Road disaster early warning method and system based on internet of things sensing analysis

Abstract

The present application relates to the field of data processing, and specifically provides a road disaster early warning method and system based on Internet of Things sensing analysis, the method comprising: collecting real-time environmental data through a multi-modal sensing device deployed in a road monitoring area; obtaining vibration fluctuation features, deformation displacement vectors, humidity gradient features and visual texture features after feature extraction; fusing the features in multi-modal space-time features to generate road state fusion features inputting a disaster prediction model for disaster probability calculation to obtain the disaster occurrence probability of the current road area, and triggering an early warning signal matched with the disaster type when the disaster occurrence probability exceeds a disaster threshold. The present application can improve the accuracy, timeliness, comprehensiveness and model adaptability of road disaster early warning, effectively ensuring the safety and smoothness of the road.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to a road disaster early warning method and system based on Internet of Things sensing analysis. BACKGROUND

[0002] In the operation and maintenance process of road infrastructure, timely early warning of road disasters is of great importance, as it is directly related to the safety of public life and property as well as the smooth operation of the transportation system. Currently, road disaster early warning mainly relies on single sensor data or simple data analysis methods. For example, some systems only rely on information obtained from vibration sensors to determine whether there is a risk of disaster on the road. This approach lacks support from multi-dimensional data, making it difficult to comprehensively and accurately reflect the actual condition of the road, resulting in low accuracy and reliability of the early warning. Some traditional early warning systems fail to fully consider the correlation and mutual influence between different data when processing data, and cannot effectively fuse multi-modal data, resulting in significant limitations in the early warning results. Moreover, existing disaster prediction models often lack a dynamic updating mechanism, and cannot adapt to changes in road conditions and new types of disasters in a timely manner, resulting in poor timeliness and adaptability of the models. Therefore, there is an urgent need for a more comprehensive, accurate and intelligent road disaster early warning method and system to improve the accuracy, timeliness and effectiveness of road disaster early warning, and to ensure the safety and smoothness of the road. SUMMARY

[0003] The present application relates to the field of data processing, in particular to a road disaster early warning method and system based on Internet of Things sensing analysis.

[0004] In the operation and maintenance process of road infrastructure, timely early warning of road disasters is of great importance, as it is directly related to the safety of public life and property as well as the smooth operation of the transportation system. Currently, road disaster early warning mainly relies on single sensor data or simple data analysis methods. For example, some systems only rely on information obtained from vibration sensors to determine whether there is a risk of disaster on the road. This approach lacks support from multi-dimensional data, making it difficult to comprehensively and accurately reflect the actual condition of the road, resulting in low accuracy and reliability of the early warning. Some traditional early warning systems fail to fully consider the correlation and mutual influence between different data when processing data, and cannot effectively fuse multi-modal data, resulting in significant limitations in the early warning results. Moreover, existing disaster prediction models often lack a dynamic updating mechanism, and cannot adapt to changes in road conditions and new types of disasters in a timely manner, resulting in poor timeliness and adaptability of the models. Therefore, there is an urgent need for a more comprehensive, accurate and intelligent road disaster early warning method and system to improve the accuracy, timeliness and effectiveness of road disaster early warning, and to ensure the safety and smoothness of the road. SUMMARY

[0003] The present application relates to the field of data processing, in particular to a road disaster early warning method and system based on Internet of Things sensing analysis.

[0004] In the operation and maintenance process of road infrastructure, timely early warning of road disasters is of great importance, as it is directly related to the safety of public life and property as well as the smooth operation of the transportation system. Currently, road disaster early warning mainly relies on single sensor data or simple data analysis methods. For example, some systems only rely on information obtained from vibration sensors to determine whether there is a risk of disaster on the road. This approach lacks support from multi-dimensional data, making it difficult to comprehensively and accurately reflect the actual condition of the road, resulting in low accuracy and reliability of the early warning. Some traditional early warning systems fail to fully consider the correlation and mutual influence between different data when processing data, and cannot effectively fuse multi-modal data, resulting in significant limitations in the early warning results. Moreover, existing disaster prediction models often lack a dynamic updating mechanism, and cannot adapt to changes in road conditions and new types of disasters in a timely manner, resulting in poor timeliness and adaptability of the models. Therefore, there is an urgent need for a more comprehensive, accurate and intelligent road disaster early warning method and system to improve the accuracy, timeliness and effectiveness of road disaster early warning, and to ensure the safety and smoothness of the road.

[0005] In a second aspect, the present application provides a computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0006] The embodiment of the present application extracts and analyzes the vibration waveform sequence and the ground deformation trajectory respectively, the pre-trained time sequence feature encoder extracts the vibration fluctuation feature, and the spatial feature encoder generates the deformation displacement vector. This processing method for different data characteristics can mine key information in the data and highlight the essential features of the data, making the subsequent analysis more targeted. The humidity distribution map is regionally segmented and the humidity gradient feature is extracted, and the visible light image stream is dynamically frame-sampled and the visual texture feature is extracted, which further refines the analysis of the data and improves the usability and value of the data. The multi-modal spatio-temporal feature fusion of multiple features generates road state fusion features, breaks the limitations of single data modalities, integrates information from multiple aspects, and makes the understanding of road conditions more comprehensive and in-depth. The road state fusion features are input into the disaster prediction model to calculate the disaster occurrence probability, and the warning signal is triggered according to the dynamically adjusted disaster threshold, while setting a hierarchical response mechanism for the warning signal and an online updating mechanism for the disaster prediction model. Based on this, in terms of the accuracy of disaster warning, the collection and fusion of multi-modal data, as well as the fine processing and feature extraction of different data, enable the disaster prediction model to more accurately capture changes in road conditions and potential disaster risks, reducing false positives and omissions. For example, by considering vibration, deformation, humidity and visual information, it can more accurately determine whether there are landslide, collapse and other disaster hazards on the road. In terms of the timeliness of the warning, by collecting data in real time and dynamically adjusting the disaster threshold, changes in the probability of disaster occurrence can be detected in time, and appropriate warning signals can be triggered quickly when a disaster is about to occur or there are signs of a disaster, giving road maintenance and traffic management departments more time to respond. In terms of the comprehensiveness of the warning, the hierarchical warning mechanism can take different measures according to different levels of disaster occurrence probability, not only reminding the road maintenance department to handle it, but also considering the impact on traffic flow, generating a path avoidance suggestion vector and adjusting the phase timing parameters of the traffic signal control node to ensure road traffic safety. In terms of the adaptability of the model, the online updating mechanism of the disaster prediction model can adjust and optimize the model according to the actual situation after the warning, so that the model can adapt to the dynamic changes of road conditions and continuously improve the accuracy and reliability of the prediction. In summary, this scheme has achieved good results in terms of the accuracy, timeliness, comprehensiveness and model adaptability of road disaster warning through a series of technical means, effectively ensuring the safety and smoothness of the road. Attached Figure Description

[0007] Figure 1 This is a flowchart of a road disaster early warning method based on Internet of Things sensing analysis provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] In this embodiment of the invention, the execution entity of the road disaster early warning method based on IoT sensor analysis is a computer system, including but not limited to servers, personal computers, laptops, tablets, and smartphones. The computer system can run independently to implement this invention, or it can connect to a network and interact with other computer systems within the network to implement this invention. The network in which the computer system resides includes, but is not limited to, the Internet, wide area networks, metropolitan area networks, local area networks, and VPN networks.

[0010] like Figure 1 As shown, the road disaster early warning method based on IoT sensor analysis provided in this embodiment of the invention includes:

[0011] Step S100: Collect real-time environmental data by deploying multimodal sensing devices in the road monitoring area. The real-time environmental data includes vibration waveform sequences, surface deformation trajectories, humidity distribution maps, and visible light image streams.

[0012] In step S100, the vibration waveform sequence reflects the road's vibration at different points in time. Roads vibrate when subjected to external forces such as vehicle traffic and surrounding construction, and these vibrations are represented by waveforms. For example, when a large truck travels on the road, it causes significant vibration. Vibration sensors collect this vibration information and convert it into time-series waveform data, i.e., the vibration waveform sequence. For instance, an accelerometer can be used to sense road vibrations. The accelerometer converts the vibration acceleration into an electrical signal, which is then processed through analog-to-digital conversion to obtain a digitized vibration waveform sequence.

[0013] Surface deformation trajectories reflect the changes in the position of the road surface over a period of time. When influenced by factors such as changes in groundwater levels and geological activity, the road surface may shift. For example, in some mountainous areas, landslides may cause the road surface to subside or rise. Displacement sensors continuously monitor these changes and record the location information at each point in time, forming a surface deformation trajectory. Alternatively, a Global Navigation Satellite System (GNSS) receiver can be used to accurately measure the surface position by receiving satellite signals, thus obtaining the surface deformation trajectory.

[0014] The humidity distribution map shows the humidity conditions at different locations within the road area. Humidity has an important impact on the stability of the road, and excessive humidity can cause the road base to soften, cracks on the road surface, and other problems. For example, during the rainy season, the humidity of the road surface and base will increase significantly. By collecting humidity data at different locations through humidity sensors distributed in the road monitoring area, these data are processed and analyzed to generate a humidity distribution map. For example, a capacitive humidity sensor can be used, which works on the principle of using humidity changes to cause changes in the capacitance value, and the humidity information is obtained by measuring the capacitance value.

[0015] The visible light image stream is video stream data obtained by continuously shooting the road area through devices such as cameras. The visible light image stream can directly reflect the actual conditions of the road, such as whether there are obstacles, whether the road surface is damaged, etc. For example, when a traffic accident occurs on the road, the camera can shoot the images of the accident scene, and these continuous image frames form the visible light image stream. For example, a high-definition network camera can be used to transmit the shot image data to the processing through the network.

[0016] Step S200: input the vibration waveform sequence into the pre-trained time sequence feature encoder for waveform feature extraction, obtain the vibration fluctuation feature, and input the ground surface deformation trajectory into the spatial feature encoder for deformation trajectory analysis to generate a deformation displacement vector.

[0017] In step S200, for the processing of the vibration waveform sequence, a pre-trained time sequence feature encoder is used for waveform feature extraction. This encoder can capture key features in the vibration waveform through learning from historical vibration waveform data. For example, when the road is subjected to external forces such as vehicle driving and surrounding construction, different frequencies and amplitudes of vibration are generated, which are manifested in the form of waveforms. The time sequence feature encoder can extract features reflecting the vibration fluctuation from these waveforms, i.e., the vibration fluctuation feature. In the specific implementation process, the pre-trained time sequence feature encoder contains multiple layers of dilated convolution kernels, which can extract waveform features of different scales and time steps through convolution operations on the input vibration waveform sequence. The principle is to use dilated convolution kernels to expand the receptive field without increasing the number of convolution kernel parameters, so as to better capture the long-time dependence of the waveform. Through further processing of the features output by the convolution layer, the vibration fluctuation feature is obtained, which can reflect the change law and trend of road vibration.

[0018] For the processing of the ground surface deformation trajectory, it is input into the spatial feature encoder for deformation trajectory analysis. The ground surface deformation trajectory records the position changes of the road surface within a period of time, and the information reflecting the deformation direction and intensity is analyzed from these trajectory data. The spatial feature encoder generates a deformation displacement vector through processing and analysis of the trajectory data.

[0019] In a specific implementation process, the spatial feature encoder first samples the ground surface deformation trajectories at equal intervals, extracts the three-dimensional coordinate data of each trajectory point and the corresponding time stamp. Then, according to the time difference between adjacent trajectory points, the displacement change rate between adjacent trajectory points is calculated. The three-dimensional coordinate data of each trajectory point and the corresponding displacement change rate are spliced to generate the displacement change vector of each trajectory point.

[0020] Next, the displacement change vectors of all trajectory points are arranged in time sequence as a displacement change sequence, and input into the graph attention network. In the graph attention network, the displacement change vector of each trajectory point is taken as the node feature of the graph, and the directed edge connection relationship between nodes is constructed according to the time sequence of adjacent trajectory points. By aggregating the features of the multi-hop neighborhood nodes of each graph node feature, the node aggregation feature containing the local deformation propagation trend is generated.

[0021] Finally, the node aggregation feature is input into the attention weight calculation layer of the graph attention network, the attention weight value between each graph node and the previous time graph node is calculated, and the node aggregation feature is weighted and summed according to the weight value to generate the deformation displacement vector reflecting the deformation direction and intensity.

[0022] Step S300: Perform region segmentation on the humidity distribution map, extract the humidity gradient features of each segmented region, and perform dynamic frame sampling on the visible light image stream to obtain an image frame sequence and input it into the image feature encoder to extract visual texture features.

[0023] In step S300, the humidity distribution map is regionally segmented to more carefully analyze the humidity conditions of different regions of the road, because the humidity changes in different regions can have different effects on the occurrence of road disasters. For example, the edge region and the center region of the road can have significant differences in humidity distribution due to differences in drainage conditions and the degree of influence from the external environment. A threshold-based segmentation method is used for regional segmentation, and the humidity distribution map is divided into multiple different regions according to the size of the humidity value. Specifically, one or more humidity thresholds are set, and when the humidity value of a region is higher or lower than the threshold, it is divided into different regions. For example, a high humidity threshold and a low humidity threshold are set, and the region with a humidity value higher than the high humidity threshold is divided into a high humidity region, the region with a humidity value lower than the low humidity threshold is divided into a low humidity region, and the region between the two is divided into a moderate humidity region. After completing the regional segmentation, the humidity gradient features of each segmented region are extracted. The humidity gradient reflects the rate of change of humidity in space, and is important for determining whether there is water accumulation or rapid drying on the road. The humidity gradient is obtained by calculating the humidity difference between adjacent pixel points. For a two-dimensional humidity distribution map, assuming that the humidity value of a pixel point is H(x, y), the humidity values of its adjacent pixel points are H(x+1, y), H(x-1, y), H(x, y+1), and H(x, y-1), then the humidity gradients of the pixel point in the x direction and the y direction are: x = H(x+1,y)-H(x-1,y), G y = H(x,y+1)-H(x,y-1). By statistically analyzing the humidity gradients of all pixel points in each segmented region, such as calculating the mean and standard deviation, the humidity gradient features of the region are obtained.

[0024] For the visible light image stream, dynamic frame sampling is performed to obtain an image frame sequence. Since the visible light image stream usually contains a large number of consecutive frames, in order to reduce the amount of calculation while retaining key information, a dynamic frame sampling method is used. Dynamic frame sampling determines the sampling interval according to the changes in the image content. When the image content changes greatly, the sampling interval is small to capture more details; when the image content changes little, the sampling interval is large. For example, during periods when vehicles frequently travel on the road and the scene changes greatly, 1 frame can be sampled every 1 frame; during periods when there are few vehicles on the road and the scene is relatively stable, 1 frame can be sampled every 5 frames. After obtaining the image frame sequence, it is input into an image feature encoder to extract visual texture features. Visual texture features can reflect the microstructure and texture information of the road surface, and play an important role in identifying whether there are cracks, potholes, and other damage on the road. The image feature encoder usually uses a convolutional neural network (CNN), which extracts features from image frames through convolutional layers, pooling layers, and other operations.

[0025] Taking the ResNet network as an example, it contains multiple residual blocks, each of which is composed of a convolutional layer, a batch normalization layer and an activation function. The input image frame is first subjected to feature extraction by the convolutional layer to obtain a feature map, then the feature map is normalized by the batch normalization layer to speed up the convergence speed of the network, and finally the activation function is introduced to introduce non-linear factors and enhance the expression ability of the network. After processing by multiple residual blocks, the network output features are visual texture features. By region segmentation of the humidity distribution map to extract humidity gradient features, and dynamic frame sampling of the visible light image stream to extract visual texture features, rich and valuable information is provided for subsequent multi-modal spatio-temporal feature fusion and road disaster warning. These features can reflect the state of the road from different angles, which helps to more accurately determine whether there is a potential disaster risk on the road.

[0026] Step S400: multi-modal spatio-temporal feature fusion of vibration fluctuation features, deformation displacement vectors, humidity gradient features and visual texture features to generate road state fusion features.

[0027] In step S400, first, the time correlation between the vibration fluctuation features and the deformation displacement vectors is processed, because although the two describe the road state from different angles, there may be differences in the time dimension. For example, when the road is affected by vehicle driving, the vibration fluctuation features will respond quickly, while the ground deformation may have a certain delay. Through time alignment operation, the two are synchronized in time so that their correlation can be better analyzed. Specifically, by matching the timestamps of the two, corresponding data with similar times are matched to generate vibration-deformation correlation features, which can reflect the mutual influence between vibration and deformation, such as whether vibration will exacerbate ground deformation, etc.

[0028] For humidity gradient features and visual texture features, since they may have inconsistent resolutions in spatial distribution, the humidity gradient features are subjected to spatial interpolation processing. Assuming that the original resolution of the humidity gradient features is low, while the resolution of the visual texture features is high, according to the known humidity gradient feature points, the humidity gradient values at other positions are estimated by methods such as bilinear interpolation, so that the resolution of the humidity gradient features is consistent with the spatial distribution of the visual texture features. The bilinear interpolation formula is: where (x, y) is the coordinate of the point to be interpolated, (x', y') is the coordinate of the surrounding known point, and u and v are relative coordinate values.

[0029] After the above processing, the vibration-deformation correlation feature, humidity gradient feature, and visual texture feature are input into the cross-modal fusion network. The network includes a channel attention module and a spatial attention module. The channel attention module is used to dynamically allocate channel weights of multi-modal features. The importance of features in different channels to the road state may be different. For example, in some cases, the features of vibration-related channels may better reflect the immediate condition of the road, and the channel attention module adjusts the weights of each channel according to the importance of the features. The spatial attention module weights the importance of the spatial position of the feature map. In the road image, the features in some key areas such as the road edge and the crack may be more important, and the spatial attention module enhances the weights of the features in these areas.

[0030] Finally, the weighted multi-modal features are channel-spliced, i.e., the features of different modalities are connected in the channel dimension to form a feature tensor containing more information. Then, the feature tensor is processed by a dimension reduction convolution layer to reduce the dimension of the features through convolution operation and remove redundant information, and finally a road state fusion feature is generated. This fusion feature integrates the advantages of multiple modal data and can more comprehensively and accurately reflect the current state of the road, providing strong support for subsequent disaster prediction.

[0031] Step S500: input the road state fusion feature into a disaster prediction model to calculate the disaster probability, and when the disaster probability exceeds the dynamically adjusted disaster threshold, trigger a warning signal matching the disaster type.

[0032] In step S500, after obtaining the road state fusion feature, it is provided as input to the disaster prediction model, which is trained to predict the probability of road disaster according to the input feature data. For example, the road state fusion feature may contain vibration fluctuation features, deformation displacement vectors, humidity gradient features, and visual texture features, and the disaster prediction model will analyze these information comprehensively. If the vibration fluctuation feature shows that the road has an abnormally strong vibration, the deformation displacement vector indicates that the ground has a large displacement change, the humidity gradient feature shows that the local humidity is abnormally high, and the visual texture feature shows that the road surface has obvious cracks, etc., these factors will make the disaster prediction model judge that the probability of disaster in the road area is high. In the specific implementation process of the disaster prediction model, a series of calculations and processing are performed on the input road state fusion feature.

[0033] The model can include structures such as gated recurrent unit layers and self-attention mechanism layers. The gated recurrent unit layers can process sequence data by sequentially processing road state fusion features according to time steps, extracting hidden state features at each time step, which contain road state information at different time points. The self-attention mechanism layer calculates the correlation weight matrix between these hidden state features, and generates the final disaster probability prediction vector by weighted summation of the correlation weight matrix. The elements in the vector represent the occurrence probability of different disaster types. After obtaining the disaster occurrence probability, it needs to be compared with the dynamically adjusted disaster threshold. The dynamically adjusted disaster threshold is determined according to the historical disaster occurrence probability dataset and actual disaster records of the current road area. First, the historical data is obtained, and the false alarm rate and the false negative rate in different probability intervals are calculated. The false alarm rate refers to the proportion of the model predicting that a disaster occurs but actually does not occur, and the false negative rate refers to the proportion of the model predicting that a disaster does not occur but actually does occur. Then, an optimization function is constructed with the constraint condition that the false alarm rate does not exceed the first threshold and the false negative rate does not exceed the second threshold, and the particle swarm optimization algorithm is used to solve the optimization function to obtain the dynamically adjusted disaster threshold corresponding to the current road area. For example, for a road that often experiences landslides, historical data shows that the false alarm rate is high in certain probability intervals, and the disaster threshold is adjusted according to these conditions to reduce the false alarm rate. When the disaster occurrence probability exceeds the dynamically adjusted disaster threshold, the warning signal matching the disaster type is triggered. The warning signal is usually divided into different warning level intervals according to the numerical range of the disaster occurrence probability, such as a first warning signal, a second warning signal, and a third warning signal. Different levels of warning signals correspond to different response measures. When the first warning signal is triggered, the real-time disaster coordinates and disaster image frames in the visible light image stream of the current road area are extracted, and these information is sent to the road maintenance terminal and the emergency communication link of the road maintenance terminal is activated. The road maintenance terminal can understand the disaster situation in time and respond according to these information. The emergency response state information fed back by the road maintenance terminal is received through the emergency communication link, associated with the disaster coordinates, and input into the warning signal continuous monitoring queue. If the disaster occurrence probability does not decrease in the warning signal continuous monitoring queue for three consecutive monitoring periods, the first warning signal is upgraded to the second warning signal. When the second warning signal is triggered, the traffic flow distribution data and vehicle position dataset of the current road area are further extracted, and the path avoidance suggestion vector is generated according to the disaster coordinates and the vehicle position dataset. For example, if a collapse disaster occurs in a certain area of the road, the driving path of the vehicle is planned to avoid the collapse area according to the position of the vehicle and the traffic flow. Then the path avoidance suggestion vector is pushed to the vehicle terminal corresponding to the vehicle position dataset to remind the vehicle driver to take appropriate measures.Meanwhile, traffic signal control nodes affected by disaster coordinates in traffic flow distribution data are extracted, phase timing parameters of the traffic signal control nodes are adjusted according to push coverage of the path avoidance recommendation vector, and the adjusted phase timing parameters are synchronized to a traffic signal control system, so as to optimize traffic flow and reduce the influence of disasters on traffic.

[0034] In the subsequent embodiment introduction, the application also sets an online updating mechanism of the disaster prediction model. After triggering the early warning signal, the vibration waveform sequence, the ground deformation trajectory and the visible light image stream of the current road area are continuously collected to generate the road state change data set after the early warning. The data set is time stamped with the triggered early warning signal type to generate the model incremental training data set with the early warning label. Every preset model updating period, the incremental training samples in the recent time window are extracted from the model incremental training data set and input into the disaster prediction model for forward propagation. The incremental loss value between the prediction probability of the disaster prediction model for the incremental training samples and the early warning label is calculated, and the model parameter fine-tuning gradient is generated according to the loss value. After the gradient clipping algorithm is used to constrain the amplitude of the model parameter fine-tuning gradient, the fully connected layer weight parameters of the disaster prediction model are updated. Finally, the updated disaster prediction model is evaluated for disaster recognition accuracy on the historical verification set, and when the disaster recognition accuracy exceeds the evaluation result of the original model, the updated disaster prediction model is deployed to the road monitoring area to improve the prediction accuracy and adaptability of the model.

[0035] In summary, the embodiment of the present application extracts and analyzes the vibration waveform sequence and the ground surface deformation trajectory respectively, the pre-trained time sequence feature encoder extracts the vibration fluctuation feature, and the spatial feature encoder generates the deformation displacement vector. This processing method for different data characteristics can mine key information in the data and highlight the essential features of the data, making the subsequent analysis more targeted. The humidity distribution map is segmented into regions and the humidity gradient feature is extracted, and the visible light image stream is dynamically frame-sampled and the visual texture feature is extracted, which further refines the analysis of the data and improves the usability and value of the data. The multi-modal spatio-temporal feature fusion of multiple features generates road state fusion features, breaking the limitations of single data modalities, integrating information from multiple aspects, and making the understanding of road conditions more comprehensive and in-depth. The road state fusion features are input into the disaster prediction model to calculate the disaster occurrence probability, and the warning signal is triggered according to the dynamically adjusted disaster threshold, and a hierarchical response mechanism of the warning signal and an online updating mechanism of the disaster prediction model are set. Based on this, in terms of the accuracy of disaster warning, the collection and fusion of multi-modal data, as well as the fine processing and feature extraction of different data, enable the disaster prediction model to more accurately capture the changes in road conditions and potential disaster risks, reducing false positives and omissions. For example, by considering vibration, deformation, humidity and visual information, it can more accurately determine whether there are landslide, collapse and other disaster hazards on the road. In terms of timeliness of warning, by collecting data in real time and dynamically adjusting the disaster threshold, changes in disaster occurrence probability can be detected in time, and appropriate warning signals can be triggered quickly when disaster is about to occur or there are signs, giving road maintenance and traffic management departments more time to respond. In terms of comprehensiveness of warning, the hierarchical warning mechanism can take different measures according to different degrees of disaster occurrence probability, not only reminding the road maintenance department to handle it, but also considering the impact on traffic flow, generating path avoidance suggestion vectors and adjusting the phase timing parameters of traffic signal control nodes to ensure road traffic safety. In terms of adaptability of the model, the online updating mechanism of the disaster prediction model can adjust and optimize the model according to the actual situation after the warning, so that the model can adapt to the dynamic changes of road conditions and continuously improve the accuracy and reliability of the prediction. In summary, this scheme has achieved good results in terms of accuracy, timeliness, comprehensiveness and model adaptability of road disaster warning through a series of technical means, which can effectively ensure the safety and smoothness of the road.

[0036] As an implementation manner, the pre-training process of the time sequence feature encoder includes:

[0037] Step S201: Obtain a historical vibration waveform data set, and the historical vibration waveform data set contains time sequence waveform data labeled with disaster types;

[0038] In practical applications, roads will produce different characteristic vibration waveforms under different disaster conditions. For example, when a landslide disaster occurs on a road, the sliding of the mountain will impact and pull the road, causing the road to vibrate at a specific frequency and amplitude. This vibration will be recorded by the vibration sensors deployed in the road monitoring area, forming time series waveform data. A large amount of such historical data needs to be collected and labeled with disaster types. The labeling process can be done by professional geology disaster experts based on relevant geological data, field survey results, etc. In this way, a historical vibration waveform data set containing multiple disaster types is obtained, and each time series waveform data in the set corresponds to a specific disaster type, providing rich and targeted samples for subsequent model training.

[0039] Step S202: Divide the historical vibration waveform data set into multiple time window segments and perform waveform amplitude normalization processing on each time window segment.

[0040] Time window division is to divide continuous time series waveform data into smaller, more manageable segments. For example, a 100-second vibration waveform data can be divided into 10 time window segments, each with a fixed time length of 10 seconds. This allows for more detailed analysis of the changes in waveform characteristics over different time periods. Waveform amplitude normalization processing is to eliminate the influence of amplitude differences between different waveforms, so that the model can focus more on the shape and frequency characteristics of the waveform. The specific method of normalization can use linear normalization, the formula is: where x is a value in the original waveform data, is the minimum value of the waveform data in the time window segment, is the maximum value, is the normalized value. Through this normalization processing, the amplitudes of all waveform data in the time window segment are mapped to the interval [0, 1], making different waveform data comparable and helping to improve the training effect of the model.

[0041] Step S203: Construct a contrast learning network containing multiple layers of dilated convolution kernels, input different time window segments of the same disaster type as positive sample pairs into the contrast learning network, and input time window segments of different disaster types as negative sample pairs into the contrast learning network.

[0042] An inflation convolution kernel is a special convolution kernel that introduces the concept of inflation rate on the basis of a normal convolution kernel, expands the receptive field of convolution by inserting zero elements between the elements of the convolution kernel, without increasing the number of parameters of the convolution kernel. For example, a normal 3x3 convolution kernel, if the inflation rate is set to 2, will actually cover a larger area during convolution calculation. Multi-layer inflation convolution kernel can capture waveform features of different scales and time steps, helping the model better learn the long-time dependence of the waveform. The goal of the contrastive learning network is to improve the model's ability to distinguish different disaster types by learning the difference between positive sample pairs and negative sample pairs. When inputting the sample, two different time window segments of the same disaster type are combined into a positive sample pair because they have similar waveform features and represent the same disaster type; while time window segments of different disaster types are combined into a negative sample pair, which have large differences in waveform features and represent different disaster types. For example, two time window segments labeled as landslide disaster are combined as a positive sample pair, and a time window segment labeled as landslide disaster and a time window segment labeled as earthquake disaster are combined as a negative sample pair. By continuously inputting these positive sample pairs and negative sample pairs to the contrastive learning network, the model can learn the waveform feature patterns corresponding to different disaster types.

[0043] Step S204: Generate a contrastive loss value by calculating the waveform feature similarity of the positive sample pair and the waveform feature difference of the negative sample pair, and optimize the parameters of the contrastive learning network based on the contrastive loss value.

[0044] Specifically, the calculation of the waveform feature similarity can use cosine similarity, the formula is: where A and B are the waveform feature vectors of the two time window segments in the positive sample pair, · represents the dot product of the vectors, and are the norms of the vectors, and sim is the cosine similarity. The closer the value of the cosine similarity is to 1, the more similar the two waveform features are. The calculation of the waveform feature difference can use Euclidean distance, the formula is: where a i and b i are the i-th elements of the waveform feature vectors of the two time window segments in the negative sample pair, n is the dimension of the vector, and dist is the Euclidean distance. The larger the value of the Euclidean distance, the greater the difference between the two waveform features. The calculation of the contrastive loss value can use a triplet loss function, the formula is: where is the waveform feature similarity of the positive sample pair, is the waveform feature similarity of the negative sample pair, is a preset boundary value for controlling the degree of difference between the positive sample pair and the negative sample pair. According to the contrast loss value, an optimization algorithm such as gradient descent is used to adjust the parameters of the contrast learning network, so that the similarity of the positive sample pair is as high as possible, and the similarity of the negative sample pair is as low as possible, thereby improving the model's ability to distinguish different disaster types.

[0045] Step S205: The output features of the dilated convolution layer in the optimized contrast learning network are taken as the vibration fluctuation feature extraction results of the time sequence feature encoder.

[0046] After the contrast learning network is trained and optimized, the dilated convolution layer inside it has learned the waveform feature patterns corresponding to different disaster types. The dilated convolution layer can extract key features of the waveform by performing convolution operations on the input time window segments. These features contain information such as frequency and amplitude changes, and can reflect the fluctuation of road vibration. Taking these output features as the vibration fluctuation feature extraction results of the time sequence feature encoder, when a new vibration waveform sequence is input to the time sequence feature encoder, the encoder can quickly and accurately extract the vibration fluctuation features according to the previously learned feature patterns, providing strong support for subsequent road disaster early warning. Through such a pre-training process, the time sequence feature encoder can better adapt to vibration waveform data of different disaster types, improve the extraction ability of road vibration features, and lay a foundation for the accuracy and reliability of the entire road disaster early warning system.

[0047] As an implementation manner, in step S200, the ground surface deformation track is input into the spatial feature encoder for deformation track analysis to generate a deformation displacement vector, including:

[0048] Step S210: The ground surface deformation track is sampled at equal intervals to extract three-dimensional coordinate data and corresponding time stamps of each track point.

[0049] In actual road monitoring scenarios, the ground deformation trajectory is continuously recorded by high-precision displacement sensors deployed around the road. These trajectory data are continuous, but need to be discretized for subsequent processing and analysis, i.e., equidistant trajectory point sampling. For example, assuming that the displacement sensor records the ground position at a frequency of once per second, and the sampling interval is set to 10 seconds, then a trajectory point will be selected every 10 seconds from the continuous trajectory data. Each trajectory point contains three-dimensional coordinate data, i.e., the x, y, z coordinates of the point in space, representing the position information in the horizontal, vertical, and vertical directions, respectively. At the same time, each trajectory point also corresponds to a timestamp, which records the specific time when the point is collected. The timestamp can be accurate to seconds or even milliseconds, facilitating subsequent analysis of ground deformation at different time points. Through such sampling and data extraction, continuous ground deformation trajectory is converted into a series of discrete trajectory point data with clear time and spatial information, providing the necessary basis for subsequent calculation of displacement change rate and generation of displacement change vector.

[0050] Step S220: Calculate the displacement change rate between adjacent trajectory points according to the time difference between the timestamp of each trajectory point and the timestamp of the previous trajectory point.

[0051] Such calculation can reflect the deformation speed of the ground in different time periods. Taking the trajectory point sampled in step S210 as an example, assume that the timestamp of a trajectory point P is , the timestamp of the previous trajectory point P is , and the time difference between the two points is . At the same time, the three-dimensional coordinates of the trajectory point P are , and the three-dimensional coordinates of the trajectory point P i-1 are , then the displacement vector between the two points is . The displacement change rate v can be obtained by dividing the modulus of the displacement vector by the time difference, i.e., . This displacement change rate can intuitively reflect the deformation speed of the ground between the two adjacent sampling time points. If the displacement change rate is large, it indicates that the ground has undergone rapid deformation in this period of time, which may pose a potential disaster risk; if the displacement change rate is small, it indicates that the ground is relatively stable.

[0052] Step S230: Concatenate the three-dimensional coordinate data of each trajectory point with the corresponding displacement change rate to generate the displacement change vector of each trajectory point.

[0053] This concatenation method combines the position information and deformation speed information of the trajectory point. Continuing with the data in steps S210 and S220, for the trajectory point P i-1 , the displacement change vector isi Its three-dimensional coordinates are (x i ,y i ,z i The rate of change of displacement is v i Then the displacement change vector v of the trajectory point i =(x i ,y i ,z i ,v i By concatenating the three-dimensional coordinates and displacement change rate, the static position information and dynamic change information of the trajectory points can be integrated into a single vector, making the information of each trajectory point richer and more comprehensive. The resulting displacement change vector can more accurately describe the state of the Earth's surface at each sampling moment, providing stronger data support for subsequent analysis of the trends and patterns of surface deformation.

[0054] Step S240: Arrange the displacement change vectors of all trajectory points in chronological order to form a displacement change sequence, and input the displacement change sequence into the graph attention network.

[0055] This step leverages the powerful capabilities of graph attention networks to process temporally ordered sequence data. Based on the displacement change vector of each trajectory point generated in step S230, they are arranged into a sequence according to the order of their timestamps, for example, {v1, v2, ..., v...}. n} where n is the total number of trajectory points. Graph attention networks are a type of neural network based on a graph structure that can automatically learn the relationships and importance between nodes in a graph. In this scenario, the displacement change vector of each trajectory point can be viewed as a node in the graph, and the temporal order relationship between nodes can be represented by edges. After inputting the displacement change sequence into the graph attention network, the network can use its internal attention mechanism to weight the displacement change vectors at different time points, highlighting important information, thereby better capturing the long-term dependencies and local change characteristics of surface deformation.

[0056] Step S250: In the graph attention network, the displacement change vector of each trajectory point is used as the graph node feature, and directed edge connections between nodes are constructed according to the time order of adjacent trajectory points.

[0057] This step constructs a suitable graph structure for the graph attention network. In the graph attention network, the displacement change vector v of each trajectory point... i Viewed as a node in a graph, the characteristics of a node are the information contained in its displacement vector. The directed edge connections between nodes are constructed based on the temporal order of adjacent trajectory points. For example, if trajectory point P... i The timestamp is later than the trajectory point P i-1 Then it will start from node Pi-1 To node P i A directed edge is constructed. This directed edge connection relationship reflects the time sequence and propagation direction of the ground deformation. By constructing such a graph structure, the graph attention network can better propagate and aggregate node features using the connection information between nodes, thereby mining the internal rules and trends of ground deformation.

[0058] Step S260: Aggregate the features of the multi-hop neighborhood nodes of each graph node feature to generate node aggregated features containing local deformation propagation trends.

[0059] This step further enhances the ability of the graph attention network to mine ground deformation information. In the graph attention network, each node has its neighborhood nodes, which are nodes directly connected to the node through directed edges. Multi-hop neighborhood nodes are nodes that can be reached through multiple edge connections. For example, for node P i , its one-hop neighborhood nodes are P i-1 and P i+1 , and its two-hop neighborhood nodes may be P i-2 and P i+2 , etc. The features of the multi-hop neighborhood nodes of each node are aggregated, and the aggregation method can use weighted summation. Specifically, for node P i , its node aggregated feature h i can be calculated by the following formula: where N(i) represents the multi-hop neighborhood node set of node P i , and a j is the attention weight of node P i relative to node P j , which represents the importance of the feature of node P i to node P i . Through this multi-hop neighborhood node feature aggregation, the node aggregated feature h i can contain more local deformation propagation trend information, such as how ground deformation gradually propagates from one region to adjacent regions.

[0060] Step S270: Input the node aggregated feature into the attention weight calculation layer of the graph attention network to calculate the attention weight value between each graph node and the graph node at the previous time; perform weighted summation on the node aggregated feature according to the attention weight value to generate a deformation displacement vector reflecting the deformation direction and intensity.

[0061] In the attention weight calculation layer of the graph attention network, the attention weight value between each graph node and the graph node at the previous time is calculated. The attention weight value represents the importance of the current node feature relative to the node feature at the previous time. For example, for node and its previous time node , attention weight value can be calculated by the following formula:

[0062] ;

[0063] where a is a learnable parameter vector, denotes concatenating the node aggregated features of node P i and node P j , and LeakyReLU is an activation function. After obtaining the attention weight value, the node aggregated features are weighted and summed according to the weight value. Assuming that the multi-hop neighborhood node set of node P i is N(i), the deformation displacement vector D i can be calculated by the following formula: This deformation displacement vector comprehensively considers the attention relationship between nodes and the local deformation propagation trend, and can accurately reflect the deformation direction and intensity of the ground at that moment, providing an important basis for road disaster warning.

[0064] As an implementation manner, in step S400, the vibration fluctuation feature, the deformation displacement vector, the humidity gradient feature, and the visual texture feature are fused in a multi-modal spatio-temporal manner to generate a road state fusion feature, including:

[0065] Step S410: performing a time alignment operation on the vibration fluctuation feature and the deformation displacement vector to generate a vibration-deformation correlation feature.

[0066] Step S410 aims to ensure that the two features are consistent in the time dimension, so as to accurately reflect the correlation between them. In the road monitoring scene, the vibration fluctuation feature and the deformation displacement vector are features that describe the road state from different angles. Due to factors such as the sampling frequency of the sensor, data transmission delay, etc., there may be inconsistencies in time. For example, the vibration sensor may collect data at a higher frequency, while the sampling frequency of the displacement sensor is relatively low, which leads to the fact that the time stamps of the vibration fluctuation feature and the deformation displacement vector may not completely correspond in the same time period. A time alignment operation is performed based on timestamp matching. First, the timestamp sequences of the vibration fluctuation feature and the deformation displacement vector are obtained. Assuming that the timestamp sequence of the vibration fluctuation feature is , and the timestamp sequence of the deformation displacement vector is . Then, the closest timestamp pairs are found by traversing the two timestamp sequences. For each timestamp t vi of the vibration fluctuation feature, the timestamp t dj with the smallest difference is found in the timestamp sequence of the deformation displacement vector.and the corresponding vibration fluctuation features and deformation displacement vectors are associated. In some cases, a completely matched timestamp cannot be found, in which case a linear interpolation method can be used to estimate the feature value at the corresponding time point. After completing the time alignment operation, the aligned vibration fluctuation features and deformation displacement vectors are combined to generate vibration-deformation correlation features. This correlation feature can reflect the mutual influence relationship between road vibration and ground deformation, for example, whether strong vibration will cause greater displacement change of the ground, providing more valuable information for subsequent disaster analysis.

[0067] Step S420: spatial interpolation processing is performed on the humidity gradient features to make the resolution of the humidity gradient features consistent with the spatial distribution of the visual texture features.

[0068] Step S420 is to eliminate the differences in spatial resolution of different features so as to perform effective feature fusion subsequently. The humidity gradient features and the visual texture features are obtained from different sensors, and their spatial resolutions can be different. For example, the humidity sensor can be distributed at certain specific positions of the road, and the spatial resolution of the humidity gradient features collected is low; while the visual texture features generated from the visible light image stream collected by the camera have a high spatial resolution. The spatial interpolation processing is performed on the humidity gradient features by using a bilinear interpolation method. The bilinear interpolation is a method of interpolation in a two-dimensional plane, which estimates the value of a to-be-interpolated point by using the values of four adjacent known points. The formula of the bilinear interpolation can be referred to the foregoing relevant introduction, which will not be described herein. According to the spatial distribution of the visual texture features, the positions that need to be interpolated are determined, and the humidity gradient values of these positions are calculated by using the bilinear interpolation formula. Through this spatial interpolation processing, the resolution of the humidity gradient features is improved to be consistent with the spatial distribution of the visual texture features, so that the two kinds of features have comparability in space, laying a foundation for subsequent cross-modal fusion.

[0069] Step S430: the vibration-deformation correlation features, the humidity gradient features and the visual texture features are input into the cross-modal fusion network, and the cross-modal fusion network comprises a channel attention module and a spatial attention module.

[0070] The purpose of the cross-modal fusion network is to organically combine features from different modalities, fully exploit the complementary information between them, and improve the judgment accuracy of the road state. The channel attention module is used to dynamically adjust the weights of different feature channels, because the importance of features in different channels to the road state may be different. For example, in the vibration-deformation correlation feature, some channels may better reflect the immediate vibration situation of the road, while other channels may be related to the long-term deformation trend of the ground. The spatial attention module is used to weight the importance of the spatial position of the feature map. In the visual texture feature, the features of the key areas such as the edge and crack of the road may better reflect the damage situation of the road. The vibration-deformation correlation feature, the humidity gradient feature after spatial interpolation processing, and the visual texture feature are input into the cross-modal fusion network at the same time. In the network, first, the three features are spliced in the channel dimension to form a feature tensor containing multiple types of information. Then, the feature tensor is processed by the channel attention module and the spatial attention module in turn. The channel attention module performs global average pooling operation on each channel of the feature tensor, compressing the feature information of each channel into a scalar. Then, through a fully connected layer and an activation function, the weight value of each channel is calculated. Finally, these weight values are multiplied with the original feature tensor channel by channel, realizing dynamic weighting of different channels. The spatial attention module performs spatial dimension operation on the feature tensor processed by the channel attention module. It performs maximum pooling and average pooling operations on the feature tensor, and then splices the two pooling results in the channel dimension. Then, through a convolution layer and an activation function, a spatial attention map is generated, which represents the importance of each spatial position in the feature map. Finally, the spatial attention map is multiplied with the feature tensor processed by the channel attention module element by element, realizing the importance weighting of the spatial position of the feature map. Through the processing of the channel attention module and the spatial attention module, the cross-modal fusion network can adaptively adjust the weights of different feature channels and spatial positions, highlight important information, and suppress irrelevant noise, thereby improving the effect of multi-modal feature fusion.

[0071] Step S440: dynamically allocate the channel weights of the multi-modal features through the channel attention module, and weight the importance of the spatial position of the feature map through the spatial attention module.

[0072] Step S440 is a further refinement of the specific functions of the channel attention module and the spatial attention module in step S430. In the channel attention module, the specific process of dynamically allocating the channel weights of the multi-modal features is as follows. Assume that the feature tensor input into the channel attention module is where C represents the number of channels, H and W represent the height and width of the feature map respectively. First, perform global average pooling operation on the spatial dimension of the feature tensor F to obtain a vector of length C , the calculation formula is: , wherein represents the element of the feature tensor F at the c-th channel, the i-th row, and the j-th column. Then, the vector z is input into a neural network containing two fully connected layers. The first fully connected layer reduces the dimension of the vector z from C to , wherein r is a reduction ratio factor, usually taking the value of 16. Then, a ReLU activation function is introduced to introduce nonlinearity. The second fully connected layer restores the dimension from to C, and maps the output value to the interval [0, 1] through a Sigmoid activation function, obtaining the channel weight vector . Finally, the channel weight vector w is multiplied with the original feature tensor F channel by channel, obtaining the feature tensor F' processed by the channel attention module, the calculation formula is: In the spatial attention module, the specific process of importance weighting of the spatial position of the feature map is as follows. Assuming that the feature tensor input into the spatial attention module is . First, the maximum pooling and average pooling operations are performed on the feature tensor F' in the channel dimension, obtaining two feature maps and . Then, the two feature maps are spliced in the channel dimension, obtaining a new feature map . Next, a convolution layer is used to reduce the number of channels of the feature map from 2 to 1, and a Sigmoid activation function is used to map the output value to the interval [0, 1], obtaining the spatial attention map . Finally, the spatial attention map M is multiplied with the feature tensor F' element by element, obtaining the feature tensor F'' processed by the spatial attention module, the calculation formula is: Through the processing of the channel attention module and the spatial attention module, the multi-modal features can be dynamically adjusted and weighted according to the importance of different feature channels and spatial positions, so as to realize more effective multi-modal feature fusion.

[0073] Step S450: The weighted multi-modal features are spliced in the channel dimension, and a road state fusion feature is generated through a dimension reduction convolution layer.

[0074] After being processed by the channel attention module and the spatial attention module, the multi-modal features have been weighted according to their channel importance and spatial position importance. At this time, these weighted features are spliced in the channel dimension. The feature information of different modalities is integrated into a unified feature tensor, so that the subsequent processing can comprehensively consider the information of multiple modalities.

[0075] After splicing, the spliced feature tensor is input into a dimension reduction convolutional layer. The main function of the dimension reduction convolutional layer is to reduce the dimension of the feature, remove redundant information, and retain features important for road state judgment. The dimension reduction convolutional layer is implemented through convolution operation, and the number of convolution kernels used is, for example, less than the number of channels of the spliced feature tensor. After processing by the dimension reduction convolutional layer, the number of channels of the feature tensor is reduced, and a feature representation with lower dimension and more compactness is obtained. This feature representation is the road state fusion feature, which integrates information from vibration, deformation, humidity, and vision, and can more comprehensively and accurately reflect the current state of the road.

[0076] As an implementation, the training process of the disaster prediction model includes the following steps:

[0077] Step S501: The road state fusion features in the historical road state data set are divided into continuous time period sequences in time sequence, and the corresponding disaster occurrence label is labeled at the end of each time period sequence.

[0078] In step S501, the road state fusion features in the historical road state data set are divided into continuous time period sequences in time sequence, and the corresponding disaster occurrence label is labeled at the end of each time period sequence, which prepares a suitable data structure for the training of the disaster prediction model. The historical road state data set contains a large number of road state fusion features, which are obtained by fusing multi-modal features such as vibration fluctuation features, deformation displacement vectors, humidity gradient features, and visual texture features, and reflect the comprehensive state of the road at different time points. These fusion features are divided according to time sequence to form continuous time period sequences. For example, assuming that the historical road state data set covers the road state information in a month, it can be divided into a sequence with one day as a time period. Each time period sequence contains a series of road state fusion features in the time period, and the corresponding disaster occurrence label is labeled at the end of the sequence. The disaster occurrence label is a binary classification label, such as "disaster occurs" marked as 1 and "no disaster occurs" marked as 0. This label is determined according to the actual disaster record, for example, if a landslide disaster occurs in a road area on a certain day, the label at the end of the corresponding time period sequence is 1. Through such division and labeling, the historical road state data is converted into input and output forms suitable for model training, providing a clear goal for subsequent model learning.

[0079] Step S502: A set proportion of road state fusion features at the front of the time period sequence is input into the bidirectional prediction network as a training set, and the bidirectional prediction network includes sequentially connected gated recurrent unit layers and self-attention mechanism layers.

[0080] Specifically, the time period sequence is divided into a training set and a test set according to a preset proportion, such as 70%. The road state fusion features of the first 70% of the time period sequence are selected as the training set for training the bidirectional prediction network. The bidirectional prediction network is a neural network capable of processing sequence data, which combines the advantages of a gated recurrent unit layer and a self-attention mechanism layer. The gated recurrent unit (GRU) is a variant of a recurrent neural network that controls the flow of information through a gating mechanism, which can effectively handle long-term dependencies in sequence data. The self-attention mechanism can automatically learn the relationship between different elements in the sequence and capture global information in the sequence. During training, the road state fusion features in the training set are input into the gated recurrent unit layer in time step order. The gated recurrent unit layer calculates the hidden state at the current time step based on the input at the current time step and the hidden state at the previous time step.

[0081] Step S503: In the pre-training phase of the bidirectional prediction network, random masking operation is performed on the road state fusion features in the training set to generate a binary mask matrix with the same dimension as the road state fusion features. The values of a certain proportion of positions in the binary mask matrix are set to zero, and the remaining positions are kept as 1. Then, the binary mask matrix and the original road state fusion features are multiplied element by element to obtain the masked features.

[0082] Random masking operation is a data augmentation method that can improve the robustness and generalization ability of the model. For example, a binary mask matrix with the same dimension as the road state fusion features is generated, and the elements in the matrix can only take 0 and 1 values. Then, according to a certain proportion, such as 30%, a certain proportion of positions in the binary mask matrix are randomly selected, and the values of these positions are set to zero, and the remaining positions are kept as 1. For example, for a road state fusion feature with a dimension of TxC (T is the time step length and C is the channel number), the generated binary mask matrix also has a dimension of TxC. Randomly set 30% of the elements to 0 and the remaining 70% of the elements to 1. Finally, the binary mask matrix and the original road state fusion features are multiplied element by element to obtain the masked features. In this way, the model needs to learn to recover the complete road state information from the partially visible features during training, thereby improving the model's ability to handle missing data and its understanding of data.

[0083] Step S504: The masked features are input into the gated recurrent unit layer in time step order, and the hidden state features at each time step are extracted in turn, and all the hidden state features at all time steps are concatenated into a time series feature.

[0084] In the bidirectional prediction network, the gated recurrent unit layer is the core component for processing sequence data. The covered features are input into the gated recurrent unit layer in time step order. At each time step, the gated recurrent unit layer calculates the hidden state of the current time step based on the input of the current time step and the hidden state of the previous time step. Through the control of the update gate and the reset gate, the gated recurrent unit can effectively capture the long-time dependency in the sequence data. The hidden state features calculated at each time step are saved in turn, and finally all the hidden state features of the time steps are spliced in the time dimension to form a complete time sequence feature. This time sequence feature contains the information of the road state at different time steps, providing input for the subsequent self-attention mechanism layer.

[0085] Step S505: input the time sequence feature into the self-attention mechanism layer, calculate the correlation weight matrix between the features of each time step in the time sequence feature, and generate the reconstructed complete road state fusion feature through weighted summation based on the correlation weight matrix.

[0086] The self-attention mechanism layer can automatically learn the correlation between different elements in the sequence. First, multiply the time sequence feature with three learnable weight matrices to get the query matrix Q, the key matrix K and the value matrix V. Then, calculate the dot product of the query matrix Q and the key matrix K, and divide by a scaling factor (d k is the dimension of the key matrix), to get the similarity score matrix. Next, convert the similarity score matrix into the correlation weight matrix through the Softmax function, so that the sum of each row of elements in the matrix is 1. Finally, multiply the correlation weight matrix with the value matrix V and sum it up to get the reconstructed complete road state fusion feature. In this way, the self-attention mechanism layer can capture the global correlation information between different time steps in the time sequence feature, and better reconstruct the complete road state fusion feature.

[0087] Step S506: input the reconstructed complete road state fusion feature and the original road state fusion feature into the mean square error calculation module, and output the feature reconstruction loss value.

[0088] Mean square error (MSE) is used to measure the difference between two vectors. Take the reconstructed complete road state fusion feature and the original road state fusion feature as input, and substitute them into the mean square error calculation formula: where y i is the i-th element in the original road state fusion feature, is the i-th element in the reconstructed complete road state fusion feature, and n is the length of the feature vector. The feature reconstruction loss value reflects the accuracy of the model in reconstructing the road state fusion feature in the pre-training stage. The smaller the loss value, the better the reconstruction effect of the model. By minimizing this loss value, the parameters of the bidirectional prediction network are optimized, enabling the model to better learn the feature patterns of the road state.

[0089] Step S507: In the optimization training phase of the bidirectional prediction network, the original road state fusion feature without covering is input into the trained gated recurrent unit layer, and the time series feature is output and the disaster probability prediction vector is generated through the self-attention mechanism layer.

[0090] After the pre-training phase ends, the model has learned certain road state feature patterns. In the optimization training phase, the original road state fusion feature without random covering operation is input into the trained gated recurrent unit layer in time step order. The gated recurrent unit layer calculates the hidden state feature of each time step in turn according to the calculation method described above, and splices all the hidden state features of the time steps into a time series feature. Then, the time series feature is input into the self-attention mechanism layer, which generates a disaster probability prediction vector by calculating the correlation weight matrix and weighted summation. Each element in this vector represents the occurrence probability of the corresponding disaster type, and the dimension of the vector is equal to the number of disaster types. For example, if there are three disaster types of landslide, collapse, and mudslide, the dimension of the disaster probability prediction vector is 3.

[0091] Step S508: Input the disaster probability prediction vector and the disaster occurrence label into the cross-entropy calculation module to output the classification loss value.

[0092] Cross-entropy is used to measure the difference between the probability distribution predicted by the model and the true label. Assuming that the disaster probability prediction vector is , where represents the predicted probability of the i-th disaster type, and the disaster occurrence label is , where is the true label (0 or 1) of the i-th disaster type. The calculation formula of the cross-entropy loss function is: The classification loss value reflects the accuracy of the model in disaster classification prediction. The smaller the loss value, the closer the prediction result of the model to the true label. By minimizing this loss value, the parameters of the bidirectional prediction network are further optimized, improving the prediction ability of the model for disaster occurrence.

[0093] Step S509: Add the feature reconstruction loss value and the classification loss value in a preset weight ratio to obtain a joint loss value. Update the network parameters of the gated recurrent unit layer and the self-attention mechanism layer through the backpropagation algorithm until the joint loss value stabilizes within a set threshold range.

[0094] First, according to the preset weight ratio, the weight of the feature reconstruction loss value is , and the weight of the classification loss value is , the feature reconstruction loss value and the classification loss value are weighted and superimposed to obtain the joint loss value: The joint loss value comprehensively considers the performance of the model in feature reconstruction and disaster classification prediction. Then, the back propagation algorithm is used to calculate the gradient of the joint loss value with respect to the network parameters of the gated recurrent unit layer and the self-attention mechanism layer. The back propagation algorithm starts from the joint loss value and gradually calculates the parameter gradient of each network layer through the chain rule. Then, according to the calculated gradient, the optimization algorithm (such as stochastic gradient descent) is used to update the network parameters. This process is repeated until the joint loss value stabilizes within the set threshold range. When the joint loss value is stable, it means that the model has converged and can better learn the relationship between road state and disaster occurrence. At this time, the training process ends and the trained disaster prediction model is obtained.

[0095] As an implementation manner, the random masking operation on the road state fusion features of the training set in the above step S503 can include the following implementation steps:

[0096] Step S5031: Randomly selecting two continuous time period sequences from the road state fusion features of the training set, respectively denoted as the first feature segment and the second feature segment.

[0097] In step S5031, two continuous time period sequences are randomly selected from the road state fusion features of the training set, respectively denoted as the first feature segment and the second feature segment. By randomly selecting different time period sequences, the diversity of the data and the generalization ability of the model can be increased. In the road disaster early warning scenario, the training set contains a large number of road state fusion features of different time periods, which reflect the comprehensive state of the road at different times. In a random manner, two continuous time period sequences are selected from the training set, for example, the training set covers the road state information within a month, and the time period sequence from the 10th day to the 15th day is randomly selected as the first feature segment, and the time period sequence from the 16th day to the 21st day is randomly selected as the second feature segment. This random selection method avoids the overfitting of the model to certain time periods caused by fixed selection of certain time periods, so that the model can learn more extensive road state change patterns.

[0098] Step S5032: Generating a horizontal masking template for the first feature segment: randomly masking a first preset proportion of continuous channel data in the channel dimension of each road state fusion feature, and retaining the original values of the remaining second preset proportion of channels.

[0099] This step aims to simulate the missing of data in the channel dimension, enhancing the robustness of the model to the missing of channel information. Assuming that the first preset proportion is 30% and the second preset proportion is 70%, for each road state fusion feature in the first feature segment, it has multiple channels, each of which represents different types of feature information, such as vibration, deformation, humidity, and other related features. Randomly select consecutive channels and set their data to zero to mask out this part of channel information. For example, a road state fusion feature has 10 channels, and the 3rd to 5th channels may be randomly selected for masking, so that the data of these three channels will not be used by the model in subsequent processing, while the original values of the remaining 7 channels are retained. In this way, the model needs to learn to recover the complete road state feature from the partially visible channel information during training, improving the processing capability for missing channel information.

[0100] Step S5033: generating a longitudinal masking template for the second feature segment: randomly masking the data of a third preset proportion of consecutive time nodes in the time step dimension of each road state fusion feature, and retaining the complete features of the remaining fourth preset proportion of time nodes.

[0101] This step simulates the missing of data in the time dimension, improving the adaptability of the model to the missing of time information. Assuming that the third preset proportion is 20% and the fourth preset proportion is 80%, for each road state fusion feature in the second feature segment, it has multiple time steps in time, and each time step records the road state information at that time. Randomly select consecutive time nodes and set the data of these time nodes to zero. For example, a road state fusion feature is recorded in 100 time steps, and the 20th to 39th time steps may be randomly selected for masking, so that the data of these 20 time steps will not be used by the model, while the complete features of the remaining 80 time steps are retained. In this way, the model can learn to infer the complete road state change trend from the information of part of the time steps, enhancing the processing capability for missing time information.

[0102] Step S5034: superimposing the horizontal masking template and the longitudinal masking template to generate a composite mask matrix, and modifying the values of the repeated masking areas in the composite mask matrix to a single masking state.

[0103] This step integrates the masking information in the channel dimension and the time dimension into a matrix. The horizontal masking template is to mask the data in the channel dimension, and the vertical masking template is to mask the data in the time dimension. Superimposing them can get a composite mask matrix that considers both the channel and time dimensions. During the superimposition process, there may be some areas that are marked as masked in both the horizontal and vertical masking templates. The values of these repeated masking areas are modified to single masking state to avoid over-masking. For example, in the composite mask matrix, a certain position is marked as masked in both the horizontal and vertical masking templates. The value of this position is uniformly processed and only one masking operation is performed to ensure reasonable processing of the data.

[0104] Step S5035: Apply the composite mask matrix to the first feature segment and the second feature segment respectively to generate two groups of masked features.

[0105] This step applies the composite mask matrix to the actual feature data to obtain the masked features. The composite mask matrix is multiplied element by element with the road state fusion features of the first feature segment, and the feature data corresponding to the positions of 0 in the matrix are set to zero, thereby realizing the masking operation of the first feature segment and obtaining the first group of masked features. Similarly, the composite mask matrix is multiplied element by element with the road state fusion features of the second feature segment to obtain the second group of masked features. In this way, the feature data of the first feature segment and the second feature segment are randomly masked in the channel and time dimensions, simulating the possible data missing situation in actual application and providing more challenging training data for the model.

[0106] Step S5036: Parallelly input the two groups of masked features into the gated recurrent unit layer to extract the first hidden state feature sequence and the second hidden state feature sequence respectively.

[0107] This step uses the gated recurrent unit layer to process the masked features and extract the hidden state information in the sequence. The gated recurrent unit (GRU) can calculate the hidden state at the current time according to the input sequence data and the hidden state at the previous time. The first group of masked features is input into the gated recurrent unit layer in time step order, and the gated recurrent unit layer calculates the hidden state at each time step in turn according to its internal calculation formula to form the first hidden state feature sequence. Similarly, the second group of masked features is input into the gated recurrent unit layer to obtain the second hidden state feature sequence. These two hidden state feature sequences contain the understanding and representation of the gated recurrent unit layer for road state features under the condition of partial data missing.

[0108] Step S5037: Alternately splice the first hidden state feature sequence and the second hidden state feature sequence in the time step dimension to generate a mixed time sequence feature.

[0109] This step fuses the hidden state feature sequences of the two different feature segments, increasing the complexity and diversity of the data. In the time step dimension, the first hidden state feature sequence and the second hidden state feature sequence are alternately arranged. For example, the first hidden state feature sequence has 10 time steps, and the second hidden state feature sequence also has 10 time steps. They are alternately spliced into a mixed time sequence feature with a length of 20, such as taking the first time step of the first hidden state feature sequence, the first time step of the second hidden state feature sequence, the second time step of the first hidden state feature sequence, and so on. Through this alternative splicing method, the mixed time sequence feature contains feature information of two different time periods, and the arrangement of the feature information is more complex, which helps the model to learn more rich feature patterns.

[0110] Step S5038: input the mixed time sequence feature into the self-attention mechanism layer, calculate the correlation degree score of each time step feature with the disaster occurrence label, and select the time step features with correlation degree scores higher than the set score to form a high-weight feature set.

[0111] This step uses the self-attention mechanism to mine the feature information closely related to disaster occurrence in the mixed time sequence feature. The self-attention mechanism can automatically learn the correlation between different elements in the sequence. The mixed time sequence feature is input into the self-attention mechanism layer, and the correlation between each time step feature and other time step features is obtained by calculating the attention weight matrix. Then, each time step feature is associated with the corresponding disaster occurrence label, and the correlation degree score is calculated. The correlation degree score can be obtained by calculating the similarity between the feature vector and the label vector, for example, using cosine similarity calculation. A score threshold can be set to filter out time step features with correlation degree scores higher than the threshold, and these features form a high-weight feature set. The features in the high-weight feature set have strong correlation with disaster occurrence.

[0112] Step S5039: input the high-weight feature set into the fully connected layer for dimension reduction processing to generate compact reconstruction features, and compare the compact reconstruction features with the original road state fusion features channel by channel.

[0113] The fully connected layer multiplies the input feature vector with a weight matrix and adds a bias to obtain an output vector. The high-weight feature set is input into the fully connected layer, and by adjusting the weight matrix of the fully connected layer, the dimension of the high-weight feature set is reduced to generate compact reconstructed features. This dimension reduction process can remove redundant information and extract key features. Then, the compact reconstructed features and the original road state fusion features are compared channel by channel in the channel dimension to calculate the difference value of each channel. For example, for a certain channel, the difference between the value of the compact reconstructed feature in the channel and the value of the original road state fusion feature in the channel is calculated. Through this comparison, the error area in the feature reconstruction process can be found.

[0114] Step S5040: According to the unmatching channel data in the comparison result, the feature reconstruction error area is located, and the channel position information of the error area is fed back to the composite mask matrix generation module to dynamically adjust the shielding proportion of the horizontal covering template and the vertical covering template in the subsequent training period.

[0115] This step dynamically adjusts the random covering operation according to the information of the feature reconstruction error area to improve the training effect of the model. After comparing the compact reconstructed features and the original road state fusion features channel by channel, the unmatching channel data is found, and the channel where these data are located is the feature reconstruction error area. The channel position information of these error areas is fed back to the composite mask matrix generation module. The composite mask matrix generation module dynamically adjusts the shielding proportion of the horizontal covering template and the vertical covering template in the subsequent training period according to this information. For example, if the error of a certain channel is large in the feature reconstruction process, the shielding proportion of this channel in the horizontal covering template may be reduced in the subsequent training to give the model more opportunities to learn the information of this channel, thereby improving the processing ability of the model for the information of this channel. Through this dynamic adjustment, the model can better adapt to the data missing situation of different channels and time dimensions, improve the generalization ability of the model, and improve the prediction accuracy of road disasters.

[0116] As an implementation manner, the determination process of the dynamically adjusted disaster threshold in step S500 can include the following steps:

[0117] Step S500a: Obtain the historical disaster occurrence probability data set and the actual disaster record of the current road area.

[0118] In the operation process of the road disaster early warning system, the disaster occurrence probability of the current road area is continuously recorded, which is calculated by the disaster prediction model according to the road state fusion features. At the same time, the actual disaster situation is recorded, including the time, type and other information of the disaster occurrence. For example, the disaster occurrence probability of the road area every month in the past year and the specific records of actual disasters such as landslides and collapses may be saved. By collecting these historical data, a historical disaster occurrence probability data set and actual disaster records are constructed, which reflect the disaster risk situation and actual disaster occurrence of the road area at different times.

[0119] Step S500b: Calculate the false positive rate and the false negative rate of different probability intervals according to the historical disaster occurrence probability data set.

[0120] This step is used to evaluate the accuracy of disaster warning under different probability thresholds. First, the historical disaster occurrence probability data set is divided into multiple probability intervals, for example, the probability range from 0 to 1 is divided into [0, 0.1), [0.1, 0.2), [0.2, 0.3) and other intervals. For each probability interval, the disaster warning situation in the interval is counted. The false positive rate refers to the proportion of the model predicting the occurrence of disasters but not actually occurring in the probability interval. Assuming that in the probability interval [0.2, 0.3), the model issued 100 disaster warnings, but only 20 disasters actually occurred, then the number of false positives is 80, and the false positive rate is 0.8. The false negative rate refers to the proportion of the model predicting no disaster but actually occurring in the probability interval. For example, 20 disasters actually occurred in the interval, but the model only warned 10 times, so the number of false negatives is 10, and the false negative rate is 0.5. By calculating the false positive rate and the false negative rate of different probability intervals, the reliability of disaster warning under different probability thresholds can be understood.

[0121] Step S500c: Construct an optimization function with the constraint condition that the false positive rate does not exceed the first threshold and the false negative rate does not exceed the second threshold.

[0122] This step converts the accuracy requirement of disaster warning into a mathematical optimization problem. The first threshold and the second threshold are pre-set upper limits of acceptable false positive rate and false negative rate, for example, the first threshold is set to 0.2 and the second threshold is set to 0.1, indicating that the maximum allowed false positive rate is 20% and the maximum allowed false negative rate is 10%. The goal of the optimization function is to find a suitable disaster threshold that can provide the most accurate disaster warning while meeting the false positive rate and false negative rate constraints. Assuming the disaster threshold is t, the false positive rate is FPR(t), and the false negative rate is FNR(t), the optimization function can be represented as: max f(t), where f(t) is an index related to the accuracy of disaster warning, such as the accuracy of the warning. At the same time, the constraint conditions are FPR(t) ≤ the first threshold and FNR(t) ≤ the second threshold. By constructing such an optimization function, the optimal disaster threshold can be found using mathematical methods.

[0123] Step S500d: Use the particle swarm optimization algorithm to solve the optimization function to obtain the dynamic disaster threshold corresponding to the current road area.

[0124] This step uses the search capability of the particle swarm optimization algorithm to find the optimal disaster threshold that meets the constraint conditions. Particle swarm optimization is a swarm intelligence-based optimization algorithm that simulates the group behavior of bird or fish swarms. In this algorithm, a group of particles is first initialized, each particle representing a possible disaster threshold t. Each particle has two attributes: position and velocity. The position represents the value of the disaster threshold, and the velocity represents the direction and speed of the particle in the search space. According to the optimization function and the constraint conditions, the fitness value of each particle is calculated, which represents the goodness or badness of the disaster threshold corresponding to the particle. Then, each particle in the particle swarm updates its own speed and position based on its own historical optimal position and the global optimal position of the group. In each iteration, the particles continuously move towards better positions until the termination condition is met, such as reaching the maximum number of iterations or the fitness value converging. Finally, the particle with the optimal fitness value is obtained, and its position corresponds to the dynamic disaster threshold of the current road area. For example, after multiple iterations, the particle swarm finds a disaster threshold t = 0.35 that meets the constraint conditions of not exceeding the first threshold in false positive rate and not exceeding the second threshold in false negative rate, and maximizes the value of the optimization function. Therefore, 0.35 is the dynamic disaster threshold for the current road area. In this way, the disaster threshold suitable for the current road area can be dynamically determined based on historical data and pre-set accuracy requirements, improving the accuracy and reliability of disaster warning.

[0125] As an implementation, in step S500, when the disaster occurrence probability exceeds the dynamically adjusted disaster threshold, a warning signal matching the disaster type is triggered, including:

[0126] Step S510: According to the numerical range of disaster occurrence probability, three consecutive warning level intervals are divided to generate a first-level warning signal, a second-level warning signal and a third-level warning signal;

[0127] Specifically, according to the disaster occurrence probability output by the disaster prediction model, different probability thresholds are set to divide the warning level intervals. For example, set the disaster occurrence probability in 0-0.3 as the third-level warning interval, 0.3-0.7 as the second-level warning interval, and 0.7-1 as the first-level warning interval. When the disaster occurrence probability is in 0-0.3, the third-level warning signal is generated, indicating that the possibility of disaster occurrence is low, but attention should be paid; when the probability is in 0.3-0.7, the second-level warning signal is generated, indicating that the possibility of disaster occurrence is medium, and monitoring and preparation measures should be strengthened; when the probability is in 0.7-1, the first-level warning signal is generated, indicating that the possibility of disaster occurrence is high, and immediate action must be taken. This grading warning mechanism can reasonably allocate resources and take appropriate response strategies according to the different degrees of disaster risk.

[0128] Step S520: When the first-level warning signal is triggered, the real-time disaster coordinates of the current road area and the disaster image frames in the visible light image stream are extracted, the disaster coordinates and the disaster image frames are sent to the road maintenance terminal, and the emergency communication link of the road maintenance terminal is activated;

[0129] When the disaster occurrence probability exceeds the lower limit of the first-level warning interval (such as 0.7), the real-time disaster coordinates of the current road area are quickly obtained from the geographic information system, which accurately indicates the specific location where the disaster may occur. At the same time, the disaster-related image frames are selected from the visible light image stream, which can directly show the road scene, such as road cracks and collapse. The disaster coordinates and disaster image frames are packaged into data information and transmitted to the road maintenance terminal through the network. The road maintenance terminal is usually a device used by road management departments or maintenance personnel. After receiving this information, they can quickly understand the specific location and on-site situation of the disaster. The emergency communication link of the road maintenance terminal is activated, which is a communication channel specially used for emergency situations to ensure that road maintenance personnel can timely communicate with relevant departments and personnel to coordinate the response to the disaster.

[0130] Step S530: Receive the emergency response state information fed back by the road maintenance terminal through the emergency communication link, and input the emergency response state information associated with the disaster coordinates into the warning signal continuous monitoring queue;

[0131] The road maintenance terminal receives the disaster coordinates and disaster image frames, evaluates the disaster situation, and takes appropriate emergency response measures. At the same time, the emergency response status information is fed back to the through the emergency communication link. The emergency response status information includes whether the rescue personnel have been dispatched, whether the rescue materials are ready, etc. After receiving this information, it is associated with the disaster coordinates, for example, each disaster coordinate is marked with the corresponding emergency response status. Then, the associated information is input into the early warning signal continuous monitoring queue. This queue is a data structure used to store and manage information related to early warning signals, facilitating continuous monitoring and tracking of early warning signals.

[0132] Step S540: When the disaster occurrence probability in the early warning signal continuous monitoring queue does not decrease for three consecutive monitoring periods, upgrade the first-level early warning signal to the second-level early warning signal;

[0133] Specifically, the disaster occurrence probability is monitored at certain time intervals (i.e. monitoring periods), and the monitoring results are updated to the early warning signal continuous monitoring queue. If the disaster occurrence probability remains in the first-level early warning interval for three consecutive monitoring periods without a downward trend, it indicates that the disaster risk has not been effectively mitigated, and may even worsen. At this time, the first-level early warning signal is upgraded to the second-level early warning signal to remind relevant personnel to take higher-level response measures. This dynamic early warning signal upgrade mechanism can adjust the warning level in a timely manner according to the changes in disaster situation, ensuring the effectiveness of response measures.

[0134] Step S550: When the second-level early warning signal is triggered, extract the traffic flow distribution data and vehicle position data set of the current road area, and generate a path avoidance recommendation vector based on the disaster coordinates and vehicle position data set;

[0135] Specifically, when the disaster occurrence probability is in the second-level early warning interval (e.g. 0.3-0.7), the traffic flow distribution data of the current road area is extracted from the traffic monitoring system, which reflects the traffic volume of different road segments on the road. At the same time, the vehicle position data set is obtained through the vehicle positioning system, which contains the real-time position information of each vehicle on the road. Combining the disaster coordinates and vehicle position data set, analyze the relative position relationship between each vehicle and the disaster occurrence location. For vehicles that may be affected by the disaster, plan a driving path that avoids the disaster area based on the traffic flow distribution data. The path avoidance recommendation vector is a vector containing information such as recommended driving direction and road segment, which indicates how the vehicle adjusts the driving route to avoid entering the disaster area. For example, the vehicle is recommended to turn right from the current position and enter another road to bypass the disaster point.

[0136] Step S560: push the path avoidance recommendation vector to the vehicle location data set corresponding to the vehicle terminal, and extract the traffic signal control nodes affected by the disaster coordinates in the traffic flow distribution data.

[0137] In step S560, the path avoidance recommendation vector is pushed to the vehicle location data set corresponding to the vehicle terminal, and the traffic signal control nodes affected by the disaster coordinates in the traffic flow distribution data are extracted. The generated path avoidance recommendation vector is sent to the vehicle location data set corresponding to the vehicle terminal through wireless communication technology. The vehicle terminal is usually a navigation device or a smart terminal installed in the vehicle, which displays the recommended driving path on the navigation interface after receiving the path avoidance recommendation vector, guiding the driver to avoid the disaster area. At the same time, the traffic signal control nodes affected by the disaster coordinates are identified from the traffic flow distribution data. These nodes are key positions in the traffic signal control system, such as traffic signal control devices at intersections. The traffic signal control nodes affected by the disaster may need to adjust the signal timing to optimize traffic flow and avoid congestion near the disaster area.

[0138] Step S570: adjust the phase timing parameters of the traffic signal control nodes according to the push coverage of the path avoidance recommendation vector, and synchronize the adjusted phase timing parameters to the traffic signal control system.

[0139] Specifically, the push coverage of the path avoidance recommendation vector is calculated, i.e. the proportion of the number of vehicles receiving the path avoidance recommendation vector to the total number of vehicles. If the push coverage is high, it means that most vehicles can receive the path avoidance recommendation, so the phase timing parameters of the traffic signal control nodes can be adjusted appropriately to coordinate the detour route of the vehicles. For example, increase the green light time of the intersection leading to the detour route, and reduce the green light time of the intersection leading to the disaster area. The adjusted phase timing parameters are sent to the traffic signal control system, which adjusts the phase and duration of the traffic signal according to the new parameters to optimize traffic flow. In this way, vehicles can be guided to avoid disaster areas while optimizing traffic signal control, reducing traffic congestion, and improving road safety and efficiency.

[0140] As an embodiment, the method provided by the embodiment of the present application can further include an online updating mechanism of the disaster prediction model, specifically, the method can include the following steps:

[0141] Step S580: after triggering the warning signal, continuously collect the vibration waveform sequence, ground deformation trajectory and visible light image stream of the current road area, and generate the post-warning road state change data set.

[0142] Specifically, when the disaster prediction model triggers an early warning signal, it means that the road may have a disaster risk, at which time the monitoring of the road status will be strengthened. Through the vibration sensors deployed in the road monitoring area, vibration waveform sequences are continuously collected, which reflect the vibration of the road after the early warning trigger, such as vibration caused by vehicle driving, geological activity, etc. Using displacement sensors, surface deformation trajectories are obtained, recording the position changes of the ground after the early warning, such as ground uplift, subsidence or horizontal displacement. At the same time, visible light image streams are collected through cameras and other devices, which intuitively show the real-time status of the road surface, such as whether there are cracks, collapses, etc. These collected data are integrated to generate a post-warning road status change dataset, which contains multi-aspect status information of the road within a period of time after the early warning trigger, providing a data basis for subsequent analysis of the trend of road status changes.

[0143] Step S590: Time stamp alignment of the post-warning road status change dataset and the triggered early warning signal type to generate a model incremental training dataset with early warning labels.

[0144] Each data in the post-warning road status change dataset has a corresponding time stamp, recording the specific time of data collection. Align these data with the triggered early warning signal type (such as landslide warning, collapse warning, etc.) by time stamp to ensure that each data point can accurately correspond to the corresponding early warning signal type. For example, if the early warning signal is triggered at a certain time, find the vibration waveform sequence, surface deformation trajectory and visible light image stream data collected before and after that time, and associate them with the early warning signal type. Through this time stamp alignment operation, each data point is added with an early warning label, indicating the early warning type corresponding to the data, thereby generating a model incremental training dataset with early warning labels. This dataset will be used as input for subsequent incremental training of the disaster prediction model, helping the model learn the characteristics of road status changes under different early warning types.

[0145] Step S5100: Every pre-set model update period, extract incremental training samples within the latest time window from the model incremental training dataset, and input the incremental training samples into the disaster prediction model for forward propagation.

[0146] Specifically, the embodiment of the present application pre-sets a model updating period, for example, model updating is performed once a week or once a month. When each updating period arrives, data in a recent time period, referred to as a time window, is extracted from the model incremental training dataset with warning labels. For example, data in the last week is selected as incremental training samples. These incremental training samples contain the latest road state information and corresponding warning labels. The incremental training samples are input into the disaster prediction model for forward propagation. Forward propagation refers to a process in which data starts from the input layer of the model, passes through each hidden layer for calculation in turn, and finally reaches the output layer to obtain a prediction result. In this process, the model processes the incremental training samples according to the current parameters and outputs the predicted disaster occurrence probability and type.

[0147] Step S5110: Calculate the incremental loss value between the prediction probability of the disaster prediction model for the incremental training samples and the warning label, and generate the model parameter fine-tuning gradient according to the incremental loss value.

[0148] Specifically, the prediction probability of the disaster prediction model for the incremental training samples is compared with the warning label, and the incremental loss value between the two is calculated using a loss function (such as a cross-entropy loss function). The incremental loss value reflects the prediction error of the model when processing the incremental training samples. According to the incremental loss value, the gradient of the model parameters is calculated using the backpropagation algorithm. The backpropagation algorithm calculates the partial derivative of the loss value with respect to each parameter of the model layer by layer from the loss value, and obtains the model parameter fine-tuning gradient. This gradient indicates the direction and magnitude of the adjustment of the model parameters, so as to reduce the incremental loss value and improve the prediction accuracy of the model.

[0149] As an implementation manner, the generation process of the model parameter fine-tuning gradient includes:

[0150] Step S5111: Select the top K samples with the largest loss value from the model incremental training dataset based on the incremental loss value, and generate a key sample set.

[0151] In step S5111, the top K samples with the largest loss values are selected from the model incremental training dataset based on the incremental loss values to generate a key sample set. This is a process of filtering the model incremental training dataset, and the purpose is to focus on samples that have a greater impact on the model performance. In the incremental training of the disaster prediction model, an incremental loss value is calculated for each sample, which reflects the difference between the model's prediction result and the true label. All samples in the model incremental training dataset are sorted in descending order of the incremental loss value. For example, in a model incremental training dataset containing 1000 samples, the incremental loss values of each sample are calculated and sorted in descending order. Then, the top K samples are selected, where K is a pre-set parameter, such as K = 100, i.e., the top 100 samples with the largest loss values are selected. These samples represent the part of the model's prediction results that are poor, and contain important information that the model may not have learned sufficiently. These samples are grouped into a key sample set, and the model will be adjusted for these samples in the future.

[0152] Step S5112: Freeze the dilated convolution kernel parameters of the time series feature encoder and the graph attention network parameters of the spatial feature encoder in the disaster prediction model, and keep the channel attention module parameters of the cross-modal fusion network in an updateable state.

[0153] In step S5112, the dilated convolution kernel parameters of the time series feature encoder and the graph attention network parameters of the spatial feature encoder in the disaster prediction model are frozen, and the channel attention module parameters of the cross-modal fusion network are kept in an updateable state, which is a limitation strategy for the update range of model parameters. In the disaster prediction model, the time series feature encoder extracts features from the vibration waveform sequence through the dilated convolution kernel, the spatial feature encoder processes the ground deformation trajectory data using the graph attention network, and the channel attention module of the cross-modal fusion network is used to dynamically allocate the channel weights of multi-modal features. Freezing the dilated convolution kernel parameters of the time series feature encoder and the graph attention network parameters of the spatial feature encoder means that these parameters will not be updated in the subsequent training process. This is because these modules have learned some stable feature representations in the previous training, and freezing them can avoid damaging these existing feature learning achievements in the incremental training process. The channel attention module parameters of the cross-modal fusion network are kept in an updateable state because this module is directly related to the fusion of multi-modal features, and adjusting its parameters can make the model better adapt to new incremental training data and optimize the fusion effect of multi-modal features. For example, when processing new road state data, the channel attention module can dynamically adjust the weights of different modal feature channels according to the characteristics of the data, improving the model's processing capability for new data.

[0154] Step S5113: input the key sample set into the frozen disaster prediction model, and extract the channel weight distribution output by the channel attention module in the cross-modal fusion network.

[0155] In step S5113, the key sample set is input into the frozen disaster prediction model, and the channel weight distribution output by the channel attention module in the cross-modal fusion network is extracted. This step is to obtain the feature information of the channel attention module when processing the key samples. The key sample set generated in step S5111 is input into the frozen model processed in step S5112. In the forward propagation process of the model, the channel attention module of the cross-modal fusion network processes the input multi-modal features and dynamically allocates channel weights according to the importance of the features. The channel weight distribution output by the channel attention module is recorded, which reflects the importance of different modal feature channels when processing the key samples. For example, when processing key samples containing vibration, deformation, humidity and visual features, the channel attention module may allocate a higher weight to the vibration-related channel, indicating that the vibration feature is more important for disaster prediction in these samples, while some humidity feature channels may be allocated a lower weight. By extracting the channel weight distribution, the focus of the model on different modal features when processing key samples can be understood, providing a basis for subsequent parameter adjustment.

[0156] Step S5114: calculate the gradient change of the channel attention module parameters relative to the incremental loss value in the direction of the incremental loss value of the channel weight distribution on the key sample set, and generate the initial fine-tuning gradient.

[0157] Specifically, first analyze the direction of the incremental loss value of the channel weight distribution on the key sample set, that is, observe how the incremental loss value changes when the channel weight changes. If increasing the weight of a certain channel leads to a decrease in the incremental loss value, it means that the weight of that channel should be increased; on the contrary, if increasing the weight of a certain channel leads to an increase in the incremental loss value, the weight of that channel should be decreased. The backpropagation algorithm is used to calculate the gradient change of the channel attention module parameters relative to the incremental loss value. The backpropagation algorithm is based on the chain rule, starting from the incremental loss value, and calculating the partial derivative of each parameter of the channel attention module with respect to the loss value layer by layer. These partial derivatives form the initial fine-tuning gradient, which indicates the direction and magnitude of the adjustment of the channel attention module parameters. For example, for a weight parameter in the channel attention module, the partial derivative of the parameter with respect to the incremental loss value is calculated by the backpropagation algorithm, and the partial derivative is the corresponding value of the parameter in the initial fine-tuning gradient. The initial fine-tuning gradient provides specific adjustment guidance for subsequent parameter update.

[0158] Step S5115: input the initial fine-tuning gradient into a sliding average filter to smooth the transient fluctuations of the initial fine-tuning gradient and generate a stabilized fine-tuning gradient.

[0159] Step S5115 is to improve the stability of the gradient. The initial fine-tuning gradient may have transient fluctuations, which may be caused by noise in the data or local instability of the model. If the initial fine-tuning gradient is directly used for parameter update, it may cause the model training process to be unstable, even oscillate. The sliding average filter is used to process the initial fine-tuning gradient. The basic principle of the sliding average filter is to weight and average the gradient values in a period of time, and the commonly used method is the exponential moving average (EMA) method. Assuming that the initial fine-tuning gradient sequence is , the calculation formula of the exponential moving average is , where v t is the sliding average gradient at time t, is a smoothing coefficient, usually taking a value between 0.9-0.99, and v0 = 0. In this way, the sliding average filter can smooth out the transient fluctuations in the initial fine-tuning gradient, making the gradient more stable. The generated stabilized fine-tuning gradient can provide more reliable guidance for the update of model parameters, and help the model converge more smoothly.

[0160] Step S5116: update the channel weight parameters of the channel attention module using the stabilized fine-tuning gradient to generate an updated channel attention module.

[0161] In step S5116, the channel weight parameters of the channel attention module are updated using the stabilized fine-tuning gradient to generate an updated channel attention module, which is the step of applying the calculated gradient to the model parameter update. The optimization algorithm (such as stochastic gradient descent) is used to update the channel weight parameters of the channel attention module according to the stabilized fine-tuning gradient. Assuming that a certain channel weight parameter of the channel attention module is w, and the value of the corresponding parameter in the stabilized fine-tuning gradient is , the formula for updating the parameter using the stochastic gradient descent algorithm is , where is the learning rate, which controls the step size of parameter update. According to this formula, all channel weight parameters of the channel attention module are updated to obtain the updated channel attention module. The updated channel attention module can better adapt to the features of the key sample set, improve the processing capability of the model for these samples, and thus improve the overall prediction performance of the model.

[0162] Step S5117: reconnect the updated channel attention module to the frozen time sequence feature encoder and spatial feature encoder to obtain a temporary disaster prediction model.

[0163] In step S5116, the channel weight parameters of the channel attention module have been updated, and the updated channel attention module is reconnected to the previously frozen temporal feature encoder and spatial feature encoder. Since the parameters of the temporal feature encoder and the spatial feature encoder remain frozen, a temporary disaster prediction model is formed. This temporary model adjusts the channel attention module of the cross-modal fusion network while maintaining the original feature extraction capability, aiming to improve the processing ability of the model for key samples and the overall disaster prediction performance. In this way, the model can be locally optimized without damaging the original model structure and feature learning results.

[0164] Step S5118: input the model incremental training dataset into the temporary disaster prediction model, and calculate the validation incremental loss value of the updated channel attention module under the unfrozen parameters.

[0165] Step S5118 is a verification step for the performance of the temporary disaster prediction model. The entire model incremental training dataset is input into the temporary disaster prediction model for forward propagation calculation. In this process, the model makes predictions based on the input data and calculates the incremental loss value between the prediction results and the true labels. This incremental loss value is called the validation incremental loss value, which reflects the performance of the updated channel attention module in processing the entire model incremental training dataset. Unlike the incremental loss value calculated on the key sample set before, the validation incremental loss value considers more data samples and can more comprehensively evaluate the improvement effect of the model. For example, if the validation incremental loss value decreases compared to before the update, it means that the updated channel attention module can better predict disaster situations when processing more data, and the performance of the model has been improved.

[0166] Step S5119: when the validation incremental loss value is less than the incremental loss value before the update, the temporary disaster prediction model is used as the current disaster prediction model; otherwise, the channel attention module parameters before the update are rolled back and the subsequent gradient update step is reduced.

[0167] Step S5119 is the evaluation of the model update effect and the adjustment strategy. The validation incremental loss value is compared with the incremental loss value before the update. If the validation incremental loss value is less than the incremental loss value before the update, it means that through the update of the channel attention module parameters, the performance of the model has been improved, at this time the temporary disaster prediction model is taken as the current disaster prediction model, and is continued to be used for subsequent disaster prediction tasks. If the validation incremental loss value is not less than the incremental loss value before the update, it means that this parameter update does not achieve the expected effect, and may even cause the model performance to decrease. In this case, the channel attention module parameters before the update are rolled back, and the model is restored to the state before the update. At the same time, the step of subsequent gradient update is reduced, for example, the learning rate is reduced to half of the original. Reducing the step can make the model more cautious in the subsequent parameter update process, avoiding the model being trapped in local optimum or having unstable performance due to too large update step. Through this evaluation and adjustment strategy, the model can be continuously optimized in the incremental training process, improving the accuracy and reliability of disaster prediction.

[0168] Step S5120: After the gradient clipping algorithm is used to constrain the magnitude of the model parameter fine-tuning gradient, the full connection layer weight parameters of the disaster prediction model are updated.

[0169] The gradient clipping algorithm is used to limit the magnitude of the model parameter fine-tuning gradient, preventing the problem of gradient explosion or gradient disappearance. In deep learning, the magnitude of the gradient may become very large or very small, causing unstable model training. The gradient clipping algorithm sets a threshold, and when the magnitude of the gradient exceeds the threshold, the gradient is scaled so that its magnitude does not exceed the threshold. For example, if the threshold is set to 1, and the gradient magnitude of a certain parameter is 2, it will be scaled to 1. After obtaining the clipped gradient, an optimization algorithm (such as stochastic gradient descent) is used to update the full connection layer weight parameters of the disaster prediction model. The full connection layer is the layer that connects the neurons in the model, and its weight parameters determine the mapping relationship between the input and the output. By updating the full connection layer weight parameters, the model can better adapt to new incremental training samples and improve the prediction ability of road disasters.

[0170] Step S5130: The updated disaster prediction model is evaluated for disaster identification accuracy on the historical validation set, and when the disaster identification accuracy exceeds the evaluation result of the original model, the updated disaster prediction model is deployed to the road monitoring area.

[0171] Specifically, the updated disaster prediction model is evaluated using a historical validation set, which is a set of pre-prepared road state data containing real disaster labels. The data in the historical validation set is input into the updated model to obtain the predicted disaster occurrence, and compared with the real disaster label to calculate the disaster identification accuracy. The accuracy is calculated, for example, by dividing the number of correctly predicted samples by the total number of samples. If the disaster identification accuracy of the updated model on the historical validation set exceeds the evaluation result of the original model, it indicates that the model has been improved through incremental training and can more accurately predict road disasters. At this time, the updated disaster prediction model is deployed to the road monitoring area to replace the original model for real-time road disaster warning work, so as to improve the accuracy and reliability of the warning.

[0172] The embodiment of the present application provides a computer system, and the computer system comprises one or more processors, a memory and one or more computer programs. Figure 2 As shown in the figure, the computer system 100 comprises a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, through a bus 102. Optionally, the computer system 100 can further comprise a transceiver 104. It should be noted that in actual application, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiment of the present application.

[0173] The embodiment of the present application provides a computer system, and the computer system comprises one or more processors, a memory and one or more computer programs. The one or more computer programs are stored in the memory and configured to be executed by the one or more processors. When the one or more programs are executed by the processor, the method provided above is implemented.

Claims

1. A road disaster early warning method based on an Internet of Things (IoT) sensing analysis, characterized by, The method comprises: Collecting real-time environmental data through a multi-modal sensing device deployed in a road monitoring area, the real-time environmental data including a vibration waveform sequence, a ground deformation trajectory, a humidity distribution map, and a visible light image stream; Inputting the vibration waveform sequence into a pre-trained time series feature encoder for waveform feature extraction to obtain vibration fluctuation features, and inputting the ground deformation trajectory into a spatial feature encoder for deformation trajectory analysis to generate a deformation displacement vector; Segmenting the humidity distribution map by region, extracting humidity gradient features of each segmented region, and dynamically sampling frames of the visible light image stream to obtain an image frame sequence and input the image frame sequence into an image feature encoder to extract visual texture features; Fusing the vibration fluctuation features, deformation displacement vector, humidity gradient features, and visual texture features to generate road state fusion features; Inputting the road state fusion features into a disaster prediction model to calculate a disaster probability, and triggering an early warning signal matching the disaster type when the disaster probability exceeds a dynamically adjusted disaster threshold; The pre-training process of the time series feature encoder comprises: Obtaining a historical vibration waveform data set, the historical vibration waveform data set including time series waveform data labeled with disaster types; Dividing the historical vibration waveform data set into multiple time window segments, and performing waveform amplitude normalization processing on each time window segment; Building a contrast learning network including multiple layers of dilated convolution kernels, inputting different time window segments of the same disaster type as positive sample pairs into the contrast learning network, and inputting time window segments of different disaster types as negative sample pairs into the contrast learning network; Calculating the waveform feature similarity of the positive sample pairs and the waveform feature difference of the negative sample pairs to generate a contrast loss value, and optimizing the parameters of the contrast learning network based on the contrast loss value; Outputting the features of the dilated convolution layer in the optimized contrast learning network as the vibration fluctuation feature extraction result of the time series feature encoder; The process of inputting the ground deformation trajectory into the spatial feature encoder for deformation trajectory analysis to generate a deformation displacement vector comprises: Sampling the ground deformation trajectory at equal intervals to extract three-dimensional coordinate data and corresponding time stamps of each trajectory point; Calculating the displacement change rate between adjacent trajectory points according to the time difference between the time stamp of each trajectory point and the time stamp of the previous trajectory point; Concatenating the three-dimensional coordinate data of each trajectory point with the corresponding displacement change rate to generate a displacement change vector for each trajectory point; Arranging the displacement change vectors of all trajectory points in chronological order into a displacement change sequence, and inputting the displacement change sequence into a graph attention network; In the graph attention network, taking the displacement change vector of each trajectory point as a graph node feature, and constructing a directed edge connection relationship between nodes according to the chronological order of adjacent trajectory points; Aggregating the features of the multi-hop neighborhood nodes of each graph node feature to generate node aggregation features containing local deformation propagation trends; The node aggregation features are input into an attention weight calculation layer of the graph attention network to calculate attention weight values between each graph node and a graph node at a previous moment; and the node aggregation features are weighted and summed according to the attention weight values to generate a deformation displacement vector reflecting deformation directions and intensities.

2. The road disaster early warning method based on the Internet of Things sensing analysis according to claim 1, characterized in that, The vibration fluctuation features, the deformation displacement vector, the humidity gradient features and the visual texture features are fused in a multi-modal spatio-temporal feature fusion manner to generate road state fusion features, including: The vibration fluctuation features and the deformation displacement vector are subjected to a time alignment operation to generate vibration-deformation correlation features; The humidity gradient features are subjected to a spatial interpolation processing to make the resolution of the humidity gradient features consistent with the spatial distribution of the visual texture features; The vibration-deformation correlation features, the humidity gradient features and the visual texture features are input into a cross-modal fusion network, the cross-modal fusion network comprising a channel attention module and a spatial attention module; The channel attention module is used to dynamically allocate channel weights of the multi-modal features, and the spatial attention module is used to weight the importance of spatial positions of feature maps; The weighted multi-modal features are subjected to channel splicing, and a dimension reduction convolution layer is used to generate the road state fusion features. 3.The road disaster early warning method based on the Internet of Things sensing analysis according to claim 1, wherein, The training process of the disaster prediction model comprises: Road state fusion features in a historical road state data set are sequentially divided into continuous time period sequences, and a corresponding disaster occurrence label is marked at the end of each time period sequence; A set proportion of road state fusion features at the front of the time period sequence is input into a bidirectional prediction network as a training set, the bidirectional prediction network comprising sequentially connected gated recurrent unit layers and self-attention mechanism layers; In a pre-training stage of the bidirectional prediction network, a random masking operation is performed on the road state fusion features of the training set to generate a binary mask matrix of the same dimension as the road state fusion features, the values of a set proportion of positions in the binary mask matrix are set to zero, and the values of the remaining positions are kept as 1, then the binary mask matrix and the original road state fusion features are multiplied element by element to obtain masked features; The masked features are input into the gated recurrent unit layers in a time step order to sequentially extract hidden state features at each time step, and all the hidden state features at the time steps are spliced into time sequence features; The time sequence features are input into the self-attention mechanism layers to calculate a correlation weight matrix between the features at the time steps in the time sequence features, and a reconstructed complete road state fusion feature is generated by weighted summation based on the correlation weight matrix; The reconstructed complete road state fusion feature and the original road state fusion feature are input into a mean square error calculation module to output a feature reconstruction loss value; In an optimization training stage of the bidirectional prediction network, the original road state fusion features without masking are input into the trained gated recurrent unit layers to output time sequence features and generate a disaster probability prediction vector through the self-attention mechanism layers; The disaster probability prediction vector and the disaster occurrence label are input into a cross-entropy calculation module to output a classification loss value. The feature reconstruction loss value and the classification loss value are superimposed in a preset weight ratio to obtain a joint loss value, and network parameters of the gated recurrent unit layer and the self-attention mechanism layer are updated through a back propagation algorithm until the joint loss value is stable within a set threshold range. 4.The road disaster early warning method based on the Internet of Things sensing analysis according to claim 3, wherein, The random covering operation on the road state fusion features of the training set comprises: Two continuous time period sequences are randomly selected from the road state fusion features of the training set, and are denoted as a first feature segment and a second feature segment respectively; A transverse covering template is generated for the first feature segment: a first preset proportion of continuous channel data is randomly masked in the channel dimension of each road state fusion feature, and the original values of the remaining second preset proportion of channels are retained; A longitudinal covering template is generated for the second feature segment: a third preset proportion of continuous time node data is randomly masked in the time step dimension of each road state fusion feature, and the complete features of the remaining fourth preset proportion of time nodes are retained; The transverse covering template and the longitudinal covering template are superimposed to generate a composite mask matrix, and the values of the repeated covering areas in the composite mask matrix are modified to single masking state; The composite mask matrix is respectively applied to the first feature segment and the second feature segment to generate two groups of covered features; The two groups of covered features are input into the gated recurrent unit layer in parallel to extract a first hidden state feature sequence and a second hidden state feature sequence respectively; The first hidden state feature sequence and the second hidden state feature sequence are alternately spliced in the time step dimension to generate a mixed time sequence feature; The mixed time sequence feature is input into the self-attention mechanism layer to calculate the correlation degree score of each time step feature and the disaster occurrence label, and a high-weight feature set is screened from the time step features with a correlation degree score higher than a set score; The high-weight feature set is input into a fully connected layer for dimension reduction processing to generate a compact reconstruction feature, and the compact reconstruction feature is compared with the original road state fusion feature channel by channel; According to the unmatching channel data in the comparison result, a feature reconstruction error area is located, and the channel position information of the error area is fed back to a composite mask matrix generation module to dynamically adjust the masking proportions of the transverse covering template and the longitudinal covering template in the subsequent training period. 5.The road disaster warning method based on the Internet of Things sensing analysis according to claim 1, wherein, The method for determining the dynamically adjusted disaster threshold value comprises: obtaining a historical disaster occurrence probability data set and actual disaster records of a current road area; calculating the false alarm rate and the missed alarm rate of different probability intervals according to the historical disaster occurrence probability data set; constructing an optimization function with the constraint condition that the false alarm rate is not more than a first threshold value and the missed alarm rate is not more than a second threshold value; solving the optimization function by using a particle swarm optimization algorithm to obtain a dynamically adjusted disaster threshold value corresponding to the current road area. 6.The road disaster early warning method based on the Internet of Things sensing analysis according to claim 1, wherein, The method for triggering a warning signal matched with the disaster type when the disaster occurrence probability exceeds the dynamically adjusted disaster threshold value comprises: dividing three continuous warning level intervals according to the numerical range of the disaster occurrence probability to generate a first-level warning signal, a second-level warning signal and a third-level warning signal. When the primary early warning signal is triggered, real-time disaster coordinates and disaster image frames in the visible light image stream of the current road area are extracted, the disaster coordinates and disaster image frames are sent to a road maintenance terminal, and an emergency communication link of the road maintenance terminal is activated; Through the emergency communication link, emergency response state information fed back by the road maintenance terminal is received, the emergency response state information is associated with the disaster coordinates, and the associated information is input into an early warning signal continuous monitoring queue; When the disaster occurrence probability in the early warning signal continuous monitoring queue does not decrease for three consecutive monitoring periods, the primary early warning signal is upgraded to a secondary early warning signal; When the secondary early warning signal is triggered, traffic flow distribution data and a vehicle position data set of the current road area are extracted, a path avoidance suggestion vector is generated according to the disaster coordinates and the vehicle position data set; The path avoidance suggestion vector is pushed to a vehicle terminal corresponding to the vehicle position data set, and a traffic signal control node affected by the disaster coordinates in the traffic flow distribution data is extracted; According to the push coverage rate of the path avoidance suggestion vector, the phase timing parameters of the traffic signal control node are adjusted, and the adjusted phase timing parameters are synchronized to a traffic signal control system. 7.The road disaster early warning method based on the Internet of Things sensing analysis according to claim 2, wherein, The method further includes an online updating mechanism of a disaster prediction model: After triggering an early warning signal, vibration waveform sequences, ground deformation trajectories and visible light image streams of the current road area are continuously collected, and a post-warning road state change data set is generated; The post-warning road state change data set is time-stamped aligned with the type of the triggered early warning signal, and a model incremental training data set with a warning label is generated; Every other preset model updating period, incremental training samples in a recent time window are extracted from the model incremental training data set, and the incremental training samples are input into the disaster prediction model for forward propagation; The incremental loss value between the prediction probability of the disaster prediction model for the incremental training samples and the warning label is calculated, a model parameter fine-tuning gradient is generated according to the incremental loss value; After the model parameter fine-tuning gradient is amplitude-constrained by using a gradient clipping algorithm, the full connection layer weight parameters of the disaster prediction model are updated; The updated disaster prediction model is used to evaluate the disaster identification accuracy on a historical verification set, and when the disaster identification accuracy exceeds the evaluation result of the original model, the updated disaster prediction model is deployed to a road monitoring area.

8. A computer system, characterized by Comprise: One or more processors; Memory; One or more computer programs; Wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, when the one or more computer programs are executed by the processor, the method of any one of claims 1-7 is implemented.

Images