Machine Learning-Based Highway Engineering Disaster Monitoring and Early Warning System

Abstract

The present invention discloses a highway engineering disaster monitoring and early warning system based on machine learning, which relates to the technical field of engineering safety monitoring and early warning. A complete solution from multi-source sensor deployment to data fusion, deep learning prediction, and automated emergency linkage is proposed for highway engineering disasters caused by extreme rainfall and multi-disaster linkage. In step one, an adaptive siting index is used to optimize sensor deployment to enhance monitoring coverage and communication robustness; in step two, the original observations are clustered, interpolated, and anomaly-checked to obtain reliable multi-source fusion data; in step three, a deep learning model is constructed based on time series or graph network to output landslide risks and dynamically adjust thresholds; in step four, urgency and emergency clustering degree are used to achieve rapid scheduling of multiple departments; in step five, it is further extended to disaster monitoring such as flood and bridge health to form a collaborative prevention and control system for all disasters, demonstrating high reliability and strong collaborative operation capabilities.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering safety monitoring and early warning, and specifically to a highway engineering disaster monitoring and early warning system based on machine learning. Background Art

[0002] In modern highway engineering, due to the highly variable natural terrain and climate characteristics in mountainous areas, hilly areas and along river sections, environmental factors such as extreme rainfall, seasonal freeze-thaw, and river water level fluctuations are extremely likely to induce disasters such as slope collapses, debris flows, subgrade settlements, and even bridge structure damages. A typical scenario is that in a high mountain highway cliff section during continuous heavy rain, the rainwater rapidly infiltrates along the rock fissures, the soil moisture content and pore water pressure suddenly increase, which not only weakens the shear strength of the slope, but also easily causes the rock mass to loosen or slip rapidly.

[0003] In addition, when the surface runoff increases sharply due to a sudden increase in rainfall, the drainage ditches, culverts and small bridges along the line may also be blocked or their structures impacted, further exacerbating the instability of the subgrade. Moreover, if the freeze-thaw cycle is superimposed with heavy rainfall, the softening of the soil, crack expansion, and pavement damage will show an accelerating cumulative trend; once short-term heavy precipitation or continuous rainy days are encountered, disasters may break out on a large scale in a short time, posing a serious threat to road traffic safety and infrastructure stability. In view of this, a large number of highway projects have deployed a variety of sensors such as rain gauges, microseismic instruments, and tilt monitoring devices in high-risk sections, hoping to reduce disaster risks through real-time capture of environmental parameters, data fusion and predictive analysis; but still need to face technical challenges such as discrete distribution of multiple sensors, huge data flow and easy packet loss, and cross-coupling of multiple disaster types.

[0004] Currently, in the field application of mountain highway engineering, there are still core problems such as insufficient mining depth of multi-source heterogeneous monitoring data and inability to comprehensively consider anomaly capture and real-time forecasting, resulting in difficulty in timely or accurately identifying key disaster signs. One particularly prominent technical problem is that in extreme heavy rain conditions, the monitoring network often experiences sensor data loss or distortion (such as sudden changes in rain and moisture content readings, and sharp increase in microseismic waveform interference). In addition, the timing of slope failures under different geological conditions is not consistent. If early warnings are simply based on single sensors or fixed rainfall thresholds, it is easy to result in high false alarm or missed alarm rates, and there is a lack of comprehensive judgment on the co-evolution of multiple disaster types.

[0005] Specifically, when heavy rain causes attenuation of remote LoRa or 5G signals and the front-end sensors cannot upload data completely, if the back-end system does not combine multi-source data for anomaly screening and complementation, it may misjudge the sudden signs of mountain landslides as communication failures or noises; on the other hand, if there is a lack of linkage monitoring of surface runoff, bridge strain or changes in the depth of frozen soil layers, it is difficult to identify in real time the interrelated disaster processes such as floods and freeze-thaws, thus missing the early warning opportunity or making improper scheduling, which may cause large-area blockages, interruptions and serious safety hazards on the road.

[0006] In summary, this technical problem brings potential risks to the highway disaster monitoring system, such as unstable data collection, insufficient information screening, and single risk determination. Once a disaster breaks out under severe weather and complex geological conditions, it is often difficult for the emergency department to take the optimal response measures in a timely manner.

[0007] Therefore, the present invention provides a highway engineering disaster monitoring and early warning system based on machine learning. Summary of the Invention

[0008] (I) Technical Problem to be Solved

[0009] In view of the deficiencies of the prior art, the present invention provides a highway engineering disaster monitoring and early warning system based on machine learning. By proposing a complete solution from multi-source sensor layout to data fusion, deep learning prediction, and automated emergency linkage, in step one, an adaptive siting index is used to optimize sensor deployment to enhance monitoring coverage and communication robustness; in step two, clustering, interpolation, and anomaly verification are performed on the original observations to obtain reliable multi-source fusion data; in step three, a deep learning model is constructed based on time series or graph network to output landslide risk and dynamically adjust the threshold; in step four, urgency and emergency clustering degree are used to achieve rapid scheduling of multiple departments; in step five, it is further extended to disaster monitoring such as flood and bridge health to form a collaborative prevention and control system for all disaster types, thus solving the technical problems described in the background art.

[0010] (II) Technical Solution

[0011] To achieve the above objectives, the present invention is realized through the following technical solutions: A highway engineering disaster monitoring and early warning system based on machine learning, including,

[0012] When extreme rainfall or geological monitoring requirements are triggered, according to the sensor siting index Differentially deploy various devices such as rain gauges, moisture content sensors, tilt monitors, and microseismic instruments, and perform preliminary data collection and timestamp calibration through networks such as LoRa or 5G, so that each monitoring node can maintain stable power supply and communication in severe weather, and integrate multi-source observed values after filtering and format conversion into an original data frame And then send it to the edge node;

[0013] For the original data frame Trigger clustering or dimensionality reduction operations to eliminate obvious invalid noise points, and use a filling algorithm to correct abnormal blank values. Subsequently, establish a high-speed screening mechanism to monitor the changes of the original data under the set threshold. If it is detected that the amplitude anomaly exceeds the allowable range, automatically call the backup sensor or perform secondary verification, and hand over the filtered and hierarchically marked multi-source fusion vector To the deep learning model for processing;

[0014] The deep learning engine is activated during high-risk periods, using multi-source fusion vectors as inputs and combining with a time series network (LSTM) or a graph structure network (GNN) to extract spatial and temporal correlation features, and using an adaptive mechanism to dynamically correct the decision threshold and constructing the landslide risk level , performing self-learning iteration on key variables under uncertain conditions, and feeding back the obtained risk assessment results together with the latest threshold iteration records to the emergency end;

[0015] When the landslide risk level exceeds the adaptive threshold , the warning mechanism immediately broadcasts an alarm to the road management platform via text messages, warning signs or in-vehicle terminals, and based on the notification urgency , the on-site status confirmation value and the emergency cluster level perform a quick sort, integrate the resources of multiple departments to trigger the entry and closure instructions of the rescue team or large equipment, and efficiently complete the emergency handling and subsequent information archiving within the closed-loop logic of high-risk sections - multi-channel warning - on-site verification - hierarchical control;

[0016] If the disaster types expand to situations such as floods, bridge structural health or freeze-thaw damage, additional monitoring equipment such as water level gauges, strain gauges or surface temperature sensors and data fusion strategies are added, and multi-task learning is used to fuse the risks of multiple disaster types to generate a combined risk level , and then automatically dispatch drainage or reinforcement resources through the emergency cluster level to achieve integrated monitoring and coordinated handling of multiple disaster types across the entire highway.

[0017] Preferably, a sensor location selection index is defined to quantify the rationality of the layout of the th sensor at a specific location ; after obtaining the sensor location selection indices of all candidate locations, the limited sensor resources are preferentially allocated to the areas with higher sensor location selection indices , and then a sensor layout plan is obtained;

[0018] Preferably, based on the completed sensor layout plan, real-time data collection and preliminary synchronization of the sensors are carried out; comparing with the determined location list , acquisition devices are deployed at each sensor location, and the obtained measurement values are combined to form an original data frame , each original data frame is attached with a location identifier and its location selection index , and the data uploaded by multiple sensors is preprocessed;

[0019] Preferably, the sensors are read separately at time the original data frames collected are merged to form a multi-source fusion vector :

[0020]

[0021] wherein: represents the monitoring value of the rain gauge at the position of the sensor ; represents the measured water content of the water content sensor at the same position; represents the surface displacement or angle recorded by the inclinometer; represents the vibration amplitude captured by the seismograph; is the site selection index;

[0022] The generalized weighted mapping formula is introduced to perform spatial fusion on multi-source data at adjacent positions, and finally a spatial difference mapping matrix can be formed:

[0023]

[0024] wherein: and are the multi-source fusion vectors of positions and respectively;

[0025] represents the monitoring difference degree of the two sensors at the same moment ; and represent the coordinate positions of the two sensors; and are fusion balance coefficients, and the values are greater than 0; is the attenuation factor;

[0026] Preferably, the spatial difference mapping matrix is used to identify significant conflicts: If the data of a certain sensor deviates greatly from that of most adjacent sensors, and the duration of the deviation exceeds the threshold , it is temporarily listed as a suspected anomaly;

[0027] If the mutation amplitude of the sensor exceeds the preset jump threshold within a short period of time, and it is judged to be extremely unreasonable after being evaluated in combination with the site selection index , the secondary reset verification mechanism of the device is triggered, or it is directly switched to the standby sensor in the same area;

[0028] To more accurately quantify whether the sensor readings constitute an anomaly, the following cross-measurement factor is defined :

[0029]

[0030] Wherein: is a multi-source fusion vector; is the reference prediction vector of the multi-source fusion vector of the sensor under normal working conditions;

[0031] represents the multi-dimensional difference between the current and the reference; is the sensor location index;

[0032] , is the weight of the cross-measurement factor, is the attenuation factor, is the integration variable;

[0033] When the cross-measurement factor is continuously lower than the set threshold for multiple times, it will be determined that there is a major abnormality in the sensor device or data, and it will be recorded in the list of suspected faults: at the same time, the standby device will be automatically called or a request for on-site manual inspection will be initiated;

[0034] If it is found that the readings of multiple surrounding sensors are abnormal at the same time, they will be included in the list of high-risk status;

[0035] Preferably, multiple monitoring points at time All the multi-source fusion vectors form a multi-source time series vector , and a preliminary construction of landslide risk prediction is carried out to align the feature order in the multi-source time series vector with the nodes of the model input layer;

[0036] The multi-source time series vector is divided into blocks by time period to form several training / prediction sequences in the form of sliding windows;

[0037] Before each batch is loaded into the prediction model, the high-risk status list and the suspected fault list are checked. If a certain sensor is in a suspected fault state, the data channel of the sensor will be temporarily masked from the input vector or the standby sensor channel will be used instead.

[0038] Preferably, forward propagation is performed batch by batch on the input multi-source time series vector to obtain the landslide risk output ;

[0039] If there are multiple sensor channels at the same location calculate the landslide risk degree ;

[0040] Preferably, if the landslide risk degree global risk threshold and last for a period of time or at critical positions;

[0041] Mark the corresponding sensors as being in a high-risk state. If multiple adjacent sensors exceed the threshold together, high-risk road segments or high-risk point sets can be generated through merging strategies or road segment aggregation markings, and the corresponding sensors or nodes can be summarized as a high-risk sensor list, and the threshold can be adaptively corrected;

[0042] Combined with the landslide risk level and known recent real feedback, partially online train the deep learning model parameters intermittently; when a new extreme rainfall situation occurs at a certain location and the model lacks relevant samples, the multi-source fusion vector and the landslide risk level can be stored in the extreme sample library.

[0043] Preferably, based on the generated high-risk sensor list and the corresponding landslide risk level , simultaneously read the system global threshold ;

[0044] For road segments where the sensors meet the landslide risk level global risk threshold , automatically send an alarm message to the road management platform and push it to the monitoring center operators; if multiple adjacent sensors exceed the threshold simultaneously, aggregate them and mark them as the same high-risk road segment;

[0045] Define the urgency index of the notification , which is used to sort multiple high-risk road segments. The specific calculation is as follows:

[0046]

[0047] In the formula: is the landslide risk level, is the current adaptive threshold, represents the distance between this location and the important target location H; is the integration variable; , is the balance coefficient, is the exponential decay factor;

[0048] Notification urgency The larger the value, it means that the risk value of this road segment exceeds the threshold by a large margin and is relatively closer to the key traffic nodes, and it is necessary to give priority to alarm and arrange emergency resources for attention;

[0049] Preferably, according to the notification urgency The list of high-priority road sections is screened, and the on-site immediate status is obtained. If obvious abnormalities are identified in the UAV images or camera images, a list of confirmed high-risk locations is constructed;

[0050] Then, immediately compare this verification result with the multi-source fusion vector data. To quantify the consistency between the image inspection and the dynamic state of the soil body, the following on-site status confirmation value is defined :

[0051]

[0052] In the formula: is the image matching degree; represents the measured value of the rainwater or sediment flow of this road section; , is the weighting coefficient;

[0053] When the on-site status confirmation value exceeds the pre-set status threshold, an emergency linkage instruction is triggered;

[0054] Preferably, based on the list of confirmed high-risk locations, if the on-site status confirmation value exceeds the pre-set status threshold and the notification urgency ranks among the top of all high-risk sections, automatically recommend closing the corresponding road section and notify the maintenance and emergency rescue departments;

[0055] Based on the notification urgency rank and schedule the priorities of multiple risks, allocate the corresponding rescue materials and personnel to the site in batches, and construct the emergency cluster degree ;

[0056] According to the emergency cluster degree evaluate the multi-department cooperation effect in real time. If it continues to decrease, issue a coordination warning;

[0057] When the critical road section is successfully de-risked or the landslide risk degree drops back below the safe range again, automatically notify all parties to end the emergency state, and start model re-training, threshold optimization, and sensor status tracking;

[0058] Preferably, install monitoring equipment for other highway disaster types, merge the newly added multi-disaster observation vectors into the model input structure, and construct multi-task learning or multi-model integration strategies for different disaster types:

[0059] Set dedicated sub-networks for flood, bridge health, and freeze-thaw in the time-series network or graph-structure network respectively, or use a high-order fusion layer to realize the joint representation of multi-disaster features;

[0060] Preferably, based on the spatio-temporal prediction of landslide risk and the emergency response mechanism for over-threshold sections, the synchronous determination of multiple risk parameters is introduced; where:

[0061] If it is defined that represents the comprehensive flood risk output of section , represents the bridge damage risk output, represents the freeze-thaw risk, together with the original landslide risk to obtain the combined risk degree by merging :

[0062]

[0063] In the formula: is the set of disaster types covered;

[0064] represents the risk value of the disaster type; represents the adaptive threshold for the disaster type ; and are the weighting and attenuation coefficients respectively, , is the integration variable;

[0065] The combined risk degree The larger the value, the higher the comprehensive risk of the section under multi-disaster conditions. When scheduling resources, richer emergency measures can be automatically incorporated to achieve the goal of coordinated disposal of multi-disasters;

[0066] When the combined risk degree exceeds the preset global threshold or the thresholds of each disaster type reach the warning standard, grading disposal is carried out with reference to the notification urgency and the emergency cluster degree to call different departments or equipment for different disaster types.

[0067] (III) Beneficial effects

[0068] The present invention provides a highway engineering disaster monitoring and early warning system based on machine learning, having the following beneficial effects:

[0069] By introducing a layout strategy with an adaptive site selection index and an Internet of Things networking method, not only can each sensor (such as a rain gauge, a soil moisture sensor, a microseismograph, etc.) be flexibly configured and complementary in a rainstorm environment, but also wide-coverage and highly reliable data transmission can be achieved through a low-power wide-area network or a 5G network, thus avoiding the problems of single arrangement of traditional monitoring points and one-sided data. This multi-source deployment can significantly improve the monitoring integrity under extreme climates;

[0070] Merge multi-source raw data into multi-source fusion vectors , and then use anomaly screening and gap filling algorithms to ensure data quality, thereby minimizing misjudgments caused by false positives or noise interference; combined with preprocessing mechanisms and redundant verification logic, subsequent deep learning models do not need to consume a lot of resources to process invalid data;

[0071] The landslide risk is calculated by hybrid model and supplemented by threshold value. The adaptive adjustment dynamically changes the warning threshold according to extreme rainfall and real-time feedback data, which not only reduces the rigidity of fixed thresholds, but also allows the model to continuously learn new samples. The coordinated analysis of microseismic, tilt and water content channels greatly reduces the overall risk of false alarms caused by the inaccuracy of a single measuring point.

[0072] exist When multiple departments are coordinated, the urgency of the notification is used to trigger , On-site status confirmation value Emergency Clustering A series of indexes, such as road section risk, on-site verification and departmental coordination capabilities, are integrated into the quantitative evaluation system, which greatly improves the resource allocation efficiency of commanders in the case of concurrent dangers on multiple road sections and the emergency clustering degree. It can dynamically depict the effectiveness of multi-party linkage and realize three-dimensional handling from a single disaster to multi-link coordination;

[0073] Expanding this system from focusing on landslide risks to floods, bridge structure health, freeze-thaw and other disasters can achieve efficient iteration and real-time response for coordinated prevention of multiple types of disasters, fully demonstrating the synergy between technical features;

[0074] In summary, its significant beneficial effects include: significantly improving the flexibility and credibility of highway disaster monitoring under extreme climate conditions, reducing missed warnings and false alarms, enhancing the efficiency of multi-departmental coordination, and improving the overall safety protection level of highways by covering multiple disaster types, demonstrating a high degree of reliability and strong collaborative operation capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 It is a schematic diagram of the highway engineering disaster monitoring and early warning process based on machine learning in the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0077] Please refer to Figure 1 , the present invention provides a highway engineering disaster monitoring and early warning system based on machine learning, including,

[0078] Step 1: When the need for extreme rainfall or geological monitoring is triggered, according to the sensor siting index Differentially deploy various devices such as rain gauges, moisture content sensors, tilt monitors, and microseismic instruments, and perform preliminary data collection and timestamp calibration through networks such as LoRa or 5G, so that each monitoring node can maintain stable power supply and communication under bad weather, and integrate multi-source observed values after filtering and format conversion into an original data frame and then send it to the edge node;

[0079] The content of the said Step 1 includes the following:

[0080] Step 101: Sensor deployment strategy and initial deployment algorithm

[0081] By establishing a unified sensor deployment strategy, ensure reasonable coverage in key areas of mountain slopes, mainly including two parts: siting planning and sensor density allocation:

[0082] To characterize the coverage quality and stability of sensors in the slope area, define the sensor siting index , used to quantify the th sensor at a specific location The rationality of the layout at the location, in order to consider the distance between the sensor and the base station, environmental load, and local gradient information during the siting process, define the sensor siting index , as follows:

[0083]

[0084] represents the candidate location of the sensor, which can be regarded as in the geographical coordinates; is the location of the base station or the main data aggregation node, which functions as a fixed coordinate determined by the communication network planning; represents the candidate point and the base station the Euclidean distance between them;

[0085] represents to the gradient vector of the coordinate, and its norm represents the rate of change of the distance function with the position fine-tuning near ; represents the candidate point the surrounding environmental load index (such as the comprehensive quantization value of average annual rainfall, geological looseness, etc.);

[0086] represents the gradient of coordinates indicating the rate of change of the environmental load in a local area; , is the balance coefficient which is used to perform weighting between communication convenience priority and high-risk environment priority 、 is the attenuation factor, which is used to control the attenuation speed of the exponential term , is the integration variable ;

[0087] After obtaining the sensor siting indices of all candidate locations the limited sensor resources are preferentially allocated to the areas with higher sensor siting indices At the same time, to avoid over-concentration or over- sparsity in the area, a minimum / maximum neighborhood limit can be superimposed subsequently to ensure effective sensing coverage between sensors and avoid self-interference;

[0088] When the sensor siting indices of multiple adjacent points in a certain area are all very high, by performing secondary screening on this area and moderately relaxing the distance constraint to avoid losing spatial diversity due to over-clustering. When the sensor siting indices of a certain area are generally low, but it still has important monitoring significance based on the terrain and soil layer structure, manual intervention deployment can be carried out based on expert experience or geological prior knowledge to ensure the integrity of the overall coverage, and then obtain the sensor layout plan;

[0089] When in use, the layout efficiency and adaptive ability of the sensors are improved, so that the layout results achieve an overall balance in both earthquake-prone areas and communication reliable areas, which can greatly reduce sensor redundancy or blind spots and improve the overall efficiency of subsequent real-time data collection.

[0090] Step 102, Multi-source data collection and preliminary synchronization

[0091] Based on the completed sensor layout plan, real-time data collection and preliminary synchronization (preliminary filtering, time calibration, etc.) of the sensors are carried out; referring to the determined location list , a variety of data collection devices such as rain gauges, moisture content sensors, tilt monitors, and microseismic instruments are deployed at each sensor location: using low-power wide area networks (such as LoRa, NB-loT) or 5G systems, the measured values such as rainfall, soil moisture content, surface displacement, and microseismic amplitude are combined to form the original data frame , where is the device number, is the acquisition time, and each original data frame is attached with a location identifier and its siting index ;

[0092] Unify the data uploaded by multiple sensors to the same timestamp format, and perform simple upper and lower limit truncation filtering on the original data frames at the edge node, discard extremely unreasonable values, and complete preprocessing. If some fields in a data frame are missing or have minor anomalies, they will be cached and marked first; During use, directly bind the generated layout coordinates and sensor feature information to the real-time data frame to ensure that data can be classified and preferentially processed quickly according to the sensor location and siting index during subsequent fusion analysis; With the help of the data frame identification mechanism, the detailed environmental information of the slope area where a certain sensor is located can be quickly located during extreme rainfall, providing an accurate geographical reference for subsequent triggering of emergency strategies.

[0093] Step 2: For the original data frames

[0094] Trigger clustering or dimensionality reduction operations to eliminate obvious invalid noise points, and use the filling algorithm to correct abnormal blank values. Subsequently, establish a high-speed screening mechanism to monitor the changes of the original data under the set threshold. If the detected amplitude anomaly exceeds the allowable range, automatically call the backup sensor or perform secondary verification, and hand over the filtered and hierarchically marked multi-source fusion vector to the deep learning model for processing;

[0095] The said Step 2 includes the following contents:

[0096] Step 201: Multi-source data fusion and unified feature mapping

[0097] According to various original data frames and their attached location information, adopt a multi-source fusion strategy to perform unified feature mapping on data such as rainfall, soil moisture content, surface displacement, and microseismic amplitude, and output the called multi-source time series vector;

[0098] Read the sensors respectivelyat time the original data frames collected , first align them on the time axis based on sensors of the same type, and perform primary merging on environmental acquisition data such as rainfall vectors, moisture content vectors, surface tilt vectors, and microseismic vectors according to the unified timestamp to form a unified multi-source fusion vector :

[0099]

[0100] In the formula: represents the rain gauge at the sensor Monitoring value of the position; Indicates the water content measured by the water content sensor at the same position; Indicates the surface displacement or angle recorded by the inclinometer; Indicates the vibration amplitude captured by the microseismograph; Is the site selection index;

[0101] By mapping each original data into a unified fusion vector , subsequent analysis can directly perform operations and clustering on sensor information in the same dimension, without having to manage multiple heterogeneous data tables separately; after the conventional data merging is completed, considering that the observation intervals of each sensor may overlap or have gaps, the following generalized weighted mapping formula is introduced Perform spatial fusion on multi-source data at adjacent positions, and finally a spatial difference mapping matrix can be formed:

[0102]

[0103] In the formula: and Are the multi-source fusion vectors of positions and respectively; Indicates the monitoring difference degree of the two sensors at the same moment ; and Indicate the coordinate positions of the two sensors; and Are the fusion balance coefficients, with values greater than 0; Is the attenuation factor, with a value greater than 0, which determines the penalty strength for large difference values;

[0104] When in use, by introducing the generalized weighted mapping formula , not only the multi-sensor values at the same position are integrated in one dimension, but also the spatial fusion based on exponential decay of the data at adjacent positions is allowed.

[0105] Step 202, Abnormality screening and high-risk data marking

[0106] Use the spatial difference mapping matrix (or directly compare the multi-source fusion vector with the multi-source fusion vectors around it) to identify significant conflicts: if the data of a certain sensor deviates significantly from the data of most adjacent sensors in terms of rainfall, water content, microseismic amplitude, etc., and the duration of the deviation exceeds the threshold , then it is temporarily listed as a suspected abnormality;

[0107] For the jumps of a single sensor itself at the time series level, a recursive weighted curve can be defined. The observed data (or the feature values after preprocessing) arriving at each new moment are superimposed, averaged, or exponentially decay-accumulated with the weighted result of the previous moment according to the set weights, forming a reference trajectory that is continuously iteratively updated over time. On this trajectory, the influence of recent data points may be given greater weights, while historical points are gradually weakened or smoothed;

[0108] If the mutation amplitude of this sensor exceeds the preset jump threshold within a short period and is judged to be extremely unreasonable after being combined with the siting index for evaluation, the secondary reset verification mechanism of this device is triggered, or it is directly switched to a standby sensor in the same area;

[0109] To more accurately quantify whether the sensor readings constitute an anomaly, the following cross-measurement factor is defined :

[0110]

[0111] In the formula: is the multi-source fusion vector; is the reference prediction vector of the multi-source fusion vector of this sensor under normal working conditions; represents the multi-dimensional difference between the current and the reference; is the sensor siting index; , is the weight of this cross-measurement factor, and the values are all greater than 0. is the attenuation factor, and the value is greater than 0, which determines the sensitivity of the numerical difference in this anomaly metric. is the integration variable;

[0112] If the siting index of this sensor is relatively high, indicating that its installation location itself is in a high-risk area, then a certain numerical fluctuation can be tolerated when evaluating the anomaly, so as to distinguish the real environmental abrupt change caused by heavy rainfall from the anomaly caused by equipment failure;

[0113] When the cross-measurement factor is continuously lower than the set threshold for multiple times, it will be determined that there is a major anomaly in this sensor device or data, and it will be recorded in the list of suspected faults: at the same time, the standby device will be automatically called or a request for on-site manual inspection will be initiated;

[0114] If it is found that the readings of multiple surrounding sensors are abnormal at the same time, but the fusion analysis shows that this may be due to the real environmental change caused by a sudden rainstorm, then it will not be judged as a failure, but will be included in the high-risk status list;

[0115] In use, by taking into account both spatial comparison and time series cross-measurement, the false alarm and misreport rates are significantly reduced, effectively distinguishing true landslide precursor signals from pure sensor anomalies; introducing a site selection index as an auxiliary indicator, such that in areas with high site selection value, the absolute judgment threshold for numerical fluctuations can be appropriately relaxed, avoiding true large-scale changes in extreme rainfall environments from being simply classified as equipment failures. A cross-measurement factor is proposed, and then the reference prediction vector is organically combined with the sensor site selection index

[0116] fully considering the importance of geographical location and the dynamics of numerical differences. Step Three: The deep learning engine is activated during high-risk periods, using the multi-source fusion vector as the input and combining with a time series network (LSTM) or a graph structure network (GNN) to extract spatial and temporal correlation features, using an adaptive mechanism to dynamically correct the decision threshold and constructing a landslide risk degree

[0117] The said Step Three includes the following contents:

[0118] Step 301: Deep learning model construction and data input

[0119] All multi-source fusion vectors of multiple monitoring points at time constitute a multi-source time series vector which contains monitoring features such as rainfall, soil water content, surface displacement, and microseismic amplitude, and undergoes multi-source fusion and anomaly rejection; and a time series deep learning model (such as LSTM) or a spatial correlation model (such as GNN) is selected for the preliminary construction of landslide risk prediction: among them,

[0120] if geographical adjacency relationships are emphasized, then GNN is focused on; if time series changes are emphasized, then LSTM is focused on, or a hybrid structure can be achieved by combining the two. Register the corresponding relationships of each data field in the prediction model configuration, making the feature order in the multi-source time series vector strictly aligned with the nodes of the model input layer, and retain the fusion mapping dimension description in the meta-information; divide the multi-source time series vector into blocks according to time periods to form several training / prediction sequences in the form of sliding windows, such as: ;

[0121] Before each batch is loaded into the prediction model, check the high-risk status list and the list of suspected faults. If a sensor is in a suspected fault state, temporarily mask the data channel from the input vector or switch to an alternative sensor channel.

[0122] During use, by setting a sliding window and handling sensors with suspected faults, the integrity of the data in time and space can be maximally preserved, preventing abnormal nodes from contaminating the model prediction results and enabling the model to have the ability to recognize spatial graph structures or long-sequence dependencies. Design an automatic mechanism for masking channels with suspected faults in the input batch management, so that the model has the ability to adaptively exclude abnormal devices, which is more flexible and stable than the single filtering of abnormal values by traditional time-series models.

[0123] Step 302, Landslide risk determination and dynamic probability output

[0124] Perform forward propagation batch by batch on the input multi-source time series vector to obtain the landslide risk output ;

[0125] The landslide risk output represents the predicted landslide probability value of the area where the sensor is located, and the numerical range is set at . If a hybrid model is used, the LSTM can be used to fit the time trend and the GNN to fit the spatial dependence, and the intermediate layers of the two are merged and output at the back end of the network ; To make the prediction more sensitive to the non-linear response caused by extreme rainfall, an attention mechanism can be introduced into the model to give higher attention to fields such as sudden changes in rainfall intensity and soil moisture content;

[0126] If there are multiple sensor channels (such as an inclinometer and a microseismograph) at the same location , the landslide risk degree can be further calculated through the following aggregation formula : Let the predicted landslide probability values of the three channels at the same location at time be respectively:

[0127]

[0128] They correspond to the predicted outputs of the microseismograph (MI), inclinometer (TI), and other comprehensive information (GI, such as the combination of moisture content and rainfall) channels. Summarize these three channels into a column vector form:

[0129]

[0130] To characterize the differences between channels, a channel difference vector can be defined:

[0131]

[0132] The norm of the difference vector can be used to measure the deviation degree of the prediction results between the microseismograph / tilt monitor and the integrated information channel at the same location; based on the above vector definition, a new landslide risk degree is introduced to measure the overall landslide risk of the location as follows:

[0133]

[0134] In the formula: is the multi-channel prediction output, which combines the landslide probabilities of different channels at the same location into a vector form with a dimension of , and each component , , are all within the interval;

[0135] is the constructed auxiliary difference vector, , is the fusion balance coefficient, , is the attenuation factor, which controls the exponential penalty strength for the channel difference , , and is the integration variable; is the weight vector, is the column vector of , ;

[0136] When in use, through forward prediction of the deep network + multi-channel weight aggregation, the overall landslide risk degree for each location is finally obtained, which is more robust compared with the traditional single prediction value; it enables the high-dimensional sensor data to be comprehensively integrated in three dimensions of space, time and channels, and ensures that the non-linear coupling effect of extreme rainfall scenarios can also be effectively captured.

[0137] Step 303, Threshold Adaptive Update and Model Iteration

[0138] Denote as the global risk threshold at the current moment, and its range is usually in ;

[0139] If the landslide risk degree is greater than the global risk threshold If it persists for a period of time or appears at critical locations (i.e., high-risk states), it is considered that the landslide risk has entered the warning state. The area with high landslide risk is listed, a risk area list is constructed, and the details of the warning trigger are recorded to retrospectively analyze whether the alarm is excessive or insufficient in the next model iteration and further correct the threshold. and network parameters;

[0140] Mark the corresponding sensors as high-risk states. If multiple adjacent sensors exceed the threshold simultaneously, high-risk road segments or high-risk point sets can be generated through merging strategies or road segment aggregation markings, and the corresponding sensors or nodes are summarized as a high-risk sensor list;

[0141] With the continuous emergence of extreme rainfall samples, if it is found that the conventional threshold triggers too many false alarms or missed alarms, based on the distribution of recent monitoring data and the actual occurrence of landslides, the threshold is adaptively corrected. For example, an adaptive update equation can be defined:

[0142]

[0143] In the formula: represents the judgment error metric obtained by comparing the output of the model with real landslide events at time ; is the preset allowable error limit, is the adjustment step size; The function is used to determine the direction of increasing or decreasing the threshold;

[0144] Combined with the landslide risk degree and the known recent real feedback (i.e., whether a landslide actually occurred), the parameters of the deep learning model can be partially online trained at regular intervals, that is, intermittently.

[0145] When a new extreme rainfall situation appears at a certain location and the model lacks relevant samples, the multi-source fusion vector and the landslide risk degree can be stored in the extreme sample library, and the fitting ability for this type of data can be enhanced during the next batch training;

[0146] When in use, through the threshold adaptive strategy, it can be dynamically corrected according to recent real landslide cases and model performance, and continuously maintain good judgment accuracy in uncertain environments such as extreme rainfall.

[0147] Double adaptive adjustment is made to the landslide risk threshold and model parameters, which can not only cope with the randomness of the natural environment, but also continuously expand the coverage of the model for extreme working conditions. The concept of an extreme sample library is introduced to highlight the key learning of rainfall / landslide conditions far beyond the normal range, meeting the rapid adaptation requirements for rare and unconventional meteorological events.

[0148] Step Four: When the landslide risk degree exceeds the adaptive threshold , the warning mechanism immediately broadcasts an alarm to the road management platform via text messages, warning signs or in-vehicle terminals, and based on the notification urgency , the on-site status confirmation value and the emergency cluster degree , a quick sort is performed to integrate the resources of multiple departments to trigger the entry and closure instructions of the emergency rescue team or large equipment, and the danger situation is efficiently disposed of and the subsequent information is archived within the closed-loop logic of high-risk sections - multi-channel warning - on-site verification - hierarchical control;

[0149] The content of the above Step Four includes the following:

[0150] Step 401: Risk broadcast and priority classification

[0151] Based on the generated high-risk sensor list (including the location ) and the corresponding landslide risk degree , the system global threshold is read simultaneously to determine which sections or nodes have reached the warning conditions;

[0152] For the sections where the sensors meet the landslide risk degree global risk threshold , an alarm message is automatically sent to the road management platform and pushed to the operators of the monitoring center, and a notice is simultaneously issued through multiple channels such as text messages, warning sign updates, and in-vehicle terminal pop-ups; if multiple adjacent sensors exceed the threshold simultaneously, they are aggregated and marked as the same high-risk section to avoid confusion caused by repeated sending of instructions;

[0153] To quantify the notification priorities of different sections, the urgency index of the notice is defined for sorting multiple high-risk sections, and the specific calculation is as follows:

[0154]

[0155] In the formula: is the landslide risk degree, is the current adaptive threshold, represents the difference between the risk degree and the corresponding threshold, represents the distance of this location from the important target location H; is the integration variable;

[0156] , is the balance coefficient, and its value is greater than or equal to 0. is the exponential decay factor, and its value is greater than 0;

[0157] Notification urgency The larger the value, the greater the amplitude by which the risk value of the road section exceeds the threshold and the relatively closer the distance to the key traffic nodes. It is necessary to give priority to warnings and arrange emergency resources for attention;

[0158] When in use, through the notification urgency , quantitative priority arrangements can be made for multiple over-threshold road sections, preventing information overload or chaos caused by broadcasting all warnings at the same time. Not only the risk value itself is concerned, but also the distance dimension between the road section and the important area is considered, so that the high-risk sections closer to the urban area or the hub can be prioritized for the response of the management department.

[0159] Step 402, Automatic verification and on-site status confirmation

[0160] According to the list of high-priority road sections screened by the notification urgency , automatic triggering of drone remote sensing inspections or image enhancement processing of roadside cameras for high-priority road sections to quickly obtain the on-site instant status;

[0161] If obvious abnormalities are identified in the drone images or camera, such as soil deformation, collapse traces, or a significant increase in the detected rainflow scouring volume, the location identifiers of relevant monitoring points or adjacent road sections , landslide risk degree , on-site image matching degree and traffic anomalies and other information are summarized to construct a list of confirmed high-risk locations;

[0162] Then immediately compare this verification result with the multi-source fusion vector data to further confirm the authenticity of the landslide risk;

[0163] When video or image acquisition is restricted by extreme weather, manual methods (rescue teams, maintenance personnel) can be used to quickly rush to the scene for actual investigation, and all newly collected environmental information is then transmitted back to the central server to provide a basis for model iteration or emergency command decision-making; To quantify the consistency between image inspection and soil dynamics, the following on-site status confirmation value is defined:

[0164]

[0165] In the formula: is the image matching degree (i.e., the difference between the soil displacement features detected in UAV or fixed camera monitoring and the normal reference state), and the range can be within ; represents the normalized result of the measured value of rainwater or sediment flow on this section of the road, relative to the historical benchmark. The higher the value, the more obvious the rain flow or sediment;

[0166] , is the weighting coefficient, and its value is greater than or equal to 0, which is used to balance image verification and actual flow observation;

[0167] When the on-site status confirmation value exceeds the pre-set status threshold, it indicates that a relatively high on-site anomaly confirmation degree has been achieved at this place, and an emergency linkage instruction is triggered;

[0168] When in use, through the dual-factor evaluation of images and flow, comparing the pure sensor data with the on-site real-shot or flow changes helps to quickly determine the authenticity of landslide threats under extreme weather. Making full use of the complementary relationship between UAVs, cameras and traditional sensor data greatly reduces the risk of false alarms or delays; no longer relying solely on single on-site pictures or manual visual judgments, but synthesizing the image matching degree + flow measurement into the on-site status confirmation value Taking into account visual information and soil and water dynamics information to form a reliable automated verification basis.

[0169] Step 403, Hierarchical Emergency Dispatch and Linkage Response

[0170] Based on the list of confirmed high-risk locations, including the on-site status confirmation value , distance information , and notification urgency ; If the on-site status confirmation value exceeds the pre-set status threshold and the notification urgency ranks among the top of all high-risk sections, it automatically recommends closing the corresponding section of the road and notifies the maintenance and emergency rescue departments. If it is only a preliminary anomaly, relatively loose control measures such as speed limits or intermittent passage can be taken;

[0171] Based on the notification urgency rank and schedule the priorities of multiple risks, and designate the corresponding rescue materials (such as engineering vehicles, retaining materials) and personnel to rush to the scene in batches;

[0172] To measure the collaborative efficiency of multiple parties such as the traffic management department, geological department, and emergency rescue team in resource scheduling, an emergency cluster degree is constructed: Among them, assuming that the set of indices of currently identified high-risk road sections is , and its size is , for each road section , the urgency has been notified and the on-site status confirmation value ;

[0173] There are several major collaborative departments (such as transportation, geology, emergency rescue, medical, etc.), and their collaborative relationship can be characterized by a collaborative matrix S. To measure the resource input or invocation degree of each department at the current moment t, a department usage matrix of the same dimension is defined , and based on this, the constructed emergency cluster degree is as follows:

[0174]

[0175] In the formula: The meaning of the Lambda-Gamma product vector is: After multiplying the urgency of notification of all road sections by the on-site status confirmation value , it is assembled into a column vector in the index order:

[0176]

[0177] The available range: Each component is not less than 0, and the larger the value, the higher the urgency and on-site confirmation degree of the corresponding road section;

[0178] is the road section weighting vector, then ; is a fixed symmetric matrix (or can be obtained from expert scoring and historical drill data); The element in the th row and th column of represents the degree of mutual cooperation between the th department and the is a department usage matrix, The element in the th row and th column of represents the degree of the department mobilizing the resources or cooperation ability of the th department at time , is the balance coefficient, and the values are all greater than 0;

[0179] Denotes the trace operation of a matrix (sum of diagonal elements), which in this scenario can be understood as the cumulative contribution of the inter-departmental collaboration matrix and the resource usage matrix on the diagonal; , is the integration variable;

[0180] If the emergency clustering degree value is large enough, it indicates that the response measures of each department to the main risk points are well coordinated in a short period of time; if the emergency clustering degree is low, it means that there may be a lag in resource allocation or information feedback, and integration and optimization at the command center level need to be carried out subsequently;

[0181] According to the emergency clustering degree to evaluate the multi-department collaboration effect in real time. If it continues to decrease, a coordination warning is issued to promote the intervention and scheduling of a higher-level command agency: if it remains stable or rises, it indicates that the emergency process is smooth;

[0182] When the critical section is successfully de-risked or the landslide risk degree drops back below the safe range again, automatically notify all parties to end the emergency state, and feedback the disposal situation and sensor updated data to Step 303 of Step Three to carry out model re-training, threshold optimization, and sensor status tracking;

[0183] When in use, with the help of the emergency clustering degree , the management department can timely insight into the rescue resource allocation efficiency under multiple risk points, and avoid rescue delays caused by information islands or departmental incoordination. Deeply integrate the risk monitoring results with emergency management to form a real closed-loop mechanism for prevention-monitoring-verification-disposal, and realize the automation and coordination of road safety guarantee.

[0184] Step Five: When the disaster type expands to situations such as flood, bridge structural health, or freeze-thaw damage, supplement corresponding monitoring equipment such as water level gauges, strain gauges, or surface temperature sensors and data fusion strategies, and use multi-task learning to fuse multi-hazard risks to generate a synthetic risk degree , and then through the emergency clustering degree automatically dispatch drainage or reinforcement resources to achieve integrated monitoring and linkage disposal of multiple hazards across the entire highway;

[0185] The said Step Five includes the following contents:

[0186] Step 501: Expansion of multi-hazard monitoring elements and data fusion

[0187] Add monitoring equipment for other types of highway disasters, such as water level gauges (flood monitoring), strain gauges (bridge structure health monitoring), surface temperature / humidity monitors (freeze-thaw disaster monitoring), etc.; in the process of fusion and anomaly screening, improve the hierarchical management and gap correction of the newly added sensor data, and maintain the fusion consistency with the existing fields (rainfall, soil moisture content, tilt monitoring, etc.); merge the newly added multi-disaster observation vectors into the model input structure, and build multi-task learning or multi-model integration strategies for different disaster types: Among them, the observation data from multi-source and multi-disaster sensors (such as landslide sensor output, flood water level gauge readings, bridge strain gauge information, etc.) are unified and merged to provide to the prediction model for training or real-time reasoning. The "model input structure" referred to here usually refers to the input tensor or vector batch received by the deep learning network (such as LSTM or GNN), which can also be regarded as the input layer or input feature container of the model;

[0188] In the time series network (such as LSTM) or graph structure network (such as GNN), set up special sub-networks for flood, bridge health, and freeze-thaw, or use high-order fusion layers to achieve joint representation of multi-disaster characteristics; maintain consistency in batch management of input data, and ensure the spatiotemporal alignment of new disaster characteristics (such as the same frequency or adjacent time window with the original water content, microseismic amplitude and other data) to avoid time series mismatch in the model, resulting in a decrease in accuracy;

[0189] When in use, by adding monitoring equipment related to multiple disaster types to the original sensor network and data fusion mechanism, it has the potential to use one network for multiple purposes, reduce duplicate deployment and enhance the overall disaster prevention capabilities of the highway.

[0190] Step 502: Multi-disaster comprehensive warning and joint emergency expansion

[0191] Based on the spatiotemporal prediction of landslide risk and the emergency response mechanism for sections exceeding the threshold, the simultaneous determination of multiple risk parameters such as flood, bridge damage, and freeze-thaw damage is introduced; among them:

[0192] If defined Indicates road segment The comprehensive flood risk output is represents the bridge damage risk output, Indicates freeze-thaw risk, and original landslide risk Combine together to get the composite risk :

[0193]

[0194] Where: is the set of disaster types covered;

[0195] Represents the risk value of the disaster type; Represents the adaptive threshold for the disaster type ; And Are the weighted and attenuation coefficients respectively, with values greater than or equal to 0, controlling the influence of each type of disaster in the synthesis; , Is the integral variable;

[0196] The synthesized risk degree The larger the value, the higher the comprehensive risk of the road section Under multi-disaster conditions. When scheduling resources, richer emergency measures (such as drainage equipment, reinforcement materials, or structural inspection teams, etc.) can be automatically included to achieve the goal of coordinated disposal of multi-disasters;

[0197] When the synthesized risk degree Exceeds the preset global threshold or the thresholds of each disaster type reach the warning standard, refer to the urgency of the notice And the emergency cluster degree For hierarchical disposal, but different departments or equipment are called for different disaster types:

[0198] If the flood risk is prominent, then call drainage equipment, plugging technology, or a professional flood peak monitoring team; if the risk of bridge damage is high, then give priority to scheduling a bridge structure inspection team or a traffic diversion plan; if the freeze-thaw threat is dominant, then promptly spread anti-slip materials on the road surface or restrict overweight vehicles from passing.

[0199] In the emergency cluster degree formula , it can be further extended to cover more department matrix entries (such as the water department, bridge maintenance team, road surface maintenance group, etc.) to form a comprehensive index for multi-department linkage;

[0200] When in use, by introducing the synthesized risk degree across disasters , at the emergency level, it no longer only responds to a single landslide danger, but can implement differential and comprehensive rescue scheduling according to the judgment results of different disasters.

[0201] When in use, by adding more division entries to the department matrix of the emergency cluster degree , cross-department and cross-disaster collaborative command is realized, enabling the highway disaster prevention system to truly have the ability of integrated joint prevention and control; on the basis of the previous landslide monitoring mechanism caused by extreme rainfall, it has been successfully extended to the comprehensive prevention and control of various disaster types such as floods, bridge structure health, and freeze-thaw damage, comprehensively improving the overall disaster prevention level of highway engineering, realizing the integrated scheduling of multi-disasters and multi-departments, and providing stronger flexibility and response ability for highway safety guarantee.

[0202] Among the outputs of various sensors, siting indices, risk assessment values, dynamic thresholds, collaborative emergency response indicators, and other types of parameters or formulas involved in this solution, different physical quantities are uniformly mapped to the same numerical range (such as ) or dimensionless scoring system through normalization, standardization, or equivalent dimensionless processing, thus effectively avoiding calculation conflicts that may be caused by differences in dimensions such as rainfall, tilt angle, and microseismic amplitude, and also ensuring that reliable numerical operations can be performed on various parameters during the implementation of this solution, without the situation of insufficient disclosure due to inconsistent dimensions.

[0203] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0204] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated here.

[0205] In several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only for some logical function divisions, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces, and the indirect couplings or communication connections of the devices or units can be in electrical, mechanical, or other forms.

[0206] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place, or they can be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0207] As described above, it is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims described above.

Claims

1. A highway engineering disaster monitoring and early warning system based on machine learning, characterized in that: Including, When triggered by extreme rainfall or geological monitoring requirements, multi-source sensors are allocated according to the site selection index, and the raw data is collected and calibrated to generate a raw data frame; among them, the sensor site selection index is defined by combining the distance between the sensor and the base station, the environmental load, and the local gradient information; among them, the raw data includes rainfall, soil water content, surface displacement, and microseismic amplitude; Construct a sensor site selection index to quantify the rationality of sensor deployment at a specific location; To simultaneously consider the distance between the sensor and the base station, the environmental load, and the local gradient information during the site selection process, a sensor site selection index is defined , as follows: ； Indicates the candidate location of the sensor, which can be regarded as in geographical coordinates ; Is the location of the base station or the main data aggregation node, acting as the fixed coordinates determined by the communication network planning; Indicates the candidate point And the base station The Euclidean distance between them; Indicates Of The gradient vector of the coordinates, whose norm Represents the rate of change of the distance function with fine-tuning of the position near ; Indicates the candidate point The environmental load index around; denote the gradient of coordinates indicating the change rate of the environmental load in a local area; is the balance coefficient which is used for weighting between communication convenience priority and high-risk environment priority is the attenuation factor , is the integration variable ; After the raw data frame arrives at the central node, noise is removed and null values are patched through clustering and interpolation algorithms, and then the abnormal screening mechanism is triggered to identify the over-limit sensors and call the standby equipment to form a multi-source fusion vector for model training; After the multi-source fusion vector enters the deep learning module, a time series network or a graph network is used to mine the space-time dependence and dynamically correct the global risk threshold, generate the landslide risk degree, and iteratively update the model parameters to adapt to extreme rainfall conditions; Immediately send multi-channel alarms after the landslide risk degree exceeds the global risk threshold, and coordinate the rescue teams and equipment to enter the high-risk sections according to the notification urgency and emergency cluster degree; When flood, bridge health, or freeze-thaw disaster demands are triggered in parallel, corresponding water level gauges, strain gauges, and temperature monitoring are added in the above-mentioned sensor layout and data fusion links. The multi-disaster risks are comprehensively constructed into a synthetic risk degree, and different supports are dispatched by means of an emergency cluster degree; among them, the emergency cluster degree is constructed used to measure the collaborative efficiency of the traffic management department, geological department, and emergency rescue team in resource allocation; If the mutation amplitude of the sensor exceeds the preset jump threshold within a short period of time and is judged to be unreasonable after being combined with the site selection index, trigger the secondary reset verification mechanism of the device, or switch to the standby sensor in the same area; When the constructed cross-measurement factor is continuously lower than the set threshold for multiple times, it is determined that the sensor device or data is abnormal, and it is recorded in the suspected fault list. At the same time, the standby device is automatically called or a request for on-site manual inspection is initiated. If it is found that the readings of multiple sensors in the surrounding area are abnormal at the same time, it is included in the high-risk status list; Define the following cross-measurement factors : ； In the formula: is the multi-source fusion vector; is the reference prediction vector of the multi-source fusion vector of the sensor under normal working conditions; represents the multi-dimensional difference between the current and the reference; is the sensor location selection index; is the weight of the cross-measurement factor, and the values are all greater than 0. is the attenuation factor, and the value is greater than 0. is the integration variable.

2. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 1, wherein: After obtaining the sensor site selection indexes of all candidate positions, allocate sensor resources according to the site selection index and obtain the sensor deployment plan.

3. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 2, wherein: Based on the completed sensor deployment plan, perform real-time data collection and preliminary synchronization of the sensors; In contrast to the determined location list, deploy acquisition devices at each sensor location, merge the obtained measurement values to form a raw data frame, attach the location identifier and its site selection index to each raw data frame, and preprocess the data uploaded by multiple sensors.

4. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 1, wherein: Read the raw data frames collected by the sensors respectively, merge them to form a multi-source fusion vector, introduce a generalized weighted mapping to perform spatial fusion on the multi-source data of adjacent positions, and form a spatial difference mapping matrix; Use the spatial difference mapping matrix to identify significant conflicts. If the data of a certain sensor deviates greatly from that of most adjacent sensors and the deviation duration exceeds the threshold, it will be temporarily listed as a suspected anomaly.

5. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 4, wherein: Construct all multi-source fusion vectors into multi-source time series vectors, and perform preliminary construction of landslide risk prediction to align the feature order in the multi-source time series vectors with the nodes in the input layer of the model; Divide the multi-source time series vectors into blocks according to time periods to form several training / prediction sequences in the form of sliding windows; Before loading the prediction model for each batch, check the high-risk status list and the list of suspected faults. If a certain sensor is in a suspected fault state, temporarily mask the data channel from the input vector or use the backup sensor channel instead.

6. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 5, characterized in that: Perform forward propagation on the input multi-source time series vectors batch by batch to obtain the landslide risk output; If there are multiple sensor channels at the same location, calculate the landslide risk degree. If the landslide risk degree ≥ the global risk threshold and lasts for a period of time or is at a critical location, mark the corresponding sensor as a high-risk state; If multiple adjacent sensors exceed the threshold together, generate high-risk road sections or high-risk point sets through a merging strategy or section aggregation marking, and summarize the corresponding sensors or nodes as a high-risk sensor list.

7. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 6, characterized in that: Dynamically correct the global risk threshold adaptively, combine the landslide risk degree with the known recent real feedback, and intermittently perform partial online training on the deep learning model parameters; When a new extreme rainfall situation occurs at a certain location and the model lacks relevant samples, the multi-source fusion vector and the landslide risk degree can be stored in the extreme sample library.

8. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 7, characterized in that: Based on generating a high-risk sensor list and the corresponding landslide risk degree: For the road sections where the sensors with landslide risk degree ≥ the global risk threshold are located, automatically send warning messages to the road management platform and push them to the monitoring center. If multiple adjacent sensors exceed the threshold at the same time, aggregate them and mark them as the same high-risk road section; Define the notification urgency for sorting multiple high-risk road sections, screen out the high-priority road section list according to the notification urgency, and give priority warnings according to the notification urgency and arrange emergency resources for attention.

9. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 8, characterized in that: Obtain the on-site immediate status and construct a list of confirmed high-risk locations, compare the verification results with the multi-source fusion vector data, and define the on-site status confirmation value to quantify the consistency between image inspection and soil body dynamics: When the on-site status confirmation value exceeds the pre-set status threshold, trigger an emergency linkage instruction; If the on-site status confirmation value exceeds the pre-set status threshold and the on-site status confirmation value ranks among the top of all high-risk sections, automatically recommend closing the corresponding road section and notify the maintenance and emergency rescue departments.

10. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 9, characterized in that: Rank multiple risks and plan the scheduling priorities based on the on-site status confirmation values, and specify the corresponding relief supplies and personnel to go to the scene in batches; Build a real-time evaluation of the emergency cluster degree to coordinate the effects of multiple departments. If it continues to decrease, issue a coordination warning; When the critical section is successfully de-risked or the landslide risk level drops below the safe range again, automatically notify all parties to end the emergency state, and start model retraining, global risk threshold optimization, and sensor status tracking.

11. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 10, characterized in that: Add monitoring equipment for other highway disaster types, merge the newly added multi-disaster observation vectors into the model input structure, and build multi-task learning or multi-model integration strategies for different disaster types: Set dedicated sub-networks for flood, bridge health, and freeze-thaw in the time series network or graph structure network respectively, or use a high-order fusion layer to achieve joint representation of multi-disaster features.

12. The machine learning-based highway engineering disaster monitoring and early warning system according to claim 11, characterized in that: Based on the spatio-temporal prediction of landslide risks and the emergency response mechanism for sections exceeding the threshold, introduce synchronous determination of multiple risk parameters, set adaptive thresholds for each disaster type respectively, and combine them to obtain a composite risk degree. The exceeded threshold is the global risk threshold; When the composite risk degree exceeds the preset global risk threshold or the thresholds of each disaster type reach the warning standard, conduct hierarchical disposal with reference to the notification urgency and emergency cluster degree, and call different departments or equipment for different disaster types.

Images