Facility anomaly real-time early warning method and system for road operation and maintenance

Abstract

This invention discloses a real-time early warning method and system for road maintenance facilities, belonging to the field of intelligent transportation technology. The method includes: constructing a facility graph model based on dual rules of physical proximity and functional dependency; introducing a dynamic weight correction mechanism to generate a dynamic weighted adjacency matrix; real-time acquisition of multi-source sensor time-series data, which, after preprocessing, is input into a time-series graph neural network; extracting spatiotemporal features of nodes through spatial feature aggregation and multi-scale temporal evolution modeling; mapping the final hidden state to a future anomaly probability sequence; dynamically adjusting the early warning threshold based on the number of historical false alarms, and comparing it with the predicted probability to trigger tiered early warnings. This invention solves the problems of difficult identification of clustered anomalies and delayed early warning response in complex environments by quantifying the implicit functional coupling relationships between facilities and fusing residual graph convolution and multi-scale dilated convolution into a TGNN architecture, achieving high-precision, adaptive proactive early warning of road facility anomalies.

Description

Technical Field

[0001] This application belongs to the field of intelligent transportation and road infrastructure operation and maintenance technology, specifically involving a method and system for real-time early warning of facility anomalies for road operation and maintenance. Background Technology

[0002] With the continuous evolution of intelligent transportation and infrastructure operation and maintenance technologies, the monitoring of road facility anomalies is gradually shifting from passive response to proactive early warning. Traditional operation and maintenance models mainly rely on manual periodic inspections or threshold alarm mechanisms based on single sensors. The former is limited by high labor costs, long coverage periods, and subjective judgment biases, while the latter, by focusing only on instantaneous conditions, struggles to identify gradual deterioration trends in facility performance. Especially in high-density urban road networks, hazards such as loose manhole covers and displaced guardrails often experience sudden failures after long-term minor changes, posing a potential threat to public safety. This places higher demands on the predictive capabilities, response timeliness, and system coordination of monitoring systems.

[0003] In particular, early warning systems for facility anomalies in road maintenance require the integration of multi-source sensor information and modeling of the spatiotemporal relationships between facilities. However, existing methods generally treat each facility as an isolated node, lacking an effective characterization of their physical proximity or functional dependence, and failing to capture coupled failure phenomena such as the cascading collapse of adjacent guardrails or the coordinated settlement of drainage well groups. Furthermore, fixed threshold strategies are ill-suited to complex environmental disturbances (such as instantaneous vibrations caused by heavy vehicle traffic), easily leading to false alarms or missed alarms; and the lack of modeling capabilities for future state evolution trends means the system can only provide "post-event alerts," unable to support proactive maintenance decisions for early intervention.

[0004] Therefore, there is an urgent need for a real-time early warning method and system for road facility anomalies that can integrate facility network topology, fuse multi-dimensional time-series sensing data, and have dynamic prediction and adaptive discrimination capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for real-time early warning of facility anomalies in road operation and maintenance, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] Firstly, a method for real-time early warning of facility anomalies for road operation and maintenance includes the following steps:

[0008] A graph model of road facilities is constructed, which includes a set of nodes and a set of edges. Nodes correspond to facilities where sensors are deployed, and edges are generated according to the dual rules of physical proximity and functional dependence between facilities. Static weights are assigned to each edge, and dynamic correction coefficients are calculated based on real-time environmental data and the service life of facilities to obtain a dynamic weighted adjacency matrix.

[0009] Real-time acquisition of multi-dimensional sensor time-series data from various facilities; preprocessing of raw data to form time window sequences;

[0010] The graph model and the time window sequence of each node are input into the temporal graph neural network, which includes a spatial feature aggregation module and a temporal evolution module. The spatial feature aggregation module uses a graph convolutional network and performs time-step spatial feature extraction based on a dynamic weighted adjacency matrix to obtain the embedding vector of each node at each time step. The temporal evolution module performs temporal modeling on the embedding vector sequence of each node and outputs the final hidden state of each node.

[0011] The final hidden state is input into a fully connected network and mapped to an anomaly probability sequence for multiple time steps within a future preset time window;

[0012] The warning threshold is dynamically adjusted based on the number of historical false alarms, and the maximum value of the predicted anomaly probability sequence at the current moment is compared with the dynamic threshold. If the threshold is exceeded, a warning work order is generated.

[0013] Preferably, the dual rules of physical proximity and functional dependency include:

[0014] If the physical distance between two facility nodes is less than 10 meters, an undirected edge is established between the corresponding nodes;

[0015] If the physical distance exceeds 10 meters but the municipal pipeline network topology database verifies that the two belong to the same continuous drainage pipeline network, or the power SCADA system data confirms that the two share the same power supply circuit, or the two are clearly located on both sides of the same bridge expansion joint, then an undirected edge will be forcibly established.

[0016] Preferably, the dynamic correction coefficient is calculated by a multilayer sensor. The input of the multilayer sensor includes the current ambient temperature, ambient relative humidity, the average number of vehicles passing through in the past 5 minutes, and the service life of the facilities corresponding to the two nodes.

[0017] The multilayer perceptron contains two hidden layers. The first hidden layer has 8 neurons and the second hidden layer has 4 neurons. The ReLU activation function is used after the hidden layers, and the Sigmoid activation function is used in the output layer. The output value is linearly transformed and mapped to the range of 0.8 to 1.2 as a dynamic correction coefficient.

[0018] The dynamic weighted adjacency matrix is ​​obtained by multiplying the static weights and dynamic correction coefficients element by element.

[0019] Preferably, the spatial feature aggregation module consists of two layers of graph convolutional networks, each of which is calculated based on a dynamic adjacency matrix with self-loops and its corresponding degree matrix.

[0020] The first layer of graph convolution maps the input features to 16 dimensions, and the second layer of graph convolution maps the 16-dimensional features to 32 dimensions. A residual connection is introduced between the two layers, which adds the output of the first layer of graph convolution to the original input after linear transformation to align the dimensions, and then uses it as the input of the second layer of graph convolution. All time steps share the same graph convolution parameters, and graph convolution operations are performed independently for each time step.

[0021] Preferably, the temporal evolution module includes a multi-scale temporal convolution submodule and a gated recurrent unit submodule;

[0022] The multi-scale temporal convolutional submodule uses three parallel causal convolutional layers, each with a kernel size of 3 and dilation rates of 1, 2, and 4, respectively. Each layer has 32 output channels. The outputs of the three scales are concatenated and then subjected to linear transformation and ReLU activation to obtain a 64-dimensional multi-scale fusion feature.

[0023] The original embedding and the multi-scale fused features are concatenated and then input into a gated linear unit. The output of the gated linear unit is added to the original embedding through a residual connection to obtain the enhanced features.

[0024] The enhanced feature sequence is input into a gated recurrent unit with a hidden state dimension of 64, and the hidden state at even time steps is added to the hidden state two time steps earlier during the iteration process to enhance long-term dependency propagation.

[0025] Preferably, the future preset time window is 15 minutes, and the anomaly probability sequence contains probability values ​​for 15 time steps, with each time step corresponding to the next minute;

[0026] When training the fully connected network, a weighted binary cross-entropy loss function is used, and positive samples and abnormal events are given a weight of 5 times. The loss function calculates the weighted cross-entropy at each time step and takes the average value as the final loss of the sample.

[0027] Preferably, dynamically adjusting the early warning threshold includes:

[0028] The number of false alarms is counted on a 24-hour or 12-hour cycle. A false alarm refers to an event in which a warning work order has been generated and the facility status has been confirmed to be normal after manual verification.

[0029] If the number of false alarms exceeds 5 or exceeds 3 within a period, the current threshold will be increased by 0.05 or 0.04.

[0030] If the number of false alarms is less than or equal to 2 or less than or equal to 1, the threshold will be lowered by 0.03 or 0.02.

[0031] The adjusted threshold is limited to a closed interval of 0.55 to 0.85.

[0032] Preferably, when acquiring multidimensional sensing time-series data in real time, all sensors are driven by the same hardware clock source, and synchronous sampling is triggered by the hardware timer of the main control chip, with the synchronization error controlled within ±2 milliseconds;

[0033] Preprocessing includes moving average filtering and temperature compensation. The moving average filtering window has a length of 5 sampling points. Temperature compensation uses a polynomial model. The compensation formula is that the compensated tilt angle is equal to the original tilt angle plus the temperature compensation coefficient multiplied by the difference between the current temperature and the calibrated reference temperature. The acceleration is handled similarly.

[0034] Preferably, for the scenario of mountainous highway tunnel groups, when constructing the graph model, virtual nodes of geological units are introduced, the tunnel entrance and exit are abstracted as independent nodes, and functional dependency edges are established with drainage ditches and slope anchor facilities within a radius of 30 meters, with the weight uniformly set to 0.6;

[0035] For the scenario of urban expressway overpasses, a three-level hierarchical subgraph is constructed, consisting of piers, beams, and expansion joints. Nodes within the same pier subgraph are fully connected and have a weight of 0.9, nodes within the same beam subgraph have a weight of 0.8, and nodes within the same expansion joint subgraph have a weight of 0.7.

[0036] For high-density overpass scenarios, the first layer of the temporal graph neural network, graph convolution, is deployed on edge nodes, while the second layer, graph convolution, and fully connected layer are deployed on the cloud. The edge nodes perform global max pooling on the 16-dimensional vector at 60 time steps and then upload the 16-dimensional aggregated vector.

[0037] Secondly, a real-time early warning system for facility anomalies in road operation and maintenance includes:

[0038] The graph model building module is used to construct nodes and edges based on facility registration information, geographic coordinates and functional dependencies, and introduces a dynamic weight correction mechanism to generate a dynamic weighted adjacency matrix.

[0039] The data acquisition and preprocessing module is deployed at each facility node and includes hardware-synchronized multi-sensor units and edge preprocessing units, which are used to acquire multi-dimensional sensor data in real time and form a time window sequence.

[0040] The temporal graph neural network processing module, deployed on a central server or edge-cloud collaborative architecture, is used to perform spatial feature aggregation and temporal evolution modeling, and output the final hidden state of each node.

[0041] The anomaly probability prediction module is used to map the final hidden state into a future anomaly probability sequence via a fully connected network;

[0042] The dynamic threshold early warning module is used to dynamically adjust the threshold based on the number of historical false alarms, compare the predicted probability with the threshold, and trigger the generation of early warning work orders and hierarchical response.

[0043] The modules are interconnected through a communication network to enable real-time early warning of road facility anomalies.

[0044] In summary, this application includes at least one of the following beneficial technical effects:

[0045] 1. This invention constructs a facility graph model based on dual rules of physical proximity and functional dependence, and introduces a dynamic weight correction mechanism driven by environmental factors and service life. It quantifies the implicit functional associations of municipal pipe networks, power supply circuits, etc., into a dynamically evolving graph topology structure. This solves the technical problem of traditional methods treating facilities as isolated nodes and failing to capture cluster anomalies such as coordinated settlement of well groups and chain collapse of guardrails. It significantly improves the ability to perceive the risk of facility coupling failure in complex environments.

[0046] 2. This invention employs a temporal graph neural network architecture consisting of a two-layer graph convolutional network with residual connections and a multi-scale temporal convolutional-gated recurrent unit. It extracts multi-scale temporal features through parallel causal convolutions with dilation rates of 1, 2, and 4, and enhances long-term dependency propagation by combining gated linear units and skip connections. This effectively solves the oversmoothing problem of deep graph convolutions and the shortcomings of single recurrent networks in simultaneously capturing short-term sudden disturbances and long-term gradual degradation trends, thus achieving high-precision modeling of the spatiotemporal evolution of facility status.

[0047] 3. This invention assigns a 5-fold weight to positive samples using a weighted binary cross-entropy loss function, overcoming the problem of insufficient model training caused by the scarcity of abnormal event samples. Combined with a step-by-step dynamic threshold adjustment mechanism based on the number of historical false alarms, the threshold parameters are automatically optimized according to the false alarm feedback within a 24-hour or 12-hour period, forming a closed-loop feedback system from anomaly prediction to operation and maintenance verification to model self-optimization. This significantly reduces the false alarm rate caused by environmental interference and achieves adaptive and highly sensitive proactive early warning for road facilities. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the real-time early warning method for facility anomalies in road maintenance according to the present invention.

[0049] Figure 2 This is a schematic diagram of the process of jointly modeling by fusing graph structures and time-series data in this invention;

[0050] Figure 3 This is a schematic diagram of the process of constructing a road infrastructure map model and collecting multi-dimensional sensor time-series data in real time in this invention;

[0051] Figure 4 This is a schematic diagram of the process in this invention for predicting the probability of future anomalies, performing dynamic threshold discrimination, and generating early warnings;

[0052] Figure 5This is a schematic diagram of the multi-level interaction relationship and data flow between edge nodes and the central server in this invention. Detailed Implementation

[0053] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of specific embodiments based on the present invention is provided in conjunction with the accompanying drawings and preferred embodiments. See attached figures for details. Figures 1 to 5 The content shown.

[0054] Example 1

[0055] A method for real-time early warning of facility anomalies for road operation and maintenance includes the following steps:

[0056] Step S1, constructing the road infrastructure map model, specifically includes the following operation process:

[0057] Step S101: First, all monitorable facilities deployed along urban main roads, secondary roads and highways are digitally registered. The registration objects include manhole covers, metal or concrete guardrail sections, street light pole bases, rainwater inlet grilles and traffic signal poles.

[0058] Each facility is assigned a unique facility identifier, which uses a 12-digit hexadecimal code. The first 4 digits represent the administrative division code where the facility is located, the middle 4 digits represent the road number to which it belongs, and the last 4 digits are the serial number of the facility on that road, ensuring that each facility can be uniquely identified throughout the entire network.

[0059] Step S102: Based on high-precision geographic information system data, extract the latitude and longitude coordinates of each registered facility, and calculate the Euclidean distance between any two facilities based on this. This distance value will serve as one of the main bases for determining whether a connection is established between nodes.

[0060] Step S103: Construct a graph node set V, where the number of nodes is equal to the number of facilities where sensors are actually deployed.

[0061] The generation of the graph edge set E follows two rules: First, if the physical distance between two facilities is less than 10 meters, an undirected edge is established between the corresponding nodes; Second, even if the physical distance between two facilities exceeds 10 meters, if the municipal pipeline network topology database verifies that they belong to the same continuous drainage pipeline network, or if the power SCADA system data confirms that they share the same power supply circuit, or if they are clearly located on both sides of the same bridge expansion joint, then an undirected edge is forcibly established.

[0062] The above rules ultimately form a static graph topology G=(V,E) that describes the spatial and functional relationships of facilities.

[0063] Step S104: Introduce a weight matrix for the graph edges. Its elements Used for quantization nodes and The physical coupling strength between them.

[0064] Weight assignment is preset based on facility type combination as follows: When and When all are adjacent guardrail sections, When both are upstream and downstream manhole covers on the same drainage pipe section, When two entities are not directly adjacent but have functional dependencies, such as streetlights sharing the same transformer for power supply, For node pairs without edge connections, .

[0065] diagonal elements Set to 1 to preserve the node's own characteristics.

[0066] The aforementioned weight matrix will be used directly as a weighted adjacency matrix in subsequent graph convolution operations, thereby reflecting the differences in the influence strength of different neighboring nodes on the central node.

[0067] Step S104A: Introduce a dynamic coupling strength self-adjustment mechanism:

[0068] Considering that the coupling strength between facilities is not constant but is affected by real-time environmental factors such as temperature and humidity, rainfall intensity, traffic flow and the aging of the facilities themselves, this embodiment further introduces a dynamic weight correction module on the basis of the static weight matrix W. This module adopts a lightweight multilayer perceptron, takes the current environmental perception data and the service life of the two facilities as input, and outputs a dynamic correction coefficient to adjust the weight of the edge.

[0069] Specifically, for any pair of connected nodes and Collect the following inputs:

[0070] The current ambient temperature T is in degrees Celsius, the ambient relative humidity H is in percentage, the average number of vehicles passing through in the past 5 minutes V is in vehicles per minute, and the node... Corresponding facility service life and nodes Corresponding facility service life All units are years.

[0071] Concatenate these five values ​​into a 5-dimensional input vector. .

[0072] The vector is input into a multilayer perceptron containing two hidden layers: the first hidden layer has 8 neurons, and the second hidden layer has 4 neurons. Each hidden layer is followed by a ReLU activation function. The output layer is a single neuron, and a Sigmoid activation function is used to obtain an intermediate value between 0 and 1. .

[0073] To limit the correction factor to the range of 0.8 to 1.2, for Perform a linear transformation to obtain the dynamic correction coefficients. , The value range is [0.8, 1.2].

[0074] For all connected node pairs in the graph, their dynamic correction coefficients are calculated one by one, and the results are used to construct a dynamic correction coefficient matrix ΔW, whose dimension is the same as the static weight matrix W. For unconnected node pairs, the corresponding values ​​in ΔW are set to 0. The final weighted adjacency matrix used for graph convolution is... It is obtained by multiplying the static weights and dynamic correction coefficients element by element, that is:

[0075]

[0076] Where ⊙ represents the Hadamard product, this dynamic weight matrix is ​​calculated in real time based on the latest environmental data and facility service life before each model inference, so that the edge weights of the graph topology can adaptively reflect the real-time changes in the physical coupling strength between facilities.

[0077] To further enhance the model's adaptive focusing ability on key neighbors, a graph attention mechanism can be superimposed in subsequent graph convolution calculations. Specifically, in each graph convolution layer, attention coefficients are dynamically calculated based on the node features of the current layer, and combined with dynamic weights as the final aggregate weights.

[0078] Specifically, for nodes and its neighboring nodes Attention coefficient The calculation method is as follows:

[0079]

[0080] in and They are nodes and The input feature vector in the current layer has a dimension that depends on the output of the previous layer; W is a shared linear transformation matrix, which maps the input features to the same high-dimensional space; a is a learnable attention parameter vector; || denotes the vector concatenation operation; LeakyReLU is a leaky ReLU activation function with a negative slope of 0.2. Represents a node The set of all neighboring nodes (including itself).

[0081] The attention coefficient mentioned in step S104A Reflects neighboring nodes For nodes The importance of.

[0082] During aggregation, dynamic weights will be used. With attention coefficient Multiplication serves as the final aggregation weight, thereby dynamically adjusting neighbor contributions based on real-time node characteristics while considering physical coupling strength. This effectively suppresses interference from noisy neighbors and highlights facilities that play a key role in anomalous evolution.

[0083] By integrating the aforementioned dynamic coupling strength self-adjustment mechanism and graph attention mechanism, the complex and time-varying interactions between facilities can be captured more accurately, thereby improving the accuracy of anomaly warnings.

[0084] Step S105: The completed graph structure data, including node attributes, edge connection relationships, and weight matrix, is stored in the graph database of the central server in the form of an adjacency list, supporting millisecond-level query response.

[0085] The system has incremental update capabilities. When new facilities are deployed or existing facilities are demolished, the graph topology reconstruction process is automatically triggered to recalculate the distances and functional dependencies between nodes, update edge connections and weight assignments, and ensure that the model input always reflects the latest actual state of the facility network.

[0086] Step S2 involves real-time acquisition of multi-dimensional sensing time-series data. The specific execution process is as follows:

[0087] Step S201: Deploy an integrated multi-sensor unit at each graph node. The unit adopts a low-power embedded system architecture, uses an ARM Cortex-M7 core as the main control chip, operates at a frequency of 480MHz, and has 256KBSRAM and 2MB Flash storage to perform data acquisition and preliminary processing tasks.

[0088] Step S202: The integrated array of the multi-sensor unit includes a MEMS triaxial accelerometer, a high-precision tilt sensor, a digital temperature and humidity sensor, and a laser displacement rate detection module.

[0089] The triaxial accelerometer has a range of ±8g, a resolution of 0.1mg, and a bandwidth of 100Hz; the tilt sensor has a measurement range of ±90 degrees and an accuracy of ±0.01 degrees; the digital temperature and humidity sensor has a temperature measurement range of -40 degrees Celsius to +85 degrees Celsius and an accuracy of ±0.3 degrees Celsius, and a humidity measurement range of 0% to 100% relative humidity and an accuracy of ±2% relative humidity; the laser displacement rate detection module has a sampling frequency of 1kHz and a resolution of 1 micrometer per second.

[0090] All sensors are connected to the main control chip via I²C and SPI buses and driven by the same hardware clock source. The hardware timer of the main control chip triggers synchronous sampling to ensure that the timing of multi-source data acquisition is strictly consistent, and the synchronization error is strictly controlled within ±2 milliseconds.

[0091] In step S203, the system continuously acquires data at a fixed frequency of 1Hz, generating a set of raw observation values ​​containing 5 dimensions per second, specifically including the tilt angle θ (in degrees), and the acceleration components of the X, Y, and Z axes. , , The units are meters per second squared, ambient temperature T is in degrees Celsius, relative humidity H is in percentage, and displacement rate of change δ is in millimeters per second.

[0092] In step S204, the raw data stream is first preprocessed at a local edge computing node. This edge node is deployed in a smart gateway device on the roadside, running a lightweight Linux system and equipped with a 4-core CPU and 8GB of memory.

[0093] The preprocessing procedure includes two operations: moving average filtering and temperature compensation.

[0094] The moving average filtering is performed independently for each sensing dimension. The window length is set to 5 sampling points corresponding to 5 seconds. That is, the arithmetic mean of the current moment and the previous 4 sampling points is taken as the filter output for that moment, which is used to suppress high-frequency vibration interference caused by the instantaneous crushing of heavy vehicles.

[0095] The temperature compensation algorithm employs a polynomial compensation model, dynamically correcting the tilt angle and acceleration readings based on the current ambient temperature T. The specific compensation formula is as follows:

[0096] Compensated tilt angle Compensated acceleration ,in and These are the temperature compensation coefficients for tilt angle and acceleration, respectively. To calibrate the reference temperature, these coefficients are obtained by performing high and low temperature chamber calibration experiments on each sensor unit before leaving the factory, and are fixed in the firmware of the sensor unit in the form of lookup tables or polynomial coefficients.

[0097] In step S205, the filtered and compensated data is organized into a time window sequence. The sequence is located in the L×d dimension space of the real number field R, where the time window length L is set to 60 to correspond to 60 seconds of historical data, and the sensing dimension d is equal to 5.

[0098] The above time window sequence As input features for subsequent neural networks.

[0099] In step S206, each time window sequence is uploaded to the central server once per minute via the MQTT protocol. During transmission, the AES-256 encryption algorithm is used to ensure data confidentiality, and a CRC32 checksum is added for integrity verification.

[0100] Step S3: Combine graph structure and time series data for joint modeling.

[0101] Step S301: Combine the static graph topology G=(V,E) constructed in step S1 with the multi-dimensional time-series data of each node collected in step S2. The inputs are fed into a pre-trained temporal graph neural network (TGNN), which employs a two-stage hybrid architecture to extract spatiotemporal features of facility status.

[0102] The node set V corresponds to the number of facilities where sensors are deployed, and the edge set E is defined by both proximity and functionality rules, including time-series data. The 60×5-dimensional time window matrix is ​​obtained after preprocessing for each node.

[0103] Step S302, the first stage of TGNN is the spatial feature aggregation module, which consists of a two-layer graph convolutional network (GCN). Input feature matrix. The dimension is Where L=60 is the time window length and d=5 is the original sensing dimension.

[0104] Since graph convolution operations are typically performed on the node feature matrix, and the input here includes a time dimension, graph convolution operations need to be performed independently for each time step.

[0105] Specifically, Consider it as L independent time slices, each slice corresponding to one The matrix is ​​denoted as , The same graph convolutional layer is applied to each time slice, meaning that all time steps share the same learnable parameters.

[0106] After the first layer of graph convolution, each time slice yields the output. Stacking these outputs along the time dimension yields a three-dimensional tensor. .

[0107] Similarly, the second layer of graph convolution is... Each time slice As input, we get After stacking, we get .

[0108] In this way, the spatial aggregation process is parameter-shared and parallel in the time dimension.

[0109] To incorporate dynamic coupling strength into the graph convolution process, this embodiment uses the dynamic weighted adjacency matrix calculated in step S104A. This serves as the basic adjacency matrix. In each layer of graph convolution, a dynamic adjacency matrix with self-loops is first constructed. , where I is the identity matrix, used to preserve the node's own characteristics. Correspondingly, its degree matrix... For a diagonal matrix, the diagonal elements The formula for calculating graph convolution is then updated to:

[0110]

[0111] in For the first The layer's output at time t has a dimension of ; It is the ReLU activation function; The dynamic weighted adjacency matrix is ​​calculated in step S104A; I is the identity matrix, with dimensions equal to 1 / 2. same; It is a dynamic adjacency matrix with self-loops; for The corresponding degree matrix is ​​a diagonal matrix, and its diagonal elements are... ; Let be the learnable weight matrix of the l-th layer, with dimension . ,in , , .

[0112] After the first layer of GCN, each node obtains a 16-dimensional intermediate representation at each time step; after the second layer of GCN, each node obtains a 32-dimensional embedding vector at each time step, denoted as... subscript Traverse the nodes, By iterating through the time steps, the vector integrates the spatial context information of the node itself and its first-order and second-order neighbors at the corresponding time.

[0113] In addition, to alleviate the oversmoothing problem caused by multi-layer GCNs, this embodiment introduces residual connections between two GCNs.

[0114] Specifically, the output of the first-layer GCN When combining with the input of the second-layer GCN, it must first be combined with the original input. Adding them together, because The feature dimension is 5, while The feature dimension is 16, and it is necessary to... Perform a linear transformation to align the dimensions. This operation is performed independently at each time step: for each time slice. First, through a learnable linear transformation Map it to Then, with the first layer output The results are summed and then used as input to the second layer of the GCN. Therefore, the actual output of the first layer of the GCN is:

[0115]

[0116] This design ensures that each node retains its initial characteristics after multi-layer aggregation, avoiding the convergence of node characteristics due to excessive smoothing, thereby maintaining the distinguishability of different facilities and laying the foundation for subsequent accurate location of abnormal facilities.

[0117] The embedding vectors of all nodes form a set. This serves as the input for the next stage of the time evolution modeling module.

[0118] Step S303: The second stage of TGNN is the time evolution modeling module, which uses a single-layer gated recurrent unit (GRU).

[0119] For each node The embedded sequence obtained after spatial aggregation Input to the GRU cell in chronological order. The GRU is accessed via a reset gate. and the update gate The calculation process for controlling the flow of information is as follows:

[0120]

[0121] In the formula, This is the hidden state from the previous moment. The current hidden state is represented by both 64-dimensional vectors. The 32-dimensional embedding vector input at the current time step; , , σ represents the learnable parameter matrix; σ represents the Sigmoid activation function; tanh represents the hyperbolic tangent activation function; ⊙ represents element-wise multiplication.

[0122] The aforementioned gating mechanism enables the network to selectively forget early information that is irrelevant to the current prediction, while retaining long-term dependency features that reflect the gradual degradation trend of the facility.

[0123] Step S303A introduces a fusion mechanism of multi-scale temporal convolution and residual gating.

[0124] To address the issue that a single GRU cannot simultaneously capture both short-term sudden disturbances and long-term gradual degradation trends, this embodiment introduces a multi-scale temporal convolutional network as a front-end feature extractor in the temporal evolution modeling module, and deeply integrates it with the GRU. Specific improvements are as follows:

[0125] (1) Multi-scale temporal feature extraction.

[0126] The embedding sequence obtained after spatial aggregation for each node ,in First, the sequence is fed into three parallel causal convolutional layers. The kernel size of each causal convolutional layer is 3, but the dilation rate is set to 1, 2 and 4 respectively. Therefore, the corresponding receptive fields are 3, 5 and 9 time steps respectively. Each causal convolutional layer contains 32 output channels. Therefore, the outputs of the three convolutional layers are concatenated in the feature dimension to obtain a 96-dimensional feature vector.

[0127] To reduce dimensionality and integrate multi-scale information, the 96-dimensional vector is input into a linear transformation layer with a weight matrix of size [size missing]. The bias is 64-dimensional, and after ReLU activation, a 64-dimensional multi-scale fused feature is obtained. This allows the model to simultaneously perceive local mutations (small expansion rate) and slow drifts (large expansion rate).

[0128] (2) Residual connection and gating fusion.

[0129] Original embedding Multiscale features Concatenating along the feature dimensions yields a vector of dimension 96. The vector is then input into a gated linear unit for feature selection, and the specific calculation is as follows:

[0130]

[0131]

[0132] in and For learnable weight matrix, and A learnable bias vector; This represents the Sigmoid activation function. This represents element-wise multiplication; This is a gating signal, controlling the proportion of information that can pass through.

[0133] Then With the original The enhanced temporal features are obtained by adding the residual connections. This operation helps to mitigate gradient vanishing and preserve the original information.

[0134] (3) Improved GRU unit.

[0135] Enhanced feature sequences Input a single-layer GRU, with the hidden state dimension of the GRU set to 64, and its internal calculation method is the same as that of the standard GRU.

[0136] To further enhance the propagation capability of long-term dependencies, skip connections are introduced during the GRU iteration process: for each time step When t is even, the hidden state at the current time step is... Hidden state two time steps prior Add, i.e., execute Then, the calculation continues for subsequent time steps, making the information flow between time steps that are far apart more direct.

[0137] Step S304: After L=60 time steps of iterative iteration, each node outputs the hidden state at the final time step. This vector integrates the historical evolution of the node itself and its spatial interaction with surrounding facilities, forming a context-aware representation of the facility's state.

[0138] All nodes Form a set This serves as the input for the subsequent anomaly probability prediction module.

[0139] Step S4, predicting the probability of future anomalies, is completed by the fully connected output layer.

[0140] Step S401: The system receives the final hidden state of each facility from the output of the Temporal Graph Neural Network (TGNN). The hidden state is a 64-dimensional vector, denoted as . The spatiotemporal evolution characteristics of the vector complex over the past 60 seconds.

[0141] Step S402, will The input undergoes a first fully connected linear transformation, which maps the 64-dimensional input to a 128-dimensional intermediate representation. The transformation process follows... ,in The learnable weight matrix is ​​128×64. It is a 128-dimensional bias vector.

[0142] Then, the hyperbolic tangent tanh activation function is applied to each element of the intermediate representation to obtain the activated 128-dimensional feature vector.

[0143] Step S403: The activated 128-dimensional feature vector is input into the second fully connected linear transformation layer. This layer maps the 128-dimensional input to a 15-dimensional output. The transformation process follows... ,in The learnable weight matrix is ​​15×128. It is a 15-dimensional bias vector.

[0144] Each component of the output vector z corresponds to the original value of the anomaly probability at each time step from 1 to 15 minutes in the future.

[0145] Step S404: Apply the Sigmoid function element-wise to the 15-dimensional vector z output by the second layer, compressing its values ​​to the interval between 0 and 1, to obtain the final anomaly probability sequence. .

[0146] in This represents the probability that the facility will experience an anomaly in the next t-minute period, where t ranges from 1 to 15, with each time step corresponding to 1 minute, and the entire prediction window is 15 minutes long.

[0147] In step S405, when constructing the training dataset, samples are extracted from historical time-series data and fault records using a sliding window method.

[0148] Specifically, for each facility, based on the timeline of its historical sensor data, any time t is selected as the prediction starting point, the multidimensional sensor time series data of the 60 consecutive seconds before the starting point is taken as the input feature, and the anomaly labels of each minute within the 15 consecutive minutes after the starting point are taken as the output target.

[0149] Anomaly tags are determined through manual on-site verification or historical fault work order records. If an abnormal event occurs within a certain minute (such as early signs like loose manhole covers or settlement of guardrail foundations), then that minute is tagged. Set it to 1, otherwise set it to 0.

[0150] This method allows for the collection of over 10,000 labeled samples.

[0151] Step S406, parameters of the fully connected layer , , , Determined through end-to-end training.

[0152] During training, the dataset is randomly divided into a training set, a validation set, and a test set. The training set is used for parameter updates, and the validation set is used to monitor model performance and determine early stopping to prevent overfitting.

[0153] The model training uses weighted binary cross-entropy as the loss function, assigning a 5x weight to positive sample anomalous events to improve the model's sensitivity to rare, high-risk events. The weighted binary cross-entropy is calculated as follows:

[0154]

[0155] Where y is the actual label (y=1 if there is an anomaly, otherwise y=0). Predict probabilities for the model.

[0156] For each sample, the weighted cross-entropy loss for 15 time steps is calculated first, and then the average value is taken as the final loss for that sample. The optimization process uses the Adam optimizer, with the learning rate set to 0.001 and the batch size to 32. Iterative training is performed until the validation set loss no longer decreases for several consecutive cycles.

[0157] Step S5: Perform dynamic threshold discrimination and generate an early warning.

[0158] In step S501, the system sets the initial value of the dynamic threshold τ to 0.65 during the initialization phase. This threshold will be compared with the abnormal probability output by the model to determine whether an early warning is triggered.

[0159] In step S502, the system continuously records the number of false alarm events in a 24-hour statistical cycle. The false alarm event refers to a situation where an early warning work order has been generated, but the actual condition of the facility is confirmed to be normal after manual on-site verification.

[0160] Every day at midnight, the system automatically reads the cumulative number of false alarms in the past 24 hours to provide a basis for threshold adjustment.

[0161] Step S503: The system dynamically adjusts the threshold based on the number of false alarms from the previous day.

[0162] The specific rules are as follows: if the number of false alarms is greater than 5 in the past 24 hours, the current threshold τ will be increased by 0.05; if the number of false alarms is less than or equal to 2, τ will be decreased by 0.03; if the number of false alarms is between 3 and 5, the threshold will remain unchanged.

[0163] The adjusted τ is limited to a closed interval of 0.55 to 0.85 to avoid the threshold being too sensitive or too insensitive.

[0164] Step S504: During real-time operation, for each facility node, obtain the anomaly probability sequence for the next 15 minutes generated in step S4. Calculate the maximum value max(P) of the sequence, which represents the highest probability of an anomaly occurring within the next 15 minutes.

[0165] In step S505, the system compares max(P) with the current dynamic threshold τ. If max(P) is greater than τ, the system determines that the facility is in a high-risk state within the next 15 minutes and immediately triggers the early warning work order generation process.

[0166] In step S506, when an early warning work order is generated, the system automatically fills in the following structured fields:

[0167] The facility's unique identifier (Facility ID), the number of minutes corresponding to the predicted peak time point (argmax(P) - representing the minute offset from the current time when the anomaly is expected to occur), the confidence level (max(P) value), the facility type, and the geographical location using the WGS-84 coordinate system.

[0168] Step S507: After the work order is generated, the Geographic Information System (GIS) service is automatically invoked to render the facility coordinates onto the electronic map, and the warning level is classified according to the predicted peak time advance.

[0169] Lead time is defined as the number of minutes corresponding to the predicted peak time. If the lead time is greater than or equal to 10 minutes, it is marked as a Level 1 warning; if the lead time is greater than or equal to 5 minutes but less than 10 minutes, it is marked as a Level 2 warning; if the lead time is less than 5 minutes, it is marked as a Level 3 warning.

[0170] Step S508: Different early warning levels correspond to different operation and maintenance response strategies.

[0171] Level 1 alerts trigger automatic dispatch to the nearest maintenance team, while also sending SMS and APP notifications to relevant personnel; Level 2 alerts only send APP notifications, requiring on-duty personnel to log in to the system for confirmation within 2 hours; Level 3 alerts only record logs for subsequent analysis and trend judgment, without triggering an immediate response.

[0172] In step S509, all generated early warning work orders are pushed to a unified operation and maintenance scheduling platform through a RESTful API interface. The interface adopts the OAuth 2.0 authentication mechanism and the data format is JSON to ensure seamless integration and data interoperability with the existing urban management system.

[0173] Example 2

[0174] In the special application scenario of mountainous highway tunnel groups, this embodiment makes adaptive adjustments to each step of Embodiment 1 to address the characteristics of high humidity, strong electromagnetic interference, and dense facilities, so as to further improve the early warning performance and reliability of the system in complex environments.

[0175] (1) Adjustments to the road facility map model constructed in step S1.

[0176] Based on the original graphical model, virtual nodes for geological units are introduced. The specific steps are as follows:

[0177] The entrance and exit of each tunnel are abstracted as independent geological unit nodes, denoted as... And establish a functionally dependent edge between it and the adjacent drainage ditch facilities and slope anchor facilities.

[0178] The term "adjacent" is defined as all drainage ditch nodes and slope anchor nodes within a radius of 30 meters centered on the center point of the tunnel entrance and exit, provided that the facility is confirmed by a geographic information system or design drawings to have an actual spatial relationship with the tunnel geological unit, such as the drainage ditch belonging to the tunnel drainage system or the anchor located within the tunnel slope protection area.

[0179] The weight of the functionally dependent edges is uniformly set to 0.6 to quantify the physical coupling strength between geological unit nodes and adjacent facilities.

[0180] By introducing virtual nodes, the graph model can more accurately reflect the potential impact of changes in geological conditions in the tunnel entrance and exit area on adjacent facilities.

[0181] (2) Adjustment of multidimensional sensing time sequence data collected in real time in step S2.

[0182] To cope with the high humidity and strong electromagnetic interference environment inside the tunnel, and to enhance the ability to detect the corrosion risk of metal components, the integrated multi-sensor unit adds two more sensors to the original five-dimensional sensor:

[0183] The barometric pressure sensor is a MEMS digital barometer with a measurement range of 300 to 1100 hPa, an accuracy of ±0.1 hPa, and a resolution of 0.01 hPa. It is connected to the main control chip via an I²C bus to monitor the atmospheric pressure difference between the inside and outside of the tunnel, and to help identify abnormal tunnel ventilation or sudden changes in external air pressure.

[0184] The corrosion current sensor uses an electrochemical sensor module with a measurement range of 0 to 100 microamps and a resolution of 0.01 microamps. It outputs an analog voltage signal, which is acquired by a 12-bit analog-to-digital converter built into the main control chip. The sampling frequency is set to 1 Hz, and it is used to evaluate the corrosion rate of metal facilities such as manhole covers and guardrails in real time.

[0185] The new sensor shares the same hardware clock source as the existing sensor, and the hardware timer of the main control chip triggers synchronous sampling in a unified manner to ensure that the data acquisition time of all sensing dimensions is consistent, and the synchronization error is still controlled within ±2 milliseconds.

[0186] Therefore, the sensing dimensions of each facility node have been expanded from the original 5 to 7, namely, the addition of air pressure value and corrosion current value.

[0187] Considering the dense facilities inside the tunnel and the potentially faster rate of anomalous evolution, the data sampling frequency was increased from 1 Hz to 2 Hz, and the time window length L was adjusted to 120, corresponding to 60 seconds of historical data (because of 2 Hz sampling, 120 points are exactly 60 seconds), ensuring that the physical time span of the input time series data remains consistent with that of Example 1.

[0188] (3) Adjustments were made to the joint modeling of the fusion graph structure and time series data in step S3.

[0189] To enhance the model's stability in the face of complex environmental noise and improve its ability to extract temporal features, the structure of the Temporal Graph Neural Network (TGNN) is adjusted as follows:

[0190] First, the spatial feature aggregation module is expanded from the original two-layer graph convolutional network (GCN) to a three-layer GCN. The output dimensions of each layer are set to 16, 32, and 64 respectively. The first GCN maps the 7-dimensional input features obtained in step S205 to 16 dimensions, the second layer maps the 16-dimensional features to 32 dimensions, and the third layer maps the 32-dimensional features to 64 dimensions. Similar to Example 1, the graph convolution operation of these three GCN layers all uses the dynamically weighted adjacency matrix calculated in step S104A. Constructed adjacency matrix Sum-degree matrix This enables each node to aggregate a wider range of spatial neighbor information and reflects the impact of dynamic coupling strength on spatial context, thereby extracting spatial context features more fully.

[0191] Secondly, the temporal evolution modeling module has been replaced by a bidirectional GRU, which consists of two independent GRU layers. The hidden state dimension of both GRU layers remains 64-dimensional. One GRU layer processes the embedded sequence obtained after spatial aggregation in forward temporal order, while the other GRU layer processes the same embedded sequence in reverse temporal order.

[0192] For each node, at each time step t, the hidden state of the forward GRU is... Hidden state of reverse GRU By concatenating the data, we obtain the final hidden state at that time step. Its dimension is 128. After L=60 time steps of iteration, the hidden state of the last time step is taken. This represents the final hidden state of the node.

[0193] In this way, we can simultaneously capture the asymptotic trend before an anomaly occurs and the brief fluctuation after it occurs, further enhancing our ability to represent time series patterns.

[0194] Since the final hidden state dimension becomes 128, in the subsequent step S4, the input dimension of the first layer of the fully connected layer is adjusted accordingly to 128, that is, the weight matrix of the first fully connected layer. The size was adjusted from 128×64 to 128×128 (the output dimension remains 128), while the second fully connected layer remained unchanged at 15×128. This adjustment allows the model to make full use of the rich temporal information after bidirectional GRU fusion.

[0195] (4) Adjustments to the dynamic threshold judgment and early warning generation in step S5.

[0196] The dynamic threshold adjustment cycle has been shortened from 24 hours to 12 hours to better adapt to the significant diurnal variations characteristic of mountainous environments, while the initial threshold τ remains set at 0.7. The false alarm statistics and threshold adjustment rules have been adjusted as follows:

[0197] If the number of false alarms exceeds 3 in the past 12 hours, the current threshold τ will be increased by 0.04.

[0198] If the number of false alarms is less than or equal to 1 in the past 12 hours, the current threshold τ will be lowered by 0.02.

[0199] If the number of false alarms is 2 or 3, the threshold remains unchanged.

[0200] The adjusted threshold remains confined to a closed interval of 0.55 to 0.85.

[0201] In addition, the early warning work order has added an environmental risk factor field to the original fields. This field integrates rainfall data (in millimeters per hour) obtained in real time from the meteorological bureau interface and geological disaster early warning levels (such as red, orange, yellow, and blue). When the work order is generated, the meteorological information of the current facility's location is automatically queried and filled into the environmental risk factor field, providing maintenance personnel with a more comprehensive basis for decision-making.

[0202] Through the aforementioned targeted adjustments, this embodiment can still maintain a high-precision facility anomaly early warning capability even in scenarios involving mountainous tunnel clusters.

[0203] Example 3

[0204] In the scenario of urban expressway interchange hubs, this embodiment optimizes the method described in Embodiment 1 to meet the actual needs of high-density facility monitoring, ultra-low latency response, and limited bandwidth resources.

[0205] (1) Adjustments to the road facility map model constructed in step S1.

[0206] This embodiment constructs a hierarchical sub-diagram based on the three-level structure of the overpass: piers, beams, and expansion joints. The specific method is as follows:

[0207] First, based on high-precision geographic information system data and design drawings, the structural unit to which each facility node belongs is marked. For example, for facilities installed on bridge piers, the bridge pier number is recorded; for facilities installed on beams, the beam number is recorded; and for facilities near expansion joints, the expansion joint number is recorded.

[0208] All facility nodes with the same pier number constitute a pier sub-drawing; all facility nodes with the same beam number constitute a beam sub-drawing; and all facility nodes with the same expansion joint number constitute an expansion joint drawing.

[0209] Within each pier subgraph, fully connected undirected edges are established between nodes, meaning that any two nodes in the pier subgraph are connected. The weight of these edges is uniformly set to 0.9 to quantify the strong physical coupling between different facilities on the same pier due to the structural integrity.

[0210] Similarly, within the beam body sub-graph and the expansion joint sub-graph, fully connected undirected edges are also established between nodes. The weight of the edges inside the beam body sub-graph is uniformly set to 0.8, and the weight of the edges inside the expansion joint sub-graph is uniformly set to 0.7. These weight values ​​reflect the differences in coupling strength between facilities within different structural units.

[0211] The connection between different bridge pier subgraphs still follows the dual rules based on proximity and functionality in Example 1. That is, if two nodes belong to different subgraphs but the physical distance is less than 10 meters, or if functional dependence is verified by data such as municipal pipelines and power supply circuits, then an edge connection is established. The weight of the edge is assigned according to the specific facility type of the node pair according to the preset rules in Example 1. This hierarchical graph model more accurately depicts the spatial organization characteristics of the overpass structure.

[0212] (2) Adjustment of multidimensional sensing time sequence data collected in real time in step S2.

[0213] The integrated multi-sensor unit adds a strain gauge array to monitor the propagation of microcracks in concrete. The strain gauge array consists of multiple resistance strain gauges, which are attached to the surface of key beams and piers. The strain is converted into a voltage signal through a Wheatstone bridge circuit.

[0214] In practice, each strain gauge is connected to a separate half-bridge circuit, using a three-wire connection to eliminate the influence of wire resistance. The bridge output is amplified by an instrumentation amplifier and then sampled by a 16-bit analog-to-digital converter built into the main control chip at a sampling frequency of 100 Hz, thereby capturing the instantaneous strain changes under high-speed vehicle impact.

[0215] Meanwhile, the edge data aggregation node uses 5G CPE equipment, which takes advantage of the low latency and high bandwidth characteristics of the 5G network to transmit the pre-processed data back to the central server with a latency of milliseconds.

[0216] In practice, each smart gateway establishes a persistent TCP connection with the central server via a 5G CPE. The data upload cycle remains once per minute, but the end-to-end transmission latency is controlled within 10 milliseconds to ensure real-time performance.

[0217] (3) Adjustments were made to the joint modeling of the fusion graph structure and time series data in step S3.

[0218] In this embodiment, the TGNN model is deployed in an edge-cloud collaborative framework to reduce the pressure on the central server and network bandwidth consumption. The first-layer graph convolutional network (GCN) is deployed and executed on the edge nodes, while the second-layer GCN and subsequent fully connected output layers are deployed on the central cloud server. Each edge node obtains two sets of information from the cloud during initialization: one is the pre-trained weight matrix of the first-layer GCN. The first is the list of IDs of its first-order neighbor nodes in the graph topology; the second is the dynamic weighted adjacency matrix. The weight submatrix corresponding to these neighboring nodes .in Step S104A calculates the dynamic weights used by the edge nodes in real time based on the latest environmental data and synchronizes them to the cloud periodically to ensure that the dynamic weights always reflect the latest coupling strength.

[0219] That is, the identifier of all nodes that are connected to this node by an edge.

[0220] During real-time operation, the edge node receives 60 seconds of time-series data collected by its own node. This data is organized into 60 time slices, each slice being a 5-dimensional vector, thus yielding a vector with dimensions of [missing information]. The matrix. At the same time, the edge node establishes real-time communication with the first-order neighbor node through the 5G device direct link, and receives the raw data (i.e., the pre-processed 5D sensing value) of the corresponding time slice of the neighbor node at each time step.

[0221] Therefore, at each time step t, the edge node possesses the feature vectors of itself and all its first-order neighbors, forming a feature vector of size . The local feature matrix, where This is the degree of this node.

[0222] Edge nodes utilize pre-trained Graph convolution is performed independently for each time slice. For time step t, a local dynamic adjacency submatrix is ​​first constructed, which directly uses the dynamic weight submatrix obtained during initialization. And based on this, the normalized Laplace is calculated. Then, according to the first-level GCN formula (using the version with a dynamic adjacency matrix, i.e.) and The 16-dimensional intermediate representation of this node at this time step is calculated. Repeat this process to obtain a 16-dimensional vector with 60 time steps.

[0223] To reduce the amount of uplink data, the edge nodes perform global max pooling on these 60 16-dimensional vectors along the time dimension. This involves taking the maximum value of each feature dimension (out of 16) over 60 time steps to obtain a 16-dimensional aggregated feature vector for each node. This vector retains the most salient spatial features within 60 seconds. Subsequently, the aggregated feature vectors of all nodes are uploaded to the central cloud server via the 5G network. Since each node only uploads 16 floating-point numbers instead of... The number of floating-point numbers reduces the amount of upstream data and significantly reduces bandwidth usage.

[0224] After receiving the 16-dimensional aggregated vectors from all nodes, the central cloud server uses them as node features as input to the second-layer GCN. The second-layer GCN also uses the pre-trained weight matrix from Example 1. The system outputs 32-dimensional features for each node. Finally, these 32-dimensional features are mapped to anomaly probability sequences for the next 15 minutes via a fully connected layer. The structure of the fully connected layer is the same as step S4 in Example 1, namely, two fully connected networks that output 15-dimensional probability values. This achieves the sharing of computational tasks and optimization of bandwidth without losing key spatial information.

[0225] (4) Adjustments to the dynamic threshold judgment and early warning generation in step S5.

[0226] This embodiment realizes the linkage between the early warning work order and the traffic signal control system. Specifically, when the system determines that a facility is under a Level 1 warning, that is, the predicted peak time advance is greater than or equal to 10 minutes, the linkage process is automatically triggered. The linkage process is implemented by calling the predefined REST API interface of the traffic signal control system. The system sends a display instruction to the variable message sign located upstream of the facility. The instruction content is in JSON format, which includes the display text "Construction ahead, drive with caution" and the duration parameter. The duration is set from the current time to the time when the warning is lifted.

[0227] At the same time, the system sends a temporary speed limit adjustment command to the signal control unit, reducing the current speed limit of the road section from the original value to 60 kilometers per hour until the warning is lifted. The basis for determining whether the warning is lifted is that the maximum value of the subsequent abnormal probability prediction sequence of the facility is lower than the current dynamic threshold for three consecutive cycles. At this time, the system automatically calls the corresponding API to cancel the display of the variable message sign and restore the original speed limit value.

[0228] All linked operations are logged and can be traced through the operation and maintenance platform.

[0229] Through the above optimizations, this embodiment can still maintain efficient and accurate anomaly warning capabilities in high-density facility scenarios such as overpasses and hubs, and achieves deep collaboration with traffic control systems.

[0230] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0231] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment includes only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for real-time early warning of facility anomalies for road operation and maintenance, characterized in that, Includes the following steps: A graph model of road facilities is constructed, which includes a set of nodes and a set of edges. Nodes correspond to facilities where sensors are deployed, and edges are generated according to the dual rules of physical proximity and functional dependence between facilities. Static weights are assigned to each edge, and dynamic correction coefficients are calculated based on real-time environmental data and the service life of facilities to obtain a dynamic weighted adjacency matrix. Real-time acquisition of multi-dimensional sensor time-series data from various facilities; preprocessing of raw data to form time window sequences; The graph model and the time window sequence of each node are input into the temporal graph neural network, which includes a spatial feature aggregation module and a temporal evolution module. The spatial feature aggregation module uses a graph convolutional network and performs time-step spatial feature extraction based on a dynamic weighted adjacency matrix to obtain the embedding vector of each node at each time step; the temporal evolution module performs temporal modeling on the embedding vector sequence of each node and outputs the final hidden state of each node. The final hidden state is input into a fully connected network and mapped to an anomaly probability sequence for multiple time steps within a future preset time window; The warning threshold is dynamically adjusted based on the number of historical false alarms, and the maximum value of the predicted anomaly probability sequence at the current moment is compared with the dynamic threshold. If the threshold is exceeded, a warning work order is generated.

2. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, The dual rules of physical proximity and functional dependency include: If the physical distance between two facility nodes is less than 10 meters, an undirected edge is established between the corresponding nodes; If the physical distance exceeds 10 meters but the municipal pipeline network topology database verifies that the two belong to the same continuous drainage pipeline network, or the power SCADA system data confirms that the two share the same power supply circuit, or the two are clearly located on both sides of the same bridge expansion joint, then an undirected edge will be forcibly established.

3. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, The dynamic correction coefficient is calculated by a multilayer sensor. The input of the multilayer sensor includes the current ambient temperature, ambient relative humidity, the average number of vehicles passing through in the past 5 minutes, and the service life of the facilities corresponding to the two nodes. The multilayer perceptron contains two hidden layers. The first hidden layer has 8 neurons and the second hidden layer has 4 neurons. The ReLU activation function is used after the hidden layers, and the Sigmoid activation function is used in the output layer. The output value is linearly transformed and mapped to the range of 0.8 to 1.2 as a dynamic correction coefficient. The dynamic weighted adjacency matrix is ​​obtained by multiplying the static weights and dynamic correction coefficients element by element.

4. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, The spatial feature aggregation module consists of two layers of graph convolutional networks. Each graph convolutional layer is calculated based on a dynamic adjacency matrix with self-loops and its corresponding degree matrix. The first layer of graph convolution maps the input features to 16 dimensions, and the second layer of graph convolution maps the 16-dimensional features to 32 dimensions. A residual connection is introduced between the two layers, which adds the output of the first layer of graph convolution to the original input after linear transformation to align the dimensions, and then uses it as the input of the second layer of graph convolution. All time steps share the same graph convolution parameters, and graph convolution operations are performed independently for each time step.

5. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, The temporal evolution module includes a multi-scale temporal convolution submodule and a gated recurrent unit submodule; The multi-scale temporal convolutional submodule uses three parallel causal convolutional layers, each with a kernel size of 3 and dilation rates of 1, 2, and 4, respectively. Each layer has 32 output channels. The outputs of the three scales are concatenated and then subjected to linear transformation and ReLU activation to obtain a 64-dimensional multi-scale fusion feature. The original embedding and the multi-scale fused features are concatenated and then input into a gated linear unit. The output of the gated linear unit is added to the original embedding through a residual connection to obtain the enhanced features. The enhanced feature sequence is input into a gated recurrent unit with a hidden state dimension of 64, and the hidden state at even time steps is added to the hidden state two time steps earlier during the iteration process to enhance long-term dependency propagation.

6. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, The preset time window is 15 minutes, and the anomaly probability sequence contains probability values ​​for 15 time steps, with each time step corresponding to the next minute. When training the fully connected network, a weighted binary cross-entropy loss function is used, and positive samples and abnormal events are given a weight of 5 times. The loss function calculates the weighted cross-entropy at each time step and takes the average value as the final loss of the sample.

7. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, Dynamically adjusting the early warning thresholds includes: The number of false alarms is counted on a 24-hour or 12-hour cycle. A false alarm refers to an event in which a warning work order has been generated and the facility status has been confirmed to be normal after manual verification. If the number of false alarms exceeds 5 or exceeds 3 within a period, the current threshold will be increased by 0.05 or 0.

04. If the number of false alarms is less than or equal to 2 or less than or equal to 1, the threshold will be lowered by 0.03 or 0.

02. The adjusted threshold is limited to a closed interval of 0.55 to 0.

85.

8. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, When acquiring multidimensional sensing time-series data in real time, all sensors are driven by the same hardware clock source, and synchronous sampling is triggered by the hardware timer of the main control chip, with the synchronization error controlled within ±2 milliseconds. Preprocessing includes moving average filtering and temperature compensation. The moving average filtering window has a length of 5 sampling points. Temperature compensation uses a polynomial model. The compensation formula is that the compensated tilt angle is equal to the original tilt angle plus the temperature compensation coefficient multiplied by the difference between the current temperature and the calibrated reference temperature. The acceleration is handled similarly.

9. The method for real-time early warning of facility anomalies for road operation and maintenance according to claim 1, characterized in that, For the scenario of mountainous highway tunnel groups, when constructing the graph model, virtual nodes of geological units are introduced, the tunnel entrance and exit are abstracted as independent nodes, and functional dependency edges are established with drainage ditches and slope anchor facilities within a radius of 30 meters, with the weight uniformly set to 0.6; For the scenario of urban expressway overpasses, a three-level hierarchical subgraph is constructed, consisting of piers, beams, and expansion joints. Nodes within the same pier subgraph are fully connected and have a weight of 0.9, nodes within the same beam subgraph have a weight of 0.8, and nodes within the same expansion joint subgraph have a weight of 0.

7. For high-density overpass scenarios, the first layer of the temporal graph neural network, graph convolution, is deployed on edge nodes, while the second layer, graph convolution, and fully connected layer are deployed on the cloud. The edge nodes perform global max pooling on the 16-dimensional vector at 60 time steps and then upload the 16-dimensional aggregated vector.

10. A real-time early warning system for facility anomalies in road maintenance, applied to the real-time early warning method for facility anomalies in road maintenance as described in any one of claims 1 to 9, characterized in that, include: The graph model building module is used to construct nodes and edges based on facility registration information, geographic coordinates and functional dependencies, and introduces a dynamic weight correction mechanism to generate a dynamic weighted adjacency matrix. The data acquisition and preprocessing module is deployed at each facility node and includes hardware-synchronized multi-sensor units and edge preprocessing units, which are used to acquire multi-dimensional sensor data in real time and form a time window sequence. The temporal graph neural network processing module, deployed on a central server or edge-cloud collaborative architecture, is used to perform spatial feature aggregation and temporal evolution modeling, and output the final hidden state of each node. The anomaly probability prediction module is used to map the final hidden state into a future anomaly probability sequence via a fully connected network; The dynamic threshold early warning module is used to dynamically adjust the threshold based on the number of historical false alarms, compare the predicted probability with the threshold, and trigger the generation of early warning work orders and hierarchical response. The modules are interconnected through a communication network to enable real-time early warning of road facility anomalies.

Images