structured modeling of the inherent physical spatial constraints of the utility tunnel (such as fire compartments, ventilation direction, and structural slope). This makes it impossible to calculate the physical attenuation and dynamic impact range of risks propagating along the tunnel, resulting in a lack of spatial specificity and temporal coordination in cross-regional linkage. Traditional monitoring systems often employ open-loop architectures, lacking state feedback and closed-loop verification from actuators such as programmable logic controllers to upper-level analysis modules. Detection thresholds cannot be adaptively adjusted online based on actual execution results, leading to insufficient long-term reliability and environmental adaptability.

[0005] In summary, the existing technologies have the following combined defects: multimodal asynchronous noise causes severe distortion in the overall state discrimination; fixed thresholds cannot track environmental baseline drift; and the lack of physical constraints on risk diffusion leads to inaccurate warning areas. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes a comprehensive monitoring and early warning method for utility tunnels, comprising the following steps:

[0007] Along the longitudinal mileage direction of the utility tunnel, the utility tunnel is divided into multiple non-overlapping topological constraint space units; various types of sensors deployed in the utility tunnel are mounted to the corresponding topological constraint space units to generate a physical mounting topology mapping table of the utility tunnel.

[0008] For each topological constraint space unit in the entity mounting topology mapping table, a dynamic sliding time window is constructed at the current moment. The sensor data within the dynamic sliding time window is preprocessed to obtain a multidimensional heterogeneous data matrix. Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window and the multidimensional heterogeneous data matrix, the attention weights in the cross-modal feature fusion process are suppressed and adjusted to generate a comprehensive state representation vector for each topological constraint space unit at the current moment.

[0009] A risk evolution knowledge ontology network with a directed graph structure for the utility tunnel is constructed. The network contains several nodes and directed edges. By calculating the cosine approximation between the comprehensive state representation vector and the baseline feature vector of each node, and combining the dynamic adaptive threshold to activate the nodes, the dynamic adaptive threshold update value of each node is obtained. The set of activated nodes is selected based on the dynamic adaptive threshold update value.

[0010] Using the active node as the diffusion source, and combining the directed edge, the ventilation direction and slope between nodes, and the spatial distance, conditional probability diffusion iteration is performed to obtain the diffusion intensity of the source node after each iteration; peak risk nodes with diffusion intensity exceeding the threshold are extracted, and graded linkage early warning instructions are generated according to the node type and the diffusion intensity.

[0011] Preferably, the preprocessing of the sensor data within the dynamic sliding time window to obtain a multidimensional heterogeneous data matrix includes:

[0012] For sensor data with sampling frequencies lower than the standard characteristic sampling frequency, linear interpolation time-domain expansion is applied. For sensor data with sampling frequencies not lower than the standard characteristic sampling frequency, local Gaussian smoothing filtering and equal-interval downsampling with the standard characteristic frequency as the period are applied sequentially to align all data to the same standard frequency, forming a multidimensional heterogeneous data matrix.

[0013] Preferably, the step of suppressing and adjusting the attention weights in the cross-modal feature fusion process based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window and the multidimensional heterogeneous data matrix, and generating a comprehensive state representation vector for each topological constraint space unit at the current time, includes:

[0014] Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window, the local information entropy deviation distribution factor of the sensor's sampled data at the current moment is obtained.

[0015] Based on a multidimensional heterogeneous data matrix, the adjusted cross-attention weights of any two sensors within a topologically constrained spatial cell at the current time are obtained.

[0016] The preprocessed sampled data of the sensor within the dynamic sliding time window at the current moment constitutes the original temporal feature vector of the sensor; the original temporal feature vector of each sensor is weighted and aggregated using the adjusted cross attention weights to obtain the state representation vector of each sensor in the topological constraint space cell at the current moment.

[0017] The mean of the state representation vectors of all sensors within the topologically constrained spatial cell at the current moment is used as the comprehensive state representation vector of the topologically constrained spatial cell at the current moment.

[0018] Preferably, the step of obtaining the local information entropy deviation distribution factor of the sensor's sampled data at the current moment based on the offset between the sensor data distribution within the dynamic sliding time window and the historical steady-state distribution includes:

[0019] For any sensor, the preprocessed sampled data sequence within the current dynamic sliding time window is discretized using an adaptive histogram binning strategy based on historical data distribution to obtain the probability density distribution of the sensor's sampled data at the current moment; based on the probability density distribution, the information entropy of the sensor's sampled data within the current dynamic sliding time window is obtained.

[0020] The absolute value of the difference between the information entropy of the data sampled by the sensor within the current dynamic sliding time window and the sensor's historical steady-state intrinsic information entropy is used as the local information entropy deviation distribution factor of the sensor's sampled data at the current moment.

[0021] Preferably, the step of obtaining the adjusted cross-attention weights of any two sensors within the topologically constrained spatial cell at the current moment based on a multidimensional heterogeneous data matrix includes:

[0022] The multidimensional heterogeneous data matrix is ​​input into the standard Transformer design algorithm. Using two trainable weight matrices, a linear mapping is performed through two independent fully connected layers to obtain the query tensor and the key tensor of each sensor.

[0023] Based on the query tensor and key tensor of the sensor, obtain the scaled dot product similarity between the query tensor and key tensor of any two sensors.

[0024] Based on the scaled dot product similarity between the query tensor and the key tensor of the sensor, the standard cross-attention weights of any two sensors are obtained.

[0025] By using the local information entropy deviation distribution factor to correct the standard cross-attention weights, the adjusted cross-attention weights of any two sensors at the current time can be obtained.

[0026] Preferably, the step of calculating the cosine approximation between the comprehensive state representation vector and the baseline feature vector of each node, and combining this with a dynamic adaptive threshold to activate the node, to obtain the dynamic adaptive threshold update value for each node, includes:

[0027] Based on the comprehensive state representation vector, obtain the cosine approximation degree between each node and the topological constraint space unit; construct a historical sliding window of cosine approximation degree; obtain the dynamic adaptive threshold update value of each node through the mean and standard deviation of the cosine approximation degree within the historical sliding window of cosine approximation degree.

[0028] Preferably, the step of filtering the set of active nodes based on the dynamically adaptive threshold update value includes:

[0029] For any topological constraint space cell, the cosine approximation degree between the topological constraint space cell and the node is compared with the dynamic adaptive threshold update value of the node: if the cosine approximation degree between the topological constraint space cell and the node is greater than or equal to the dynamic adaptive threshold update value of the node, the node is activated; if the cosine approximation degree between the topological constraint space cell and the node is less than the dynamic adaptive threshold update value of the node, the node is not activated; all activated nodes constitute the set of activated nodes within the current dynamic sliding time window.

[0030] Preferably, the step of using the activated node as the diffusion source point, combined with directed edges, ventilation direction and slope between nodes, and spatial distance, to perform conditional probability diffusion iteration to obtain the diffusion intensity of the source node after each iteration includes:

[0031] The iteration is initiated with the activation node as the diffusion source point. The risk evolution knowledge ontology network is retrieved to obtain the directed evolution weights of the directed edges between the source nodes, and the ventilation direction, wind speed and structural slope data of the topological constraint space unit corresponding to the source node are read simultaneously.

[0032] In each iteration, for each target source node, the comprehensive diffusion increment of the target source node is obtained; based on the comprehensive diffusion increment and the diffusion intensity of the target source node in the previous round, the diffusion intensity of the target source node after the current round is obtained; the above iterative operation is continuously executed until the preset maximum number of iterations is reached or the change in diffusion intensity is detected to be lower than the preset convergence threshold, thereby outputting the diffusion intensity of each source node after each iteration.

[0033] Preferably, obtaining the comprehensive diffusion increment of the target source node includes:

[0034] Traverse all source nodes pointing to the target source node, multiply the current diffusion intensity of each source node by the corresponding directional evolution weight, airflow correction factor, slope correction factor and distance attenuation factor in turn, and then add all the product results to obtain the comprehensive diffusion increment of the target source node.

[0035] Preferably, the step of extracting peak risk nodes whose diffusion intensity exceeds a threshold and generating graded linkage early warning instructions based on node type and diffusion intensity includes:

[0036] After traversing all source nodes after the iteration is completed, the source nodes whose diffusion intensity exceeds the preset diffusion peak threshold are identified as peak risk nodes; sorted by diffusion intensity from high to low, a graded linkage early warning instruction is generated for each peak risk node according to the risk node type and the preset risk classification rules.

[0037] The present invention has the following beneficial effects:

[0038] By using a risk evolution knowledge ontology network, the cosine approximation degree of the comprehensive state representation vector and the node baseline feature vector is calculated, and the adaptive threshold is dynamically updated based on historical steady-state data, replacing the traditional fixed threshold. This avoids the risk of frequent false activations caused by the overall temperature rise during the summer high-temperature period, or the risk of missed detection in the winter low-temperature environment, and significantly improves the robustness of node activation.

[0039] Using the activated node as the diffusion source, and combining the directed edge causal weight, the airflow correction factor determined by the ventilation direction and wind speed, the slope correction factor determined by the structural slope, and the geodesic distance attenuation factor, conditional probability diffusion iteration is performed. This process integrates the inherent physical constraints of the utility tunnel into the risk reasoning process, accurately predicts the propagation intensity and impact range of risks in adjacent spatial units, solves the defect of traditional single-point alarms that cannot predict cross-regional risk transmission, and avoids early warning omissions in downwind diffusion scenarios or excessive alarms in upwind areas.

[0040] Based on diffusion intensity and node type, hierarchical linkage early warning commands are generated, and closed-loop verification is achieved through actuator status feedback: for the cumulative frequency of invalid linkages, the node sensitivity coefficient is dynamically adjusted to raise the activation threshold and reduce false activations; for execution anomalies, backup equipment is automatically switched or the early warning level is raised. This breaks the limitations of traditional open-loop systems, enabling the threshold to be adaptively optimized online based on real feedback from the physical world, significantly improving the system's environmental adaptability and early warning reliability in long-term operation. Attached Figure Description

[0041] Figure 1 This is a flowchart illustrating the steps of a comprehensive monitoring and early warning method for utility tunnels according to an embodiment of the present invention. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0043] Please see Figure 1 The diagram illustrates a flowchart of a comprehensive monitoring and early warning method for utility tunnels according to an embodiment of the present invention. The method includes the following steps:

[0044] S001. Along the longitudinal mileage direction of the utility tunnel, divide the utility tunnel into multiple non-overlapping topological constraint space units; deploy various types of sensors in the utility tunnel and mount them to the corresponding topological constraint space units to generate a physical mounting topology mapping table of the utility tunnel.

[0045] It should be noted that the various sensing nodes deployed within the utility tunnel, such as temperature and humidity sensors, hazardous gas detectors, video acquisition equipment, and electromechanical equipment status monitoring terminals, are highly overlapping in physical space. However, their installation location information reporting mechanisms are inconsistent, and some older sensors cannot report precise spatial coordinates. If the potentially missing coordinates reported independently by each sensor are used directly as the positioning basis, a large number of fragmented data blocks with inconsistent spatial locations will be generated, making it impossible to correlate and analyze subsequent multimodal data under unified spatial constraints. Therefore, the first step should be to address the standardization and discretization of the physical space of the utility tunnel and the establishment of the physical mounting relationship between sensing devices and discrete spatial units.

[0046] Since the utility tunnel has been designed and constructed in accordance with fire safety regulations, ventilation system zoning, and waterproof structure, it has been equipped with physical separation structures such as fire-resistant isolation belts, independent ventilation sections, and waterproof gates. These structures constitute the natural physical boundaries inside the utility tunnel. Therefore, the above-mentioned physical separation structures should be regarded as hard constraints for spatial discretization.

[0047] Specifically, along the longitudinal mileage direction of the utility tunnel, the continuous tunnel segment between two adjacent physical separation structures is taken as a spatial unit, and the entire physical corridor of the utility tunnel is divided into multiple non-overlapping topological constraint spatial units.

[0048] After the segmentation is completed, each topological constraint space cell is assigned a unique cell identifier, and the starting and ending mileage coordinates of each topological constraint space are recorded to form a spatial cell boundary dataset.

[0049] The units for the starting and ending mileage coordinates are meters. This discretization method ensures a strict correspondence between the spatial units and the actual physical structure of the utility tunnel, and makes adjacent spatial units have a clear topological adjacency relationship.

[0050] It should be noted that after spatial discretization is completed, the various types of deployed sensing devices need to be mapped to the corresponding topological constraint space units. Since some sensors can report their installation location coordinates normally, while others cannot report them due to hardware limitations, it is necessary to use a fallback extraction method based on environmental sensing information to establish an entity mounting topology mapping table.

[0051] Specifically, for sensors capable of reporting location coordinates, the coordinates reported by the sensors are compared with the established spatial cell boundary dataset to determine spatial inclusion.

[0052] If the location coordinates reported by the sensor fall within the starting mileage and ending mileage interval of a certain topological constraint space cell, then the device identifier of the sensor is attached to the entity list of the topological constraint space cell.

[0053] For sensors that have not reported their location coordinates or whose location coordinates are significantly outside the tunnel boundary, a fallback strategy is implemented. The geometric center coordinates of the compartment where the sensor is located are inherited as proxy location coordinates, and then the sensor is mounted to the corresponding topological constraint space unit according to the same spatial inclusion rule.

[0054] The geometric center coordinates of the compartment are provided by the tunnel design drawings or preliminary survey data, in meters, and are pre-stored in the configuration table during the system initialization phase. The final generated entity mounting topology mapping table contains three fields: the identifier of the topology constraint space unit, the physical space coordinate boundary of the topology constraint space unit, and a list of identifiers of all entity equipment mounted in the topology constraint space unit.

[0055] At this point, the entity mounting topology mapping table of the utility tunnel is obtained.

[0056] S002. For each topological constraint space unit in the entity mounting topology mapping table, a dynamic sliding time window is constructed at the current moment. The sensor data within the dynamic sliding time window is preprocessed to obtain a multidimensional heterogeneous data matrix. Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window and the multidimensional heterogeneous data matrix, the attention weights in the cross-modal feature fusion process are suppressed and adjusted to generate a comprehensive state representation vector for each topological constraint space unit at the current moment.

[0057] It should be noted that for each topological constraint space cell in the physical mounting topology mapping table of the utility tunnel, multimodal heterogeneous data generated in real-time by various sensors such as temperature and humidity, harmful gas concentration, and electromechanical status mounted within the topological constraint space cell are extracted. Due to differences in hardware sampling mechanisms and communication protocols, these sensors have significantly different original sampling frequencies, and the sampling times of each sensor are independent of each other. This results in the data collected in the same time period being distributed asynchronously and at unequal intervals on the time axis. This asynchronicity directly manifests as irregular spike oscillations in the comprehensive state representation vector. Therefore, it is necessary to first perform time-series alignment and frequency unification of the multimodal heterogeneous data through a dynamic sliding time window.

[0058] Specifically, a time parameter Tw is preset, and a dynamic sliding time window with the current time t as the right boundary and a length of Tw is constructed.

[0059] The window length Tw can be set according to actual needs, and it needs to meet the following conditions: greater than the maximum sampling interval of the sensor and less than the time resolution required for risk response. This embodiment of the invention does not impose too many restrictions here, for example, it can be 60 seconds.

[0060] For the raw data sequence collected by any sensor mounted within a topologically constrained space cell, different sampling processes are applied based on the relationship between the sensor's sampling frequency and the preset standard feature sampling frequency:

[0061] If the sensor's sampling frequency is lower than the preset standard feature sampling frequency, the original data sequence collected by the sensor will be extended in the time domain using a linear interpolation algorithm within a dynamic sliding time window; if the sensor's sampling frequency is greater than or equal to the preset standard feature sampling frequency, the original data sequence collected by the sensor will be subjected to local Gaussian smoothing filtering and sampling points will be extracted at equal intervals with the standard feature sampling frequency as the downsampling period.

[0062] After the above processing, the original data sequences collected by all sensors within the dynamic sliding time window are projected onto the same standard feature sampling frequency surface, resulting in the preprocessed sampled data sequence of each sensor within the current dynamic sliding time window. The preprocessed sampled data sequences of all sensors within the current dynamic sliding time window are formed into a multidimensional heterogeneous data matrix that is strictly aligned on the time axis, denoted as the multidimensional heterogeneous data matrix of each topological constraint space unit.

[0063] The linear interpolation algorithm and local Gaussian smoothing filtering are existing technologies, and will not be elaborated on in this embodiment. The standard feature sampling frequency can be set according to actual needs, and must meet the following conditions: not lower than the minimum time resolution required by the system and not exceeding the computing resource carrying capacity. This embodiment of the invention does not impose too many restrictions, for example, 1 Hz. The standard deviation of the window width of the Gaussian smoothing kernel can be set according to actual needs, and must meet the following conditions: greater than the hardware noise correlation time and less than the effective signal change time scale. This embodiment of the invention does not impose too many restrictions, for example, 0.5 seconds.

[0064] For any sensor, the preprocessed sampled data sequence within the current dynamic sliding time window is discretized using an adaptive histogram binning strategy based on historical data distribution to obtain the probability density distribution of the sensor's sampled data at the current moment; based on the probability density distribution, the information entropy of the sensor's sampled data within the current dynamic sliding time window is obtained.

[0065] In the system initialization phase, the distribution range of the sampling data of each sensor is statistically analyzed, and the boundaries of K bins are determined by equal frequency quantiles. The specific number of bins K can be set according to actual needs, and the conditions must be met: ensuring the stability of the information entropy estimation and sufficiently reflecting the distribution shape. This embodiment of the invention does not impose too many restrictions here, for example, it can be 20, dimensionless. In the early stage of system operation, a period of normal working condition historical data with a duration of not less than 3600 seconds is taken, and the historical steady-state intrinsic information entropy of each sensor is calculated according to the same binning strategy.

[0066] The absolute value of the difference between the information entropy of the data sampled by the sensor within the current dynamic sliding time window and the sensor's historical steady-state intrinsic information entropy is used as the local information entropy deviation distribution factor of the sensor's sampled data at the current moment.

[0067] The specific formula is as follows:

[0068]

[0069] In the formula, This represents the local information entropy deviation distribution factor of the sensor's sampled data at the current moment; The information entropy represents the data sampled by the sensor within the current dynamic sliding time window; This represents the historical steady-state intrinsic information entropy of the sensor.

[0070] Among them, the local information entropy deviation distribution factor is used to quantify the structural shift of the data probability distribution within the current dynamic sliding time window relative to the historical steady-state distribution; when the data distribution within the window deviates significantly from the historical steady-state pattern due to abnormal spikes, drift, or changes in the actual physical state, The offset occurred, directly leading to Increase.

[0071] It should be noted that the high-frequency quantization spikes and local physical measurement anomalies introduced by asynchronous sampling can cause the information entropy of the sampled data from the relevant sensors to deviate significantly from the historical steady-state benchmark within the dynamic sliding time window. The magnitude of the local information entropy deviation distribution factor directly reflects the degree of unstructured shift in the probability distribution of the sampled data. The larger the value, the heavier the noise-dominated component in the current channel data. If these highly biased sampled data are fed into cross-modal cross-attention calculation without suppression, they will excessively dominate the attention weight allocation, causing the fused integrated state representation vector to exhibit large and irregular oscillations.

[0072] Based on this, it is necessary to utilize the acquired local information entropy deviation distribution factor to apply an exponential decay penalty to the response intensity of high-deviation sensors during the cross-attention calculation process, thereby generating a low-noise comprehensive state representation vector.

[0073] Specifically, for any topologically constrained spatial cell, the multidimensional heterogeneous data matrix of the topologically constrained spatial cell is denoted as... ,in This represents the total number of sensors mounted within a topologically constrained space cell. This indicates the number of aligned time sampling points within the current dynamic sliding time window;

[0074] Multidimensional heterogeneous data matrix The input to the standard Transformer design algorithm (Transformer, a deep learning model based on self-attention mechanism) utilizes two trainable weight matrices and performs linear mapping through two independent fully connected layers to obtain the query tensor and key tensor of each sensor.

[0075] Here, two trainable weight matrices are defined as follows: and Mapping Dimension The empirical value of 64 is used as the preset parameter, but it can also be adjusted by the implementer based on computing resources and feature capacity.

[0076] Based on the query tensor and key tensor of the sensor, obtain the scaled dot product similarity between the query tensor and key tensor of any two sensors.

[0077] Based on the scaled dot product similarity between the query tensor and the key tensor of the sensor, the standard cross-attention weights of any two sensors are obtained.

[0078] By using the local information entropy deviation distribution factor to correct the standard cross-attention weights, the adjusted cross-attention weights of any two sensors at the current time can be obtained.

[0079] The specific formula is as follows:

[0080]

[0081] In the formula, This represents the adjusted cross-attention weights between the i-th and j-th sensors; This represents the standard cross-attention weight between the i-th sensor and the j-th sensor; The local information entropy deviation distribution factor of the j-th sensor is represented by . Represents an exponential function with the natural constant as its base; express The attenuation intensity coefficient can be set according to actual needs, and it needs to meet the following conditions: effectively suppressing abnormal channels without destroying the weight stability of normal channels. This embodiment of the invention does not impose too many restrictions here, for example, it can be set to 0.5.

[0082] The Transformer standard design algorithm is existing technology and will not be elaborated upon in this embodiment. When the local information entropy deviation distribution factor approaches zero, the exponential decay term approaches 1, and the cross-attention weight of the sensor remains almost unchanged. When the local information entropy deviation distribution factor increases significantly due to non-steady-state noise, this term decreases rapidly, forcibly compressing the sensor's contribution in cross-modal aggregation. The penalty is applied independently to each key channel in the form of column scaling, and the suppression amounts of different channels do not interfere with each other. After column scaling, the sum of the elements in each row of the cross-attention weight matrix is ​​no longer equal to 1, but the change in the row sum in subsequent weighted aggregation is only equivalent to scaling the overall query response amplitude and does not change the relative contribution ratio between the information of each sensor.

[0083] The preprocessed sampled data within the dynamic sliding time window of the sensor at the current moment constitutes the sensor's original temporal feature vector; the original temporal feature vector of each sensor is weighted and aggregated using adjusted cross-attention weights to obtain the state representation vector of each sensor within the topologically constrained spatial cell at the current moment; the specific formula is as follows:

[0084]

[0085] In the formula, This represents the state representation vector of the j-th sensor at the current moment; This represents the number of all sensors within a topologically constrained spatial cell. This represents the adjusted cross-attention weights between the i-th and j-th sensors; This represents the original time-series feature vector of the j-th sensor.

[0086] The mean of the state representation vectors of all sensors within the topologically constrained spatial cell at the current moment is used as the comprehensive state representation vector of the topologically constrained spatial cell at the current moment.

[0087] Thus, the comprehensive state representation vector of each topological constraint space element at the current moment is obtained.

[0088] S003. Construct a risk evolution knowledge ontology network with a directed graph structure for the utility tunnel. The network includes several nodes and directed edges. By calculating the cosine approximation degree between the comprehensive state representation vector and the baseline feature vector of each node, and combining the dynamic adaptive threshold to activate the nodes, obtain the dynamic adaptive threshold update value of each node. Filter out the set of activated nodes based on the dynamic adaptive threshold update value.

[0089] It should be noted that the comprehensive state representation vector is obtained by cross-modal fusion and information entropy deviation penalty adjustment of multimodal heterogeneous data, and its value has low noise and high stability in the feature space. However, the comprehensive state representation vector itself is only a fusion of the states of multiple sensors within the topologically constrained spatial unit at the current moment. It does not contain the causal evolution logic between risk types, nor can it express the propagation direction and intensity of risk between different topologically constrained spatial units. If an isolated threshold judgment is directly applied to this vector, the system will degenerate into a single-point anomaly detector. When multiple physical quantities deviate slightly but causally related in the early stage of composite risk, it will be unable to aggregate and reason about the weak signals scattered in each modality according to the risk evolution path, leading to missed warnings. At the same time, the topologically constrained spatial units at different locations in the utility tunnel have significant differences in physical structure, equipment type, and risk evolution path; therefore, a structured knowledge network is needed to carry the differentiated risk evolution logic.

[0090] Based on this, it is necessary to construct a directed graph structured risk evolution knowledge ontology network, calculate the cosine approximation degree of the comprehensive state representation vector and network nodes, and implement node activation by combining the dynamic adaptive threshold generated based on historical steady-state characteristics, so as to provide rigid source points for subsequent conditional probability diffusion.

[0091] Specifically, a directed graph structure risk evolution knowledge ontology network is constructed. The network node set contains two types of nodes: one type is basic risk factor nodes, which correspond to the types of physical state anomalies in the utility tunnel that can be directly or indirectly observed by sensors, such as abnormal local temperature rise, increased combustible gas concentration gradient, deterioration of cable joint insulation resistance, and stagnation of airflow in ventilation sections; the other type is composite risk event nodes, which correspond to higher-level accident types that may be caused by the transmission of multiple basic risk factors in the physical causal chain, such as electrical fires, diffusion of toxic fumes, and local compartment hypoxia.

[0092] The directed edge set of a network represents the causal evolution relationship from cause nodes to result nodes. For example, there is a directed edge from the node of deteriorating insulation resistance of a cable joint to the node of electrical fire, indicating that the former will induce the latter under certain conditions; there is a directed edge from the node of abnormal local temperature rise to the node of airflow stagnation in the ventilation section, indicating that the abnormal temperature may cause the ventilation equipment to trip in a cascading manner, thus causing airflow stagnation. These directed edges together constitute the causal connectivity topology of the risk evolution knowledge ontology network.

[0093] In the initial construction phase of the risk evolution knowledge ontology network, the edge weights of the adjacency matrix are pre-defined by expert rules: It is an asymmetric matrix with elements Indicates from node To the node The directed evolution weights, if the node To the node If there is a causal evolutionary relationship confirmed by experts, then Initially set as an interval The expert experience values ​​within the system are specifically determined by the fire protection specifications for utility tunnels, risk assessment reports, and historical accident cases. For example, the boundary weight for the progression from cable joint insulation resistance deterioration to electrical fire is 0.7, and the boundary weight for the progression from abnormal local temperature rise to airflow stagnation in ventilation sections is 0.4. These values ​​are dimensionless. If no causal relationship exists, then... .

[0094] The adjacency matrix is ​​loaded as a cold start scheme during system initialization, and then the edge weights are updated online based on the subsequent closed-loop verification results: when the causal path represented by a directed edge is verified as having a high probability of occurrence in the actual warning, the edge weight is increased by a small amount, and vice versa. The increment and decrement step size can be set according to actual needs, and the following conditions must be met: ensure learning stability and do not over-respond to a single event. This embodiment of the invention does not impose too many restrictions here, for example, it is taken as 0.05, which is dimensionless.

[0095] Each node in the risk evolution knowledge ontology network pre-stores a corresponding baseline feature vector, with dimensions consistent with the comprehensive state representation vector, for cosine approximation calculation. The baseline feature vector is composed of typical multimodal fusion feature templates corresponding to the risk type represented by the node in the historical operation data or simulation data of the utility tunnel. During the system initialization phase, for each type of risk factor or risk event, the comprehensive state representation vectors of at least 20 typical samples in the historical case library are selected, and the dimensional mean of these vectors is calculated to obtain the baseline feature vector corresponding to the node, which is then stored in the knowledge base. Optionally, after the system has run for a preset period, the baseline feature vectors of each node are recalculated using the accumulated confirmed real risk event samples to achieve long-term maintenance of the knowledge base.

[0096] Based on the comprehensive state representation vector, the formula for calculating the cosine approximation degree between each node and the topological constraint space element is as follows:

[0097]

[0098] In the formula, This represents the comprehensive state representation vector of the topological constraint space element at the current moment; Represents a node The baseline feature vector; For vector dot product, and These are the Euclidean norms of the comprehensive state representation vector and the baseline eigenvector, respectively. Represents a node Cosine approximation of the topologically constrained space cell, range of values The closer a value is to 1, the more likely it is to represent the overall state and nodes of the topologically constrained space cell. The risk types they represent tend to be more consistent in their characteristic directions.

[0099] It should be noted that a dynamic adaptive threshold, generated based on historical steady-state characteristics, is introduced to determine whether a node is activated. Since the operating status of the utility tunnel experiences slow baseline drift due to seasonal and diurnal load fluctuations and equipment aging, a fixed threshold may frequently result in false activations during the high-temperature summer months due to the overall upward shift of the temperature baseline, and may miss detections during the cold winter months. Therefore, the generation and online adjustment of the dynamic adaptive threshold directly affect the accuracy and robustness of node activation.

[0100] In the initial stage, for each node, a comprehensive state representation vector sample set corresponding to the risk type of the node is extracted from its corresponding historical steady-state dataset. The cosine similarity statistical distribution between the sample set and the baseline feature vector of the node is calculated, and the upper bound of the 95% confidence interval of the distribution is taken as the initial dynamic adaptive threshold of the node.

[0101] Taking a node with a local temperature anomaly as an example, if 500 cosine similarity values ​​of the node are obtained from historical steady-state operation data, and these values ​​follow an approximately normal distribution with a mean of 0.3 and a standard deviation of 0.1, then the upper bound of the 95% confidence interval is approximately 0.4645. At this time, the initial dynamic adaptive threshold of the node is initialized to 0.46. The specific confidence level can be set according to actual needs, and this embodiment of the invention does not impose too many restrictions here.

[0102] During continuous system operation, the specific process of adjusting the dynamic adaptive threshold through online self-learning is as follows:

[0103] Construct a [structure] with the current time as its right boundary and a length of [length]. The cosine approximation history sliding window is used to obtain the cosine approximation between the comprehensive state representation vector of each topological constraint space unit and the node. Regardless of whether the node is activated, the cosine approximation value is incorporated into the historical sliding distribution statistics of the node.

[0104] in, Specific settings can be configured according to actual needs, and the following conditions must be met: ensuring statistical stability and tracking the rate of environmental change. This embodiment of the invention does not impose excessive limitations here; for example, 1000 sample points can be used. Whenever the number of samples within the historical sliding window of the cosine approximation reaches... Then, a rolling update method is used to discard the oldest sample and add the latest sample.

[0105] By analyzing the distribution of cosine approximation within the historical sliding window of cosine approximation, the dynamic adaptive threshold update value of each node is obtained.

[0106] The specific formula is as follows:

[0107]

[0108] In the formula, Represents a node The dynamic adaptive threshold update value; This represents all topologically constrained spatial cells and nodes within the history sliding window of the cosine approximation. The mean of the cosine approximation values ​​between them; This represents all topologically constrained spatial cells and nodes within the history sliding window of the cosine approximation. The standard deviation of the cosine approximation values ​​between them; This is the sensitivity coefficient, which can be set according to actual needs. It needs to meet the following conditions: balance the sensitivity to real anomalies and the ability to suppress noise. This embodiment of the invention does not impose too many restrictions here, and the initial value is 2.0.

[0109] The update process can be performed each time a new cosine approximation is obtained, allowing the threshold to migrate smoothly with the slow, normal drift of the environment and device state.

[0110] For any topological constraint space cell, connect the topological constraint space cell with the node. cosine approximation , and nodes Dynamic adaptive threshold update value Comparison:

[0111] like Then the node Activated, its activation strength is recorded. ;otherwise, All activated nodes and their activation strengths constitute the set of activated nodes within the current dynamic sliding time window.

[0112] At this point, the set of active nodes within the current dynamic sliding time window is obtained.

[0113] S004. Using the active node as the diffusion source point, and combining the directed edge, the ventilation direction and slope between nodes, and the spatial distance, perform conditional probability diffusion iteration to obtain the diffusion intensity of the source node after each iteration; extract the peak risk nodes whose diffusion intensity exceeds the threshold, and generate graded linkage early warning instructions according to the node type and the diffusion intensity.

[0114] It should be noted that several active risk factor nodes correspond to abnormal physical state trends confirmed by multimodal fusion features occurring within different topologically constrained spatial units. However, as a continuous, confined physical space, the evolution of risks within the utility tunnel is naturally constrained by physical constraints such as airflow along the tunnel direction, structural slope, and geometric path distance. For example, abnormal heating of cable joints may lead to a gradual increase in the temperature field of adjacent sections, and the leakage of toxic and harmful gases will form a concentration diffusion zone along the ventilation airflow direction. These processes are not instantaneous and synchronous across the entire space but rather attenuate and propagate along the topological path according to physical laws. If a fixed-level alarm is generated only for a single isolated point after the node is activated, the system will be unable to predict the propagation intensity and impact frontier of the risk in adjacent spatial units. In scenarios where toxic fumes spread downwind or fire spreads along the cable tunnel, this could easily lead to missed warnings for compartments about to be threatened or excessive alarms for compartments far from the risk source.

[0115] Based on this, it is necessary to use the activated nodes as the diffusion source points, and perform conditional probability diffusion iteration on the constructed risk evolution knowledge ontology network, which is jointly constrained by the physical correction factor and the geodesic distance attenuation coefficient, to extract peak risk nodes and generate early warning instructions with spatial targeting, intensity classification and linkage direction.

[0116] Specifically, all nodes with activation strength greater than 0 are formed into a source set for diffusion; for each source node, its initial diffusion strength is directly taken as the activation strength; for other non-source nodes, the initial diffusion strength is set to 0.

[0117] Among them, the adjacency matrix of the risk evolution knowledge ontology network is obtained from the nodes. To the node Directed edge evolution weights This represents the cause node under ideal causal logic that ignores physical space decay. For the result node The conditional probability propagation strength is determined, and the adjacency matrix is ​​updated online during system operation based on the closed-loop verification results, using the latest weights at the current moment.

[0118] It should be noted that, since the actual risk transmission intensity within the utility tunnel is significantly modulated by the physical transmission conditions along the tunnel direction, the physical correction factor between any two topologically constrained spatial units can be obtained from the utility tunnel design drawings and ventilation system operation data. The physical correction factor represents the physical enhancement or inhibition factor of risk transmission between two node-corresponding topologically constrained spatial units, which is determined by the ventilation flow direction and wind speed between the two topologically constrained spatial units, as well as the longitudinal height difference and the density characteristics of flammable or toxic gases between the two topologically constrained spatial units. Therefore, the physical correction factor can be composed of airflow correction factor and slope correction factor.

[0119] For the airflow correction factor, if the node Corresponding topological constraint space cells and nodes The corresponding topologically constrained spatial cells are located in the same ventilation section and the airflow direction is from the node Blow towards the node If the airflow velocity between two topologically constrained spatial cells exceeds a preset low wind speed threshold, then the airflow correction factor is an enhancement coefficient greater than 1, and the node is obtained. Corresponding topological constraint space cells and nodes The formula for calculating the airflow correction factor between the corresponding topologically constrained space cells is as follows:

[0120]

[0121] In the formula, Represents a node Corresponding topological constraint space cells and nodes The corresponding airflow correction factor between topologically constrained space cells; This indicates the node that the ventilation monitoring system reports in real time. The corresponding topological constraint space cell blows towards the node The airflow velocity between the corresponding topologically constrained spatial units; Indicates the reference airflow wind speed; Indicates the airflow amplification gain parameter;

[0122] The reference airflow velocity can be set according to actual needs. This embodiment of the invention does not impose many restrictions here, for example, it can be set to 1.0 m / s. The airflow amplification gain parameter can be set according to actual needs, and it needs to meet the condition that it matches the cross-sectional size of the pipe gallery and the conventional ventilation volume. This embodiment of the invention does not impose many restrictions here, for example, it can be set to 0.6.

[0123] If node Corresponding topological constraint space cells and nodes The airflow directions between the corresponding topologically constrained spatial cells are opposite, i.e., by the nodes The corresponding topological constraint space cell blows towards the node The corresponding topologically constrained spatial unit implies that the risk headwind transmission is suppressed; therefore, the node Corresponding topological constraint space cells and nodes The formula for calculating the airflow correction factor between the corresponding topologically constrained space cells is as follows:

[0124]

[0125] In the formula, Represents a node Corresponding topological constraint space cells and nodes The corresponding airflow correction factor between topologically constrained space cells; This indicates the node that the ventilation monitoring system reports in real time. The corresponding topological constraint space cell blows towards the node The airflow velocity between the corresponding topologically constrained spatial units; Indicates the reference airflow wind speed; Indicates the airflow amplification gain parameter; This represents an exponential function with the natural constant as its base.

[0126] If node Corresponding topological constraint space cells and nodes There are no nodes with obvious directional airflow or ventilation monitoring system reporting in real time between the corresponding topologically constrained spatial units. The corresponding topological constraint space cell blows towards the node If the airflow speed between the corresponding topological constraint space cells is less than the preset low wind speed threshold, then the node... Corresponding topological constraint space cells and nodes The airflow correction factor between the corresponding topologically constrained space cells is 1.

[0127] The low wind speed threshold can be set according to actual needs and must meet the following conditions: the lowest wind speed that can be sensed for directional airflow in the pipe gallery. This embodiment of the invention does not impose too many restrictions here, for example, it can be 0.3 meters per second.

[0128] For the slope correction factor, if the node Corresponding topological constraint space cells and nodes The airflow direction between the corresponding topologically constrained space cells is along the nodes. To node When the airflow direction is upward, gas diffusion in the uphill direction is hindered by the terrain, so the slope correction factor is a suppression coefficient less than 1; if the airflow direction is along the node... To node When the slope is downward, the slope correction factor is an enhancement coefficient greater than 1; therefore, the node Corresponding topological constraint space cells and nodes The expression for calculating the slope correction factor between the corresponding topologically constrained spatial cells is as follows:

[0129]

[0130] In the formula, Represents a node Corresponding topological constraint space cells and nodes The slope correction factor between corresponding topological constraint space units; Represents a node Corresponding topological constraint space cell to node The slope angle between the corresponding topological constraint space cells; This represents the slope sensitivity coefficient.

[0131] Where the topological constraint space units corresponding to the two nodes do not have direct physical adjacency or the risk transmission is not mainly gaseous diffusion, the slope correction factor is 1; the slope sensitivity coefficient can be set according to actual needs, and needs to meet the condition of matching the gradient correlation with the actual gas monitoring data in the pipe gallery. This embodiment of the invention does not impose too many restrictions here, for example, it is taken as 1.5.

[0132] Nodes are obtained based on airflow correction factors and slope correction factors. Corresponding topological constraint space cells and nodes The specific formula for the physical correction factor between the corresponding topological constraint space elements is as follows:

[0133]

[0134] In the formula, Represents a node Corresponding topological constraint space cells and nodes The corresponding physical correction factor between topological constraint space units; Represents a node Corresponding topological constraint space cells and nodes The slope correction factor between corresponding topological constraint space units; Represents a node Corresponding topological constraint space cells and nodes The airflow correction factor between the corresponding topologically constrained space cells.

[0135] It should be noted that, as risks propagate along the utility tunnel space, in addition to the modulation of causal intensity and physical directionality, they also exhibit inherent attenuation with the increase of propagation geometric distance; therefore, the geodesic distance between the geometric centers of two topologically constrained spatial units can be used as a measure of spatial propagation distance.

[0136] Specifically, set nodes The geometric center coordinates of the corresponding topologically constrained space cell are: ,node The geometric center coordinates of the corresponding topologically constrained space cell are: Then the geodesic distance between the two is the sum of the lengths of the polyline paths of all topologically constrained spatial elements along the longitudinal axis of the utility tunnel; that is, if the node With nodes If the corresponding topological constraint space cells are adjacent, the geodesic distance between them is directly equal to the distance between their geometric center coordinates. If there are other intermediate topological constraint space cells between them, the geodesic distance is the sum of the center distances of all the constraint space cells along the shortest topological path. Then the node... Corresponding topological constraint space cells and nodes The expression for calculating the distance attenuation factor between the corresponding topologically constrained spatial cells is as follows:

[0137]

[0138] In the formula, Represents a node Corresponding topological constraint space cells and nodes The distance attenuation factor between the corresponding topological constraint space units; Represents the physical space conduction attenuation coefficient; Represents a node Corresponding topological constraint space cells and nodes The geodesic distance between corresponding topological constraint space cells; This represents an exponential function with the natural constant as its base.

[0139] The physical space conduction attenuation coefficient can be set according to actual needs, and it needs to meet the condition that it is consistent with the spatial attenuation rate of sensor response intensity in historical risk events in the utility tunnel. This embodiment of the invention does not impose too many restrictions here, for example, it can be taken as 0.02 per meter. The unit is meters.

[0140] It should be noted that during the conditional probability diffusion iteration process, which is jointly constrained by the physical correction factor and the geodesic distance attenuation coefficient, on the risk evolution knowledge ontology network, in the... During step-by-step diffusion, the source node in the network The received diffusion intensity increment comes from all pointing nodes. source node The diffusion intensity in the previous step needs to be corrected by a combination of causal weighting, physical correction, and distance attenuation.

[0141] The expression for calculating the diffusion intensity of source node j after the (t+1)th iteration is:

[0142]

[0143] In the formula, This represents the diffusion intensity of source node j after the (t+1)th iteration; This represents the diffusion intensity of source node j after the t-th iteration; Indicates from the source node To the source node The directed evolution weights; Represents the source node Corresponding topological constraint space cell and source node The distance attenuation factor between the corresponding topological constraint space units; Represents the source node Corresponding topological constraint space cell and source node The corresponding physical correction factor between topological constraint space units; This represents the set of all source nodes pointing to source node j within the current dynamic sliding time window. This represents the diffusion retention coefficient.

[0144] Wherein, when t=0, the initial diffusion intensity of the source node is directly taken as its activation intensity; the diffusion retention coefficient can be set according to actual needs, and needs to meet the condition: control the smoothness of diffusion iteration. In this embodiment of the invention, no restrictions are imposed, for example, it is taken as 0.7; the above-mentioned preset total diffusion steps can be set according to actual needs, and needs to meet the condition: the maximum propagation depth of the coverage network and no excessive iteration leading to numerical saturation. In this embodiment of the invention, no restrictions are imposed, for example, it is taken as 5.

[0145] After each iteration, determine whether the maximum change in diffusion intensity of each source node in the network is lower than the preset convergence threshold. If so, terminate the iteration early; otherwise, continue execution until the preset total number of diffusion steps is reached.

[0146] The convergence threshold can be set according to actual needs, and it needs to meet the condition that it is less than the minimum perceptible resolution of the system for changes in diffusion intensity. This embodiment of the invention does not impose too many restrictions here, for example, it can be 0.001.

[0147] After the iteration is completed, all source nodes are traversed, and the source node whose diffusion intensity exceeds the preset diffusion peak threshold is taken as the peak risk node.

[0148] The diffusion peak threshold can be set according to actual needs, and it needs to meet the condition of effectively distinguishing significant risks from background noise. This embodiment of the invention does not impose too many restrictions here, for example, it can be set to 0.3.

[0149] All peak risk nodes are sorted from high to low according to their diffusion intensity values, and corresponding graded linkage early warning instructions are generated for each peak risk node based on its node type and preset risk classification rules.

[0150] The node types include basic risk factor nodes and composite risk event nodes. Composite risk event nodes have higher warning levels than general basic risk factor nodes due to the severity of their potential consequences. The matching relationship between warning levels and linkage operation sequences is pre-stored in the system linkage strategy library in the form of a static configuration table, which is compiled by safety engineers based on emergency plans.

[0151] It should be noted that traditional open-loop systems cannot confirm whether the actuator has actually acted after the command is issued, and the threshold cannot be adaptively adjusted according to the execution result, resulting in insufficient long-term operational reliability and environmental adaptability. Therefore, it is necessary to introduce actuator status feedback to perform closed-loop verification of the linkage execution result and adjust the dynamic adaptive threshold through online self-learning.

[0152] Specifically, the system extracts instruction information line by line from the instruction set obtained in the above steps. Each instruction contains the identifier of the peak risk node it affects, the identifier of the corresponding utility tunnel topology constraint space unit, the diffusion intensity value, the instruction warning level, and the linkage operation sequence number. Based on the linkage operation sequence number, the system retrieves the corresponding specific operation sequence from the preset linkage strategy library. This sequence includes the start / stop or opening adjustment actions of actuators such as exhaust fans, fire dampers, emergency lighting, and drainage pumps in the specified space unit.

[0153] After retrieval, the system sends the operation sequence frame by frame to the programmable logic controller (PLC) deployed in the local control cabinet of the corresponding space unit via industrial Ethernet using the Modbus TCP protocol. Upon receiving the instruction frame, the PLC immediately drives the mounted actuator to operate and continuously collects the actual execution status signals (such as valve open signal, motor running current value, and air valve closed signal) returned by the limit switches, current transformers, or status contacts of each actuator.

[0154] The controller assembles the collected execution status signals according to a preset status feedback data frame format and transmits them back to the monitoring and early warning system through the same communication link. The system parses the transmitted data frames, extracts the execution status code corresponding to each action command, and compares the execution status code bit by bit with a preset success status mask to confirm the linkage execution result.

[0155] If the execution status codes of all action commands in the operation sequence match the success status mask, the linkage is considered successfully executed, and the system records the linkage completion event and its completion time. If the execution status code of at least one action command indicates an execution error (including equipment failure, action timeout, or communication timeout), the linkage is considered an execution error, and the system triggers a policy correction process: First, the action command with the execution error is resent; if the execution status code is still abnormal after resentment, the system automatically checks whether the space unit is configured with a backup actuator. If there is a backup exhaust fan or backup drainage pump, the original command is switched to the backup actuator and resent; if there is no backup device or the backup device is also abnormal, the system raises the warning level of the command by one level and writes this abnormal event into the abnormal event log (recording the peak risk node identifier, abnormal type, and occurrence time), and pushes an alarm notification to the operation and maintenance terminal.

[0156] The maximum number of retransmissions can be set according to actual needs, and the following conditions must be met: ensuring communication reliability and avoiding network congestion caused by invalid retransmissions. This embodiment of the invention does not impose too many restrictions here, for example, it can be set to 3 times.

[0157] During the closed-loop verification process, the system also makes an additional determination of a type of invalid linkage based on the feedback status code: when the execution status code indicates that the device is already in the target state or the command is repeated, that is, when the target state required by the instruction is consistent with the current actual state of the actuator and the actuator has not undergone a state transition, the system marks this linkage as invalid linkage, indicating that the activation of the node that triggered the linkage is a pseudo-activation and no real physical state change has occurred.

[0158] The system constructs an invalid linkage sliding counting window for each risk node, with a window duration of [length missing]. The specific settings can be configured according to actual needs, and the following conditions must be met: covering the typical operating cycle of the utility tunnel. This embodiment of the invention does not impose too many restrictions here, for example, 3600 seconds.

[0159] Statistical analysis of the same risk node within the window The cumulative number of times marked as invalid linkage .when Reaching the preset frequency threshold When the threshold is low, it indicates that the dynamic adaptive threshold of the node is too low in the current environment, and its ability to suppress steady-state noise or non-risk fluctuations is insufficient.

[0160] Among them, frequency threshold The specific settings can be configured according to actual needs, and the following conditions must be met: reflecting the upper limit of the system's tolerable false alarm frequency. This embodiment of the invention does not impose too many restrictions here, for example, taking 5 times.

[0161] The online self-learning module performs threshold adjustment: A dynamic adaptive threshold update mechanism based on historical sliding window statistics has been established in the above steps. The dynamic adaptive threshold is:

[0162]

[0163] In the formula, and The mean and standard deviation of the cosine approximation values ​​within the window. This is the sensitivity coefficient; the online self-learning module will connect the nodes. Sensitivity coefficient Increase step size The calculation formula is:

[0164]

[0165] In the formula The sensitivity coefficient before adjustment; This is the adjusted sensitivity coefficient; The step size for adjusting the sensitivity coefficient in a single step can be set according to actual needs, and the following conditions must be met: finely adjust the threshold with a sufficiently small step size and avoid excessive penalty. This embodiment of the invention does not impose too many restrictions here, for example, it can be 0.1.

[0166] The increased limit is within a range not exceeding 5.0. Because... For positive, Increasing the threshold directly raises the dynamic adaptive threshold, raising the threshold for the node to be activated and reducing the probability of pseudo-activation under subsequent similar data distributions.

[0167] If, during long-term system operation, it is confirmed through manual routine inspections or post-incident tracing that a certain risk node has been missed (not activated when the actual risk event occurred), the β value of that node can be manually lowered through the system operation and maintenance interface, with the step size being the same each time. , The lower limit is controlled at 1.0, and this manual correction operation is also recorded in the online self-learning log.

[0168] Through the combined operation of the closed-loop verification and threshold self-learning described above, the real response state from the physical world on the actuator side is continuously fed back into the risk judgment threshold update on the sensing side, forming a hardware-level end-to-end closed-loop path. The output of this step is a linked execution result record and an updated dynamic adaptive threshold parameter set, which are used in the next complete detection cycle.

[0169] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for integrated monitoring and early warning of utility tunnels, characterized in that, include: Along the longitudinal mileage direction of the utility tunnel, the utility tunnel is divided into multiple non-overlapping topologically constrained spatial units; Various types of sensors are deployed in the utility tunnel and mounted to the corresponding topological constraint space units to generate a physical mounting topology mapping table for the utility tunnel. For each topological constraint space cell in the entity mounting topology mapping table, a dynamic sliding time window is constructed based on the current moment. The sensor data within the dynamic sliding time window is preprocessed to obtain a multidimensional heterogeneous data matrix. Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window and the multidimensional heterogeneous data matrix, the attention weights in the cross-modal feature fusion process are suppressed and adjusted to generate a comprehensive state representation vector for each topological constraint space unit at the current moment. A risk evolution knowledge ontology network with a directed graph structure is constructed for the utility tunnel, the network containing several nodes and directed edges; By calculating the cosine approximation between the comprehensive state representation vector and the baseline feature vector of each node, and combining the dynamic adaptive threshold to activate the node, the dynamic adaptive threshold update value of each node is obtained. The set of active nodes is selected based on dynamically adaptive threshold update values; Using the active node as the diffusion source, and combining the directed edge, the ventilation direction and slope between nodes, and the spatial distance, conditional probability diffusion iteration is performed to obtain the diffusion intensity of the source node after each iteration. Extract peak risk nodes where the diffusion intensity exceeds the threshold, and generate graded linkage early warning commands based on node type and diffusion intensity.

2. The integrated monitoring and early warning method for utility tunnels according to claim 1, characterized in that, The preprocessing of sensor data within the dynamic sliding time window to obtain a multidimensional heterogeneous data matrix includes: For sensor data with sampling frequencies lower than the standard characteristic sampling frequency, linear interpolation time-domain expansion is applied. For sensor data with sampling frequencies not lower than the standard characteristic sampling frequency, local Gaussian smoothing filtering and equal-interval downsampling with the standard characteristic frequency as the period are applied sequentially to align all data to the same standard frequency, forming a multidimensional heterogeneous data matrix.

3. The integrated monitoring and early warning method for utility tunnels according to claim 1, characterized in that, Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window and the multidimensional heterogeneous data matrix, the attention weights in the cross-modal feature fusion process are suppressed and adjusted to generate a comprehensive state representation vector for each topological constraint space unit at the current time, including: Based on the offset between the sensor data distribution and the historical steady-state distribution within the dynamic sliding time window, the local information entropy deviation distribution factor of the sensor's sampled data at the current moment is obtained. Based on a multidimensional heterogeneous data matrix, the adjusted cross-attention weights of any two sensors within a topologically constrained spatial cell at the current time are obtained. The preprocessed sampled data of the sensor within the dynamic sliding time window at the current moment constitutes the original temporal feature vector of the sensor; the original temporal feature vector of each sensor is weighted and aggregated using the adjusted cross attention weights to obtain the state representation vector of each sensor in the topological constraint space cell at the current moment. The mean of the state representation vectors of all sensors within the topologically constrained spatial cell at the current moment is used as the comprehensive state representation vector of the topologically constrained spatial cell at the current moment.

4. The integrated monitoring and early warning method for utility tunnels according to claim 3, characterized in that, The method for obtaining the local information entropy deviation distribution factor of the sensor's sampled data at the current moment based on the offset between the sensor data distribution within the dynamic sliding time window and the historical steady-state distribution includes: For any sensor, the preprocessed sampled data sequence within the current dynamic sliding time window is discretized using an adaptive histogram binning strategy based on historical data distribution to obtain the probability density distribution of the sensor's sampled data at the current moment; based on the probability density distribution, the information entropy of the sensor's sampled data within the current dynamic sliding time window is obtained. The absolute value of the difference between the information entropy of the data sampled by the sensor within the current dynamic sliding time window and the sensor's historical steady-state intrinsic information entropy is used as the local information entropy deviation distribution factor of the sensor's sampled data at the current moment.

5. The integrated monitoring and early warning method for utility tunnels according to claim 3, characterized in that, The method of obtaining the adjusted cross-attention weights of any two sensors within a topologically constrained spatial cell at the current moment based on a multidimensional heterogeneous data matrix includes: The multidimensional heterogeneous data matrix is ​​input into the standard Transformer design algorithm. Using two trainable weight matrices, a linear mapping is performed through two independent fully connected layers to obtain the query tensor and the key tensor of each sensor. Based on the query tensor and key tensor of the sensor, obtain the scaled dot product similarity between the query tensor and key tensor of any two sensors. Based on the scaled dot product similarity between the query tensor and the key tensor of the sensor, the standard cross-attention weights of any two sensors are obtained. By using the local information entropy deviation distribution factor to correct the standard cross-attention weights, the adjusted cross-attention weights of any two sensors at the current time can be obtained.

6. The integrated monitoring and early warning method for utility tunnels according to claim 1, characterized in that, The process involves calculating the cosine approximation between the comprehensive state representation vector and the baseline feature vector of each node, and then using a dynamic adaptive threshold to activate the nodes to obtain the dynamic adaptive threshold update value for each node. This includes: Based on the comprehensive state representation vector, obtain the cosine approximation degree between each node and the topological constraint space unit; construct a historical sliding window of cosine approximation degree; obtain the dynamic adaptive threshold update value of each node through the mean and standard deviation of the cosine approximation degree within the historical sliding window of cosine approximation degree.

7. The integrated monitoring and early warning method for utility tunnels according to claim 1, characterized in that, The process of selecting the set of active nodes based on dynamically adaptive threshold updates includes: For any topological constraint space cell, the cosine approximation degree between the topological constraint space cell and the node is compared with the dynamic adaptive threshold update value of the node: if the cosine approximation degree between the topological constraint space cell and the node is greater than or equal to the dynamic adaptive threshold update value of the node, the node is activated; if the cosine approximation degree between the topological constraint space cell and the node is less than the dynamic adaptive threshold update value of the node, the node is not activated; all activated nodes constitute the set of activated nodes within the current dynamic sliding time window.

8. The integrated monitoring and early warning method for utility tunnels according to claim 1, characterized in that, The process involves using the active node as the diffusion source point, combining directed edges, ventilation direction and slope between nodes, and spatial distance, to perform conditional probability diffusion iteration, obtaining the diffusion intensity of the source node after each iteration, including: The iteration is initiated with the activation node as the diffusion source point. The risk evolution knowledge ontology network is retrieved to obtain the directed evolution weights of the directed edges between the source nodes, and the ventilation direction, wind speed and structural slope data of the topological constraint space unit corresponding to the source node are read simultaneously. In each iteration, for each target source node, the comprehensive diffusion increment of the target source node is obtained; based on the comprehensive diffusion increment and the diffusion intensity of the target source node in the previous round, the diffusion intensity of the target source node after the current round is obtained; the above iterative operation is continuously executed until the preset maximum number of iterations is reached or the change in diffusion intensity is detected to be lower than the preset convergence threshold, thereby outputting the diffusion intensity of each source node after each iteration.

9. The integrated monitoring and early warning method for utility tunnels according to claim 8, characterized in that, The process of obtaining the comprehensive diffusion increment of the target source node includes: Traverse all source nodes pointing to the target source node, multiply the current diffusion intensity of each source node by the corresponding directional evolution weight, airflow correction factor, slope correction factor and distance attenuation factor in turn, and then add all the product results to obtain the comprehensive diffusion increment of the target source node.

10. The method for integrated monitoring and early warning of utility tunnels according to claim 1, characterized in that, The process involves extracting peak risk nodes where the diffusion intensity exceeds a threshold, and generating tiered, linked early warning commands based on node type and diffusion intensity, including: After traversing all source nodes after the iteration is completed, the source nodes whose diffusion intensity exceeds the preset diffusion peak threshold are identified as peak risk nodes; sorted by diffusion intensity from high to low, a graded linkage early warning instruction is generated for each peak risk node according to the risk node type and the preset risk classification rules.

Images