Pavement collapse risk prediction method and system fusing ground penetrating radar and traffic load

Abstract

The application discloses a pavement collapse risk prediction method and system fusing geological radar and traffic load, and belongs to the field of pavement collapse risk prediction. The method fuses an initial collapse risk prediction value obtained after prediction by a trained random forest model with a feature difference change value by collecting real-time remote sensing images of a to-be-measured pavement and identifying the feature difference change value, and then uses a gradient boosting decision tree model which can capture the complex nonlinear relationship between geological data and traffic load data to predict the pavement collapse risk of multi-source data composed of the geological data and the traffic load data, thereby solving the problem that the prior art cannot capture the complex nonlinear relationship between geological data and traffic load data due to the dependence on a single data source and the use of simple linear regression or empirical formula in the prediction model, and cannot accurately predict the pavement collapse risk.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of road surface collapse risk prediction, in particular to a road surface collapse risk prediction method and system fusing geological radar and traffic load. BACKGROUND

[0002] With the acceleration of urbanization and the continuous growth of traffic flow, road infrastructure is facing increasingly severe challenges, and road surface collapse accidents occur frequently, which has a great impact on traffic safety, urban operation and residents' life. In order to ensure road safety and stability, road surface collapse risk prediction has become a key research direction in the field of road engineering, and its importance is increasingly prominent, which has attracted widespread attention from the academic and engineering circles. From the development trend, multi-source data fusion and intelligent technology application have become an important way to improve the prediction accuracy and reliability, which is expected to provide more scientific and effective decision support for road maintenance management.

[0003] In the prior art, road surface collapse risk prediction mostly relies on a single data source, such as only using geological data obtained by geological exploration or simply relying on traffic flow monitoring data for evaluation. When only using geological data, it is impossible to reflect the dynamic influence of traffic load on the road surface in real time, such as sudden changes in traffic flow, fluctuations in road pressure caused by concentrated driving of heavy vehicles, etc. While relying only on traffic load data, the potential role of geological conditions is ignored, such as changes in soil mechanical properties caused by changes in soil moisture, gradual expansion of soil cavities, etc. In addition, most existing prediction models use simple linear regression or empirical formula, which is difficult to capture the complex nonlinear relationship between geological data and traffic load data.

[0004] Therefore, the prior art has the problem of inaccurate road surface collapse risk prediction due to the reliance on a single data source and the use of simple linear regression or empirical formula in the prediction model, which cannot capture the complex nonlinear relationship between geological data and traffic load data. SUMMARY

[0005] The present application provides a road surface collapse risk prediction method and system fusing geological radar and traffic load, which can solve the problem of inaccurate road surface collapse risk prediction due to the reliance on a single data source and the use of simple linear regression or empirical formula in the prediction model, which cannot capture the complex nonlinear relationship between geological data and traffic load data, by fusing the initial collapse risk prediction value obtained by predicting the feature difference change value through the trained random forest model with the feature difference change value obtained by identifying the feature difference change value from the real-time remote sensing image of the road to be measured.

[0006] To solve the above technical problems, an embodiment of the present application provides a road surface collapse risk prediction method fusing geological radar and traffic load, comprising:

[0007] obtain geological data and traffic load data of the to-be-tested road surface at the current moment; wherein, the geological data includes soil humidity value and soil cavity value; the traffic load data includes traffic flow and road surface pressure;

[0008] input the geological data and traffic load data of the to-be-tested road surface at the current moment into the trained random forest model, so that the trained random forest model performs road surface collapse risk prediction according to the geological data and traffic load data of the to-be-tested road surface at the current moment, and obtains an initial collapse risk prediction value;

[0009] collect real-time remote sensing images of the to-be-tested road surface, perform horizontal feature extraction on the real-time remote sensing images, identify feature difference change values on the same horizontal plane in the real-time remote sensing images, and perform fusion processing on the initial collapse risk prediction value according to the feature difference change values, to obtain a first road surface collapse risk prediction value;

[0010] input the geological data and traffic load data of the to-be-tested road surface at the current moment into the trained gradient boosting decision tree model, so that the trained gradient boosting decision tree model performs road surface collapse risk prediction according to the geological data and traffic load data of the to-be-tested road surface at the current moment, to obtain a second road surface collapse risk prediction value;

[0011] calculate a final road surface collapse risk prediction value from the first road surface collapse risk prediction value and the second road surface collapse risk prediction value;

[0012] compare the final road surface collapse risk prediction value with a preset risk level interval, determine a road surface collapse risk level, and regulate the to-be-tested road surface according to the road surface collapse risk level.

[0013] Further, the step of collecting real-time remote sensing images of the to-be-tested road surface, performing horizontal feature extraction on the real-time remote sensing images, and identifying feature difference change values on the same horizontal plane in the real-time remote sensing images specifically includes:

[0014] perform grid processing on the real-time remote sensing images, and extract horizontal direction features in the grid images;

[0015] determine a circumcircle of each horizontal direction feature according to the position of each horizontal direction feature in the grid, to obtain coordinate positions of circumcircle points of each horizontal direction feature;

[0016] determine a spatial distance value between two circumcircle points as a feature difference change value on the same horizontal plane according to the coordinate positions of the two circumcircle points.

[0017] Further, the step of fusing the initial collapse risk prediction value according to the feature difference change value to obtain a first road collapse risk prediction value specifically comprises:

[0018] According to the initial collapse risk prediction value, determine the collapse risk point of the road to be tested in the real-time remote sensing image, and determine the coordinate position of the collapse risk point;

[0019] Calculate the spatial distance value between the coordinate position of each collapse risk point and the coordinate position of any excenter point as a reference change value;

[0020] Superimpose and fuse the reference change value and the feature difference change value in space distance, and take the value exceeding the preset threshold after superimposition and fusion as the first road collapse risk prediction value.

[0021] Further, the trained random forest model comprises a plurality of decision trees;

[0022] The trained random forest model performs road collapse risk prediction according to the geological data and traffic load data of the road to be tested at the current time to obtain a first road collapse risk prediction value, comprising:

[0023] Input the soil moisture value, soil cavity value, traffic flow and road pressure of the road to be tested at the current time into each decision tree of the trained random forest model;

[0024] For each decision tree, recursively traverse the soil moisture value, soil cavity value, traffic flow and road pressure using the built-in feature splitting rule to determine the target subnode, and take the collapse risk prediction value stored in the target subnode as the collapse risk prediction output value of the decision tree; wherein the collapse risk prediction value stored in the target subnode is determined by calculating the average value of the actual collapse risk values of all training samples falling into the subnode during the training process of the random forest model;

[0025] According to the collapse risk prediction output value of each decision tree, calculate the average value of the collapse risk prediction output value, and take the average value of the collapse risk prediction output value as the first road collapse risk prediction value.

[0026] Further, the model training of the random forest model comprises:

[0027] Obtain the soil moisture value, soil cavity value, traffic flow data, road pressure and corresponding ground collapse actual risk value of different road sections at different times to generate a training set and a validation set;

[0028] According to the training set, repeatedly perform the decision tree construction operation until the number of decision trees reaches a first threshold, to obtain a trained random forest model;

[0029] According to the verification set, the performance index evaluation calculation is performed on the trained random forest model, and the hyperparameters of the trained random forest model are optimized according to the performance evaluation index, to obtain the trained random forest model;

[0030] The decision tree construction operation includes:

[0031] For the root node of each decision tree, a plurality of samples are randomly selected from the training set with replacement, and a plurality of features are randomly selected as a plurality of candidate split features;

[0032] According to the candidate split features of the root node, the sample data of the root node is split to obtain the sample split results of each candidate split feature, and the mean square error of the sample split results of each candidate split feature is calculated according to the ground collapse actual risk value of the sample split results of each candidate split feature. The candidate split feature with the minimum mean square error is selected as the optimal split feature of the root node;

[0033] According to the optimal split feature, the sample data of the root node is split to generate a child node and sample data of the child node;

[0034] The candidate split feature selection, optimal split feature determination and sample data splitting operations are repeated for the child node until the depth of the current decision tree reaches the preset depth threshold, to obtain a single decision tree.

[0035] Further, the trained gradient boosting decision tree model includes a plurality of decision trees;

[0036] The trained gradient boosting decision tree model performs road collapse risk prediction according to the geological data and traffic load data of the road to be measured at the current time to obtain a second road collapse risk prediction value, including:

[0037] The soil moisture value, soil cavity value, traffic flow and road pressure of the road to be measured at the current time are input into each decision tree of the trained gradient boosting decision tree model;

[0038] For each decision tree of the gradient boosting decision tree model, the soil moisture value, soil cavity value, traffic flow and road pressure are recursively traversed using the built-in feature split rule to determine the target child node, and the residual prediction value stored in the target child node is taken as the output value of the decision tree. The residual prediction value is the residual mean value of the collapse risk prediction value and the collapse risk actual value of the sample falling into the node in the previous decision tree in the training stage of the gradient boosting decision tree model;

[0039] According to the output value of each decision of the gradient boosting decision tree model, the second road collapse risk prediction value is calculated.

[0040] Further, the calculation formula of the second road surface collapse risk prediction value is:

[0041] ;

[0042] Wherein, is the second road surface collapse risk prediction value; X is a feature composed of the soil humidity value, the soil cavity value, the traffic flow and the road pressure of the current moment of the to-be-tested road surface, X=[soil humidity value, soil cavity value, traffic flow, road pressure]; is the initial collapse risk prediction value, which is the mean value of the actual collapse risk values of all training samples in the gradient boosting decision tree model training stage; is the preset learning rate of the gradient boosting decision tree model; M is the total number of decision trees; is the output value of the mth decision tree, that is, when X falls in the mth decision tree, the residual prediction value stored in the node.

[0043] Further, the model training of the gradient boosting decision tree model comprises:

[0044] Obtain the soil humidity value, soil cavity value, traffic flow data, road pressure and corresponding ground collapse actual risk value of different road sections at different moments to generate a training set and a validation set;

[0045] According to the training set, perform decision tree iteration construction operation until the iteration number reaches the second threshold value, to obtain the trained gradient boosting decision tree model;

[0046] According to the validation set, perform performance index evaluation calculation on the trained gradient boosting decision tree model, and according to the performance evaluation index, optimize the hyperparameters of the trained gradient boosting decision tree model to obtain the trained gradient boosting decision tree model;

[0047] Wherein, the decision tree iteration construction operation comprises:

[0048] Calculate the mean value of the actual collapse risk values of all samples in the training set, and take the mean value as the initial collapse risk prediction value;

[0049] For each decision tree, calculate the residual value of each sample in the current decision tree according to the collapse risk prediction value of each sample in the previous decision tree and the actual collapse risk value of each sample; wherein, when the first decision tree is constructed, the collapse risk prediction value of each sample in the previous decision tree is the initial collapse risk prediction value;

[0050] For each to-be-split node of the decision tree, repeatedly perform node splitting operation until the depth of the current decision tree reaches the preset depth threshold value to obtain the current decision tree;

[0051] According to a feature split rule of the current decision tree, a collapse risk prediction value update calculation is performed on each sample of the current decision tree to obtain a collapse risk prediction value of each sample of the current decision tree;

[0052] The node split operation includes:

[0053] A plurality of features are randomly selected as a plurality of candidate split features, and the node samples are classified for each candidate feature to obtain a child node sample split result of each candidate split feature; the child node sample split result is a mean value of sample residual error values in the child node, and the mean value of the sample residual error values in the child node is taken as a residual prediction value of the node.

[0054] According to the child node sample split results of the candidate split features, a mean square error of the child node sample split results of the candidate split features is calculated, and a candidate split feature with the minimum mean square error is taken as an optimal split feature, and then a feature split rule is determined according to the optimal split feature.

[0055] Further, the calculation formula of the final road collapse risk prediction value is:

[0056] ;

[0057] Wherein, is the final road collapse risk prediction value; is the first road collapse risk prediction value; is the second road collapse risk prediction value; X is composed of the soil humidity value, the soil cavity value, the traffic flow and the road pressure at the current time of the road to be measured, X=[soil humidity value, soil cavity value, traffic flow, road pressure]; and are the first risk weight and the second risk weight, respectively.

[0058] Further, the comparison of the final road collapse risk prediction value with the preset risk level interval to determine the road collapse risk level includes:

[0059] The final road collapse risk prediction value is compared with the upper limit index and the lower limit index of the preset risk level interval to determine the road collapse risk level; wherein, the preset risk level interval includes: a low risk level interval, a medium risk level interval and a high risk level interval; the upper limit index of the low risk level interval is less than the lower limit index of the medium risk level interval; the upper limit index of the medium risk level interval is less than the lower limit index of the high risk level interval.

[0060] On the basis of the above method embodiment, the present application provides a system embodiment;

[0061] An embodiment of the present application provides a pavement collapse risk prediction system fusing ground penetrating radar and traffic load, comprising: a data acquisition module, an initial risk prediction module, a first risk prediction module, a second risk prediction module, a final risk prediction module and a pavement risk monitoring module.

[0062] The data acquisition module is configured to acquire geological data and traffic load data of the pavement at a current time, wherein the geological data comprises soil humidity values and soil cavity values, and the traffic load data comprises traffic flow and road pressure.

[0063] The initial risk prediction module is configured to input the geological data and the traffic load data of the pavement at the current time into a trained random forest model, so that the trained random forest model performs pavement collapse risk prediction according to the geological data and the traffic load data of the pavement at the current time, and obtains an initial collapse risk prediction value.

[0064] The first risk prediction module is configured to collect a real-time remote sensing image of the pavement, perform horizontal feature extraction on the real-time remote sensing image, identify a feature difference change value in the real-time remote sensing image on the same horizontal plane, and perform fusion processing on the initial collapse risk prediction value according to the feature difference change value, to obtain a first pavement collapse risk prediction value.

[0065] The second risk prediction module is configured to input the geological data and the traffic load data of the pavement at the current time into a trained gradient boosting decision tree model, so that the trained gradient boosting decision tree model performs pavement collapse risk prediction according to the geological data and the traffic load data of the pavement at the current time, and obtains a second pavement collapse risk prediction value.

[0066] The final risk prediction module is configured to calculate a final pavement collapse risk prediction value from the first pavement collapse risk prediction value and the second pavement collapse risk prediction value.

[0067] The pavement risk monitoring module is configured to compare the final pavement collapse risk prediction value with a preset risk level interval, determine a pavement collapse risk level, and regulate the pavement according to the pavement collapse risk level.

[0068] Compared with the prior art, the embodiment of the present application has the following beneficial effects:

[0069] The application can obtain the first road surface collapse risk prediction value and the second road surface collapse risk prediction value by collecting real-time remote sensing images of the to-be-measured road surface, identifying the feature difference change value, performing fusion processing on the initial collapse risk prediction value obtained after prediction by the trained random forest model and the feature difference change value, and using the gradient boosting decision tree model which can capture the complex nonlinear relationship between geological data and traffic load data, and then calculating the final road surface collapse risk prediction value. By using the trained random forest model and the gradient boosting decision tree model which can capture the complex nonlinear relationship between geological data and traffic load data, the multi-source data composed of geological data and traffic load data is predicted for road surface collapse risk, which not only avoids the limitation of relying on a single data source and ignoring other influencing factors, but also avoids the problem that the prediction model only uses simple linear regression or empirical formula and cannot capture the complex nonlinear relationship between geological data and traffic load data, improves the accuracy of road surface collapse risk prediction, and solves the problem that the existing technology cannot capture the complex nonlinear relationship between geological data and traffic load data due to the dependence on a single data source and the use of simple linear regression or empirical formula in the prediction model, resulting in inaccurate road surface collapse risk prediction. In addition, the final prediction value is compared with the preset risk level interval to determine the risk level and regulate the to-be-measured road surface accordingly, which further enhances the effectiveness of risk prevention and control and provides a strong guarantee for road safety. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 A flowchart of a road surface collapse risk prediction method fusing geological radar and traffic load provided by the embodiment of the application is shown in the figure.

[0071] Figure 2 A module diagram of a road surface collapse risk prediction system fusing geological radar and traffic load provided by the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0073] In the description of the application, it should be understood that the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features.

[0074] Embodiment 1:

[0075] Reference Figure 1The flowchart of the pavement collapse risk prediction method combining ground penetrating radar and traffic load provided by the embodiment of the present application; in order to solve the problem that the prior art cannot capture the complex nonlinear relationship between geological data and traffic load data due to the dependence on a single data source and the use of simple linear regression or empirical formula in the prediction model, resulting in inaccurate prediction of pavement collapse risk, the method comprises at least the following steps:

[0076] Step S1: acquiring geological data and traffic load data of the to-be-tested pavement at the current time; wherein the geological data includes soil humidity value and soil cavity value; the traffic load data includes vehicle flow and road pressure.

[0077] In the embodiment, the soil humidity value can be obtained by using soil humidity sensors buried at different depths under the to-be-tested pavement. For example, high-precision time domain reflectometry (TDR) sensors are installed at 20 cm, 50 cm and 80 cm below the pavement. These sensors calculate the dielectric constant of the soil by transmitting electromagnetic pulses and measuring their propagation time in the soil, and then obtain the soil humidity value.

[0078] In the embodiment, the soil cavity value can be detected by using a ground penetrating radar device. The ground penetrating radar transmits high-frequency electromagnetic waves to the underground, and when the electromagnetic waves encounter abnormal structures such as soil cavities, reflected waves are generated. By receiving and analyzing the characteristics of these reflected waves, such as the intensity of the reflected waves, the two-way travel time, etc., combined with professional radar image processing software, the position, size and shape of the soil cavity can be determined. The operator moves the ground penetrating radar antenna at a certain interval and speed to scan the to-be-tested pavement, and the scanning data is transmitted to the computer connected thereto in real time, and the soil cavity value related data is obtained after software processing.

[0079] In the embodiment, the vehicle flow can be obtained by installing vehicle detection devices such as geomagnetic sensors or video monitoring cameras at the entrance and exit of the to-be-tested pavement. The geomagnetic sensor detects the passage of vehicles by sensing the change of the earth's magnetic field caused by the passage of vehicles, transmits the detection signal to the counter, and counts the number of vehicles passing through in a unit of time, thereby obtaining the vehicle flow data. The video monitoring camera uses image recognition technology to analyze the video pictures in real time, recognizes and counts the number of passing vehicles. The camera transmits the video data to the background server, and the vehicle recognition software running on the server processes the data and outputs the vehicle flow information.

[0080] In the embodiment, the road surface pressure can be laid with pressure sensors, such as piezoelectric pressure sensors, in the road surface structure layer. When a vehicle travels on the road surface, the pressure sensor is subjected to pressure to generate an electric signal, and the size of the electric signal is proportional to the pressure. The pressure sensor transmits the electric signal to a data acquisition system, which processes the signal through amplification, filtering, etc., and calculates the pressure value of the road surface according to the pre-calibrated pressure-electric signal relationship.

[0081] Step S2: input the geological data and traffic load data of the to-be-tested road surface at the current time into the trained random forest model, so that the trained random forest model performs road surface collapse risk prediction according to the geological data and traffic load data of the to-be-tested road surface at the current time, and obtains an initial collapse risk prediction value.

[0082] In the embodiment, the trained random forest model includes a plurality of decision trees; the trained random forest model performs road surface collapse risk prediction according to the geological data and traffic load data of the to-be-tested road surface at the current time, and obtains an initial collapse risk prediction value, including:

[0083] The soil humidity value, soil cavity value, vehicle flow and road surface pressure of the to-be-tested road surface at the current time are input into each decision tree of the trained random forest model;

[0084] For each decision tree, the soil humidity value, soil cavity value, vehicle flow and road surface pressure are recursively traversed using the built-in feature splitting rule to determine a target sub-node, and the collapse risk prediction value stored in the target sub-node is taken as the collapse risk prediction output value of the decision tree; wherein the collapse risk prediction value stored in the target sub-node is determined by calculating the average value of the collapse risk actual values of all training samples falling into the sub-node during the training process of the random forest model.

[0085] According to the collapse risk prediction output value of each decision tree, the average value of the collapse risk prediction output values is calculated, and the average value of the collapse risk prediction output values is taken as the initial collapse risk prediction value.

[0086] For example, the trained random forest model includes 50 decision trees;

[0087] For a single decision tree, the soil humidity value (such as 25%), the soil cavity value (such as 0.2 cubic meters), the traffic flow (such as 1000 vehicles / hour during the morning rush hour), and the road surface pressure (such as 50000 pascals when a large vehicle passes) are input into the decision tree, and recursive traversal is performed according to the built-in feature splitting rules in the decision tree. For example, the feature splitting rule may be to first judge according to the soil humidity value. If the soil humidity value is less than 25%, go to the left child tree, otherwise go to the right child tree. In the child tree, further split according to the traffic flow, etc. After a series of recursive traversals, the target child node is determined. Assuming that the collapse risk prediction value stored in this target child node is 0.3 (this value is determined by calculating the average of the actual collapse risk values of all training samples falling into the child node during the model training process).

[0088] According to the method of calculating the collapse risk prediction output value of a single decision tree, the collapse risk prediction output values of 50 decision trees are calculated respectively. For example, the collapse risk prediction output values of the 50 decision trees are [0.3, 0.25, 0.32, 0.28, …], the average of these collapse risk prediction output values is calculated by adding all the values and dividing by 50, and the initial collapse risk prediction value is obtained.

[0089] In this embodiment, the model training of the random forest model comprises:

[0090] Obtain the soil humidity value, soil cavity value, traffic flow data, road surface pressure, and corresponding ground collapse actual risk value of different road sections at different times to generate a training set and a validation set.

[0091] According to the training set, repeatedly perform a decision tree construction operation until the number of decision trees reaches a first threshold, and obtain a trained random forest model.

[0092] According to the validation set, perform performance index evaluation and calculation on the trained random forest model, and according to the performance evaluation index, optimize the hyperparameters of the trained random forest model to obtain a trained random forest model.

[0093] The decision tree construction operation comprises:

[0094] For the root node of each decision tree, a plurality of samples are randomly selected with replacement from the training set, and a plurality of features are randomly selected as a plurality of candidate splitting features.

[0095] According to the candidate split feature of the root node, the sample data of the root node is split to obtain a sample split result of each candidate split feature, and according to the ground collapse actual risk value of the sample split result of each candidate split feature, the mean square error of the sample split result of each candidate split feature is calculated, and the candidate split feature with the minimum mean square error is taken as the optimal split feature of the root node;

[0096] According to the optimal split feature, the sample data of the root node is split to generate a child node and sample data of the child node;

[0097] The candidate split feature selection, optimal split feature determination and sample data splitting operations are repeatedly performed on the child node until the depth of the current decision tree reaches the preset depth threshold, and a single decision tree is obtained.

[0098] In the embodiment, the performance evaluation index calculation method includes but is not limited to root mean square error calculation and mean absolute error calculation.

[0099] For example, the model training of the random forest model includes the following steps:

[0100] Step 1: Data collection and set division:

[0101] Step 1.1: Data collection; install various monitoring devices on multiple different road sections (for example, 10 road sections) to obtain data at different times;

[0102] For soil moisture values, soil moisture sensors are set at different positions and depths on each road section to record humidity values at different times (such as different time periods every day: early morning, noon, evening, and late night), for example, the soil moisture value of road section A is 30% in the early morning and decreases to 20% at noon;

[0103] Use geological exploration equipment to obtain soil cavity values, and mark whether there is a cavity in each road section and the size of the cavity, for example, road section B has a cavity with a value of 0.3 cubic meters;

[0104] Statistical traffic data by traffic monitoring system, distinguish weekdays and non-working days, different time periods (morning peak, noon, evening peak, etc.) of traffic, for example, the traffic of road section C during the morning peak of weekdays is 1200 vehicles per hour;

[0105] Install pressure sensors under the road surface to measure the pressure on the road surface and record the pressure values when different types of vehicles (such as cars, trucks, etc.) pass through. When a truck passes through road section D, the pressure on the road surface is 60,000 pascals;

[0106] The actual ground collapse risk value of each road section at these times is evaluated and recorded by professionals, and the road sections with no obvious collapse signs are marked as low risk (e.g. risk value is 0.1), the road sections with slight collapse signs (e.g. small cracks) are marked as medium risk (e.g. risk value is 0.5), and the road sections with serious collapse conditions (e.g. large pits) are marked as high risk (e.g. risk value is 0.9);

[0107] Step 1.2: Generating training set and validation set; all the collected soil moisture values, soil cavity values, traffic volume data, road surface pressure and corresponding actual ground collapse risk values of road sections at different times are summarized, and the data is randomly divided according to a certain proportion (e.g. 80% of the data as training set, 20% of the data as validation set) to obtain the training set and the validation set;

[0108] Step 2: Decision tree construction operation (constructing a single decision tree):

[0109] Step 2.1: root node operation; for each decision tree, start constructing the root node, randomly select a certain number (e.g. 50) of samples from the training set with replacement, randomly select several features (assume to select soil moisture value, traffic volume and road surface pressure) as candidate split features, and split the 50 sample data of the root node for each candidate split feature; taking soil moisture value as an example, set different humidity intervals (e.g. less than 25% as one group, greater than or equal to 25% as another group), and divide the samples into different groups to obtain the sample split results of each candidate split feature, calculate the mean square error according to the actual ground collapse risk value in the sample split result of each candidate split feature, for example, calculate the mean square error of the risk value in each group and the actual risk value after splitting the two groups of samples according to the soil moisture value, compare the mean square error of each candidate split feature, and select the candidate split feature with the smallest mean square error (assume to be traffic volume) as the optimal split feature of the root node;

[0110] Step 2.2: child node operation; according to the optimal split feature (traffic volume) determined by the root node, split the sample data of the root node to generate child nodes and corresponding child node sample data; for example, samples with traffic volume less than 1000 vehicles / hour are divided into left child nodes, and samples with traffic volume greater than or equal to 1000 vehicles / hour are divided into right child nodes; for left child nodes and right child nodes, the following operations are repeatedly executed: randomly select several features (e.g. soil cavity value and road surface pressure) as candidate split features again, determine the optimal split feature according to the above method of calculating the mean square error, split the sample data according to the optimal split feature to generate new child nodes and child node sample data, and continue to repeat the above operation for the child nodes until the depth of the current decision tree reaches the preset depth threshold (assume the preset depth threshold is 5 layers), at which time the construction of a single decision tree is completed;

[0111] Step 3: constructing a random forest model

[0112] The decision tree construction operation is repeatedly performed, and one decision tree is constructed each time until the number of decision trees reaches a first threshold (assuming that the first threshold is 100), thereby obtaining the trained random forest model.

[0113] Step S3: collecting a real-time remote sensing image of the to-be-measured road surface, performing horizontal feature extraction on the real-time remote sensing image, identifying a feature difference change value on the same horizontal plane in the real-time remote sensing image, and performing fusion processing on the initial collapse risk prediction value according to the feature difference change value, thereby obtaining a first road surface collapse risk prediction value.

[0114] In this embodiment, there are two aspects. The first aspect is that the real-time remote sensing image is subjected to grid processing, and horizontal direction features in the grid image are extracted; a circumcircle of each horizontal direction feature is determined according to the position of each horizontal direction feature in the grid, thereby obtaining the coordinate position of the circumcircle point of each horizontal direction feature; and the spatial distance value between two circumcircle points is determined according to the coordinate positions of the two circumcircle points, thereby serving as a feature difference change value on the same horizontal plane. The second aspect is that a collapse risk point of the to-be-measured road surface in the real-time remote sensing image is determined according to the initial collapse risk prediction value, and the coordinate position of the collapse risk point is determined; the spatial distance value between the coordinate position of each collapse risk point and the coordinate position of any circumcircle point is calculated, thereby serving as a reference change value; and the reference change value and the feature difference change value are subjected to spatial distance superposition fusion, and a value exceeding a preset threshold after superposition fusion is taken as a first road surface collapse risk prediction value.

[0115] In this embodiment, the real-time remote sensing image of the to-be-measured road surface is collected, and the feature difference change value is identified. The initial collapse risk prediction value obtained after prediction by the trained random forest model is subjected to fusion processing with the feature difference change value, thereby solving the problem that the prior art cannot capture the complex nonlinear relationship between geological data and traffic load data due to the dependence on a single data source and the use of a simple linear regression or empirical formula in the prediction model, and the problem of inaccurate road surface collapse risk prediction.

[0116] Step S4: inputting the geological data and traffic load data of the to-be-measured road surface at the current time into the trained gradient boosting decision tree model, so that the trained gradient boosting decision tree model performs road surface collapse risk prediction according to the geological data and traffic load data of the to-be-measured road surface at the current time, thereby obtaining a second road surface collapse risk prediction value.

[0117] In this embodiment, the trained gradient boosting decision tree model includes a plurality of decision trees.

[0118] The trained gradient boosting decision tree model performs road collapse risk prediction according to the geological data and traffic load data of the road to be measured at the current moment, and obtains a second road collapse risk prediction value, including:

[0119] The soil moisture value, soil cavity value, traffic flow and road pressure of the road to be measured at the current moment are input into each decision tree of the trained gradient boosting decision tree model;

[0120] For each decision tree of the gradient boosting decision tree model, the soil moisture value, soil cavity value, traffic flow and road pressure are recursively traversed using the built-in feature splitting rule to determine the target subnode, and the residual prediction value stored in the target subnode is taken as the output value of the decision tree; wherein the residual prediction value is the mean residual of the collapse risk prediction value and the actual collapse risk value of the sample falling into the node in the previous decision tree in the training stage of the gradient boosting decision tree model.

[0121] According to the output value of each decision of the gradient boosting decision tree model, a second road collapse risk prediction value is calculated.

[0122] In this embodiment, the calculation formula of the second road collapse risk prediction value is:

[0123] ;

[0124] Wherein, is the second road collapse risk prediction value; X is a feature composed of the soil moisture value, soil cavity value, traffic flow and road pressure of the road to be measured at the current moment, X = [soil moisture value, soil cavity value, traffic flow, road pressure]; is the initial collapse risk prediction value, which is the mean value of the actual collapse risk values of all training samples in the training stage of the gradient boosting decision tree model; is the preset learning rate of the gradient boosting decision tree model; M is the total number of decision trees; is the output value of the mth decision tree, that is, the residual prediction value stored in the node when X falls into the mth decision tree.

[0125] In this embodiment, the model training of the gradient boosting decision tree model includes:

[0126] Obtain the soil moisture value, soil cavity value, traffic flow data, road pressure and corresponding actual ground collapse risk value of different road sections at different times to generate a training set and a validation set;

[0127] According to the training set, perform decision tree iteration construction operation until the iteration number reaches a second threshold value, and obtain the trained gradient boosting decision tree model;

[0128] According to the verification set, performance index evaluation calculation is performed on the trained gradient boosting decision tree model, and according to the performance evaluation index, the hyperparameters of the trained gradient boosting decision tree model are optimized to obtain the trained gradient boosting decision tree model;

[0129] The decision tree iterative construction operation includes:

[0130] The mean value of the actual values of the collapse risk of all samples in the training set is calculated, and the mean value is taken as the initial prediction value of the collapse risk;

[0131] For each decision tree, the residual value of each sample in the current decision tree is calculated according to the collapse risk prediction value of each sample in the previous decision tree and the actual value of the collapse risk of each sample; wherein, when the first decision tree is constructed, the collapse risk prediction value of each sample in the previous decision tree is the initial prediction value of the collapse risk;

[0132] In the embodiment, the calculation formula of the residual value of each sample in the current decision tree is: ; In the formula, is the residual value of the i th sample in the current decision tree, is the actual value of the collapse risk of the i th sample, is the collapse risk prediction value of the i th sample in the previous decision tree;

[0133] For each node to be split of the decision tree, the node splitting operation is repeatedly performed until the depth of the current decision tree reaches the preset depth threshold, and the current decision tree is obtained;

[0134] According to the feature splitting rule of the current decision tree, the collapse risk prediction value update calculation is performed for each sample of the current decision tree to obtain the collapse risk prediction value of each sample of the current decision tree;

[0135] In the embodiment, the calculation formula of the collapse risk prediction value update calculation for each sample of the current decision tree is: ; In the formula, is the i th sample, is the i th sample collapse risk prediction value of the current decision tree, is the i th sample collapse risk prediction value of the previous decision tree, is the preset learning rate of the gradient boosting decision tree model; M is the total number of decision trees; is the output value of the m th decision tree, that is, the residual prediction value stored by the node when the i th sample falls in the node of the m th decision tree, and the residual prediction value stored by the node is the residual mean value of the sum of the residual values of all samples in the node;

[0136] The node splitting operation includes:

[0137] Randomly select a plurality of features as a plurality of candidate split features, classify the node samples for each candidate feature, and obtain the sub-node sample split results of each candidate split feature; wherein the sub-node sample split result is the mean of the sample residual values in the sub-node, and the mean of the sample residual values in the sub-node is taken as the residual prediction value of the node;

[0138] According to the sub-node sample split results of each candidate split feature, the mean square error of the sub-node sample split results of each candidate split feature is calculated, and the candidate split feature with the minimum mean square error is taken as the optimal split feature, and then the feature split rule is determined according to the optimal split feature.

[0139] Step S5: Calculate the final road collapse risk prediction value based on the first road collapse risk prediction value and the second road collapse risk prediction value;

[0140] In this embodiment, the calculation formula of the final road collapse risk prediction value is:

[0141]

[0142] wherein, is the final road collapse risk prediction value; is the first road collapse risk prediction value; is the second road collapse risk prediction value; X is a feature composed of the soil moisture value, soil cavity value, traffic flow and road pressure of the road to be measured at the current time, X = [soil moisture value, soil cavity value, traffic flow, road pressure]; and are the first risk weight and the second risk weight, respectively.

[0143] Step S6: Compare the final road collapse risk prediction value with the preset risk level interval to determine the road collapse risk level, and regulate the road to be measured according to the road collapse risk level.

[0144] In this embodiment, the comparison of the final road collapse risk prediction value with the preset risk level interval to determine the road collapse risk level comprises:

[0145] Compare the final road collapse risk prediction value with the upper limit index and the lower limit index of the preset risk level interval to determine the road collapse risk level; wherein the preset risk level interval includes a low risk level interval, a medium risk level interval and a high risk level interval; the upper limit index of the low risk level interval is less than the lower limit index of the medium risk level interval; the upper limit index of the medium risk level interval is less than the lower limit index of the high risk level interval.

[0146] ​For example, the low-risk level interval can be set as [0, 0.3); the medium-risk level interval can be set as [0.3, 0.7); and the high-risk level interval can be set as [0.7, +∞). When the road collapse risk level is a low-risk level, the normal road inspection frequency is maintained, and the maintenance work is carried out according to the normal road maintenance plan. When the road collapse risk level is a medium-risk level, the road inspection and geological detection frequencies are increased, and the potential problems of the road, such as small cracks and local subsidence, are solved to repair and reinforce the road in time. When the road collapse risk level is a high-risk level, the high-risk road section is closed or partially closed, and the high-risk road section is repaired or reconstructed.

[0147] In the embodiment, the initial collapse risk prediction value obtained by collecting the real-time remote sensing image of the to-be-tested road, identifying the feature difference change value, and performing prediction by the trained random forest model is fused with the feature difference change value, and then the gradient boosting decision tree model capable of capturing the complex nonlinear relationship between the geological data and the traffic load data is used to predict the road collapse risk of the multi-source data composed of the geological data and the traffic load data. Not only the limitation of relying on a single data source and ignoring other influencing factors is avoided, but also the problem that the prediction model cannot capture the complex nonlinear relationship between the geological data and the traffic load data by using a simple linear regression or an empirical formula is solved. The accuracy of the road collapse risk prediction is improved, and the problem that the road collapse risk prediction is inaccurate due to the dependence on a single data source and the use of a simple linear regression or an empirical formula in the prediction model is solved. In addition, the final prediction value is compared with the preset risk level interval to determine the risk level and regulate the to-be-tested road according to the risk level, which further enhances the effectiveness of risk prevention and control and provides a powerful guarantee for road safety.

[0148] Embodiment 2

[0149] Referring to Figure 2 The module diagram of the road collapse risk prediction system provided by the embodiment of the present application is provided. In order to solve the problem that the road collapse risk prediction is inaccurate due to the dependence on a single data source and the use of a simple linear regression or an empirical formula in the prediction model, the system at least includes a data acquisition module, an initial risk prediction module, a first risk prediction module, a second risk prediction module, a final risk prediction module, and a road risk monitoring module.

[0150] The data acquisition module is configured to acquire geological data and traffic load data of the to-be-tested road surface at a current time; wherein the geological data comprises soil humidity values and soil cavity values; and the traffic load data comprises traffic flow and road surface pressure;

[0151] The initial risk prediction module is configured to input the geological data and the traffic load data of the to-be-tested road surface at the current time into the trained random forest model, so that the trained random forest model performs road surface collapse risk prediction according to the geological data and the traffic load data of the to-be-tested road surface at the current time, and obtains an initial collapse risk prediction value;

[0152] The first risk prediction module is configured to collect a real-time remote sensing image of the to-be-tested road surface, perform horizontal feature extraction on the real-time remote sensing image, identify a feature difference change value in the real-time remote sensing image on the same horizontal plane, and perform fusion processing on the initial collapse risk prediction value according to the feature difference change value, to obtain a first road surface collapse risk prediction value;

[0153] The second risk prediction module is configured to input the geological data and the traffic load data of the to-be-tested road surface at the current time into the trained gradient boosting decision tree model, so that the trained gradient boosting decision tree model performs road surface collapse risk prediction according to the geological data and the traffic load data of the to-be-tested road surface at the current time, and obtains a second road surface collapse risk prediction value;

[0154] The final risk prediction module is configured to calculate a final road surface collapse risk prediction value from the first road surface collapse risk prediction value and the second road surface collapse risk prediction value.

[0155] The road surface risk monitoring module is configured to compare the final road surface collapse risk prediction value with a preset risk level interval, determine a road surface collapse risk level, and regulate the to-be-tested road surface according to the road surface collapse risk level.

[0156] The above-described specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present application, and it should be understood that the above-described embodiments are merely specific embodiments of the present application and are not intended to limit the protection scope of the present application. It should be particularly pointed out that any modifications, equivalent replacements, improvements, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load, characterized in that, include: Obtain the geological data and traffic load data of the road surface to be tested at the current moment; wherein, the geological data includes soil moisture value and soil porosity value; the traffic load data includes traffic flow and road surface pressure; The geological data and traffic load data of the road surface to be tested at the current moment are input into the trained random forest model so that the trained random forest model can predict the road collapse risk based on the geological data and traffic load data of the road surface to be tested at the current moment and obtain the initial collapse risk prediction value. Real-time remote sensing images of the road surface to be tested are acquired, horizontal features are extracted from the real-time remote sensing images, the feature difference change values ​​on the same horizontal plane in the real-time remote sensing images are identified, and the initial collapse risk prediction value is fused based on the feature difference change values ​​to obtain the first road surface collapse risk prediction value. The geological data and traffic load data of the road surface to be tested at the current moment are input into the trained gradient boosting decision tree model so that the trained gradient boosting decision tree model can predict the road collapse risk based on the geological data and traffic load data of the road surface to be tested at the current moment and obtain the second road collapse risk prediction value. The first and second road surface collapse risk prediction values ​​are used to calculate the final road surface collapse risk prediction value. The final predicted value of road collapse risk is compared with the preset risk level range to determine the road collapse risk level, and the road surface to be tested is adjusted according to the road collapse risk level. The trained random forest model comprises several decision trees. Based on the geological and traffic load data of the road surface at the current moment, the trained random forest model predicts the road subsidence risk to obtain a first road subsidence risk prediction value. This includes: inputting the soil moisture value, soil void value, traffic flow, and road pressure of the road surface at the current moment into each decision tree of the trained random forest model; for each decision tree, recursively traversing the soil moisture value, soil void value, traffic flow, and road pressure using built-in feature splitting rules to determine target child nodes, and using the subsidence risk prediction value stored in the target child node as the subsidence risk prediction output value of that decision tree; wherein, the subsidence risk prediction value stored in the target child node is determined during the training of the random forest model by calculating the average of the actual subsidence risk values ​​of all training samples falling into that child node; and calculating the mean of the subsidence risk prediction output values ​​based on the subsidence risk prediction output values ​​of each decision tree, and using the mean of the subsidence risk prediction output values ​​as the first road subsidence risk prediction value.

2. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load as described in claim 1, characterized in that, The steps of acquiring real-time remote sensing images of the road surface to be measured, extracting horizontal features from the real-time remote sensing images, and identifying the feature difference changes on the same horizontal plane in the real-time remote sensing images specifically include: The real-time remote sensing image is processed into a grid, and the horizontal features in the gridded image are extracted. Based on the position of each horizontal feature in the grid, the outer center circle of the horizontal feature is determined, and the coordinate position of the outer center circle point of each horizontal feature is obtained; The spatial distance between two concentric points is determined by their coordinate positions and used as the characteristic difference variation value on the same horizontal plane.

3. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load according to claim 2, characterized in that, The step of fusing the initial collapse risk prediction value based on the feature difference change value to obtain the first road collapse risk prediction value specifically includes: Based on the initial predicted collapse risk value, the collapse risk points of the road surface to be tested in the real-time remote sensing image are determined, and the coordinates of the collapse risk points are determined. Calculate the spatial distance between the coordinates of each collapse risk point and the coordinates of any outer center point, and use this as a reference change value; The reference change value and the feature difference change value are spatially superimposed and fused, and the value that exceeds a preset threshold after superposition and fusion is used as the first road collapse risk prediction value.

4. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load as described in claim 1, characterized in that, The training of the random forest model includes: Obtain soil moisture value, soil porosity value, traffic flow data, road surface pressure and corresponding actual risk value of ground collapse at different times for different road sections, and generate training set and validation set; Based on the training set, the decision tree construction operation is repeated until the number of decision trees reaches the first threshold, and the trained random forest model is obtained. Based on the validation set, the performance metrics of the trained random forest model are evaluated and calculated. Based on the performance evaluation metrics, the hyperparameters of the trained random forest model are optimized to obtain a well-trained random forest model. The decision tree construction operation includes: For the root node of each decision tree, several samples are randomly selected from the training set with replacement, and several features are randomly selected as several candidate splitting features. Based on the candidate splitting features of the root node, the sample data of the root node is split to obtain the sample splitting results of each candidate splitting feature. Based on the actual risk value of ground collapse of the sample splitting results of each candidate splitting feature, the mean square error of the sample splitting results of each candidate splitting feature is calculated, and the candidate splitting feature with the smallest mean square error is taken as the optimal splitting feature of the root node. Based on the optimal splitting feature, the sample data of the root node is split to generate child nodes and sample data of the child nodes. Repeatedly perform candidate splitting feature selection, optimal splitting feature determination, and sample data splitting operations on child nodes until the depth of the current decision tree reaches a preset depth threshold, thus obtaining a single decision tree.

5. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load according to claim 4, characterized in that, The trained gradient boosting decision tree model includes several decision trees; The trained gradient boosting decision tree model predicts road collapse risk based on the current geological and traffic load data of the road surface under test, and obtains a second road collapse risk prediction value, including: The soil moisture value, soil porosity value, traffic flow, and road pressure of the road surface at the current moment are input into each decision tree of the trained gradient boosting decision tree model; For each decision tree in the gradient boosting decision tree model, the built-in feature splitting rules are used to recursively traverse the soil moisture value, soil void value, traffic flow and road pressure to determine the target child node. The residual prediction value stored in the target child node is used as the output value of the decision tree. The residual prediction value is the mean residual between the predicted value of collapse risk and the actual value of collapse risk of the sample falling into the node in the previous decision tree during the training phase of the gradient boosting decision tree model. The second road surface collapse risk prediction value is calculated based on the output value of each decision in the gradient boosting decision tree model.

6. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load according to claim 5, characterized in that, The formula for calculating the second road surface collapse risk prediction value is as follows: ； in, X is the predicted value of the second road surface collapse risk; X is a feature composed of the soil moisture value, soil void value, traffic flow and road pressure of the road surface at the current moment, X=[soil moisture value, soil void value, traffic flow, road pressure]; The initial predicted value of the collapse risk is the average of the actual collapse risk values ​​of all training samples during the training phase of the gradient boosting decision tree model. The preset learning rate for gradient boosting decision tree models; M is the total number of decision trees; This is the output value of the m-th decision tree, that is, the residual prediction value stored in the node when X falls into the child node of the m-th decision tree.

7. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load as described in claim 6, characterized in that, The training of the gradient boosting decision tree model includes: Obtain soil moisture value, soil porosity value, traffic flow data, road surface pressure and corresponding actual risk value of ground collapse at different times for different road sections, and generate training set and validation set; Based on the training set, perform iterative decision tree construction operations until the number of iterations reaches the second threshold, and then obtain the trained gradient boosting decision tree model. Based on the validation set, the performance metrics of the trained gradient boosting decision tree model are evaluated and calculated. Based on the performance evaluation metrics, the hyperparameters of the trained gradient boosting decision tree model are optimized to obtain the trained gradient boosting decision tree model. The decision tree iterative construction operation includes: Calculate the mean of the actual values ​​of collapse risk for all samples in the training set, and use the mean as the initial predicted value of collapse risk; For each decision tree, the residual value of each sample in the current decision tree is calculated based on the predicted collapse risk value of each sample in the previous decision tree and the actual collapse risk value of each sample; wherein, when constructing the first decision tree, the predicted collapse risk value of each sample in the previous decision tree is the initial predicted collapse risk value. For each node to be split in the decision tree, repeat the node splitting operation until the depth of the current decision tree reaches the preset depth threshold, and then obtain the current decision tree. Based on the feature splitting rules of the current decision tree, the collapse risk prediction value is updated and calculated for each sample of the current decision tree to obtain the collapse risk prediction value of each sample of the current decision tree. The node splitting operation includes: Several features are randomly selected as candidate splitting features. For each candidate feature, the node samples are classified to obtain the splitting results of the child node samples of each candidate splitting feature. The splitting result of the child node samples is the mean of the sample residuals within the child node. The mean of the sample residuals within the child node is used as the residual prediction value of that node. Based on the splitting results of the child node samples of each candidate splitting feature, the mean square error of the splitting results of the child node samples of each candidate splitting feature is calculated, and the candidate splitting feature with the smallest mean square error is taken as the optimal splitting feature. Then, based on the optimal splitting feature, the feature splitting rule is determined.

8. The method for predicting road subsidence risk by integrating ground-penetrating radar and traffic load according to claim 7, characterized in that, The step of comparing the final predicted road collapse risk value with a preset risk level range to determine the road collapse risk level includes: The final predicted value of road collapse risk is compared with the upper and lower limits of the preset risk level range to determine the road collapse risk level; wherein, the preset risk level range includes: low risk level range, medium risk level range and high risk level range; the upper limit index of the low risk level range is less than the lower limit index of the medium risk level range; the upper limit index of the medium risk level range is less than the lower limit index of the high risk level range.

9. A road subsidence risk prediction system integrating ground-penetrating radar and traffic load analysis, characterized in that: include: The system includes a data acquisition module, an initial risk prediction module, a first risk prediction module, a second risk prediction module, a final risk prediction module, and a road surface risk monitoring module. The data acquisition module is used to acquire the geological data and traffic load data of the road surface to be tested at the current moment; wherein, the geological data includes soil moisture value and soil porosity value; the traffic load data includes traffic flow and road surface pressure. The initial risk prediction module is used to input the geological data and traffic load data of the road surface to be tested at the current moment into the trained random forest model, so that the trained random forest model can predict the road collapse risk based on the geological data and traffic load data of the road surface to be tested at the current moment and obtain the initial collapse risk prediction value. The first risk prediction module is used to acquire real-time remote sensing images of the road surface to be tested, extract horizontal features from the real-time remote sensing images, identify the feature difference change values ​​on the same horizontal plane in the real-time remote sensing images, and fuse the initial collapse risk prediction value according to the feature difference change values ​​to obtain the first road surface collapse risk prediction value. The second risk prediction module is used to input the geological data and traffic load data of the road surface to be tested at the current moment into the trained gradient boosting decision tree model, so that the trained gradient boosting decision tree model can predict the road collapse risk based on the geological data and traffic load data of the road surface to be tested at the current moment, and obtain the second road collapse risk prediction value. The final risk prediction module is used to calculate the final road collapse risk prediction value from the first road collapse risk prediction value and the second road collapse risk prediction value. The road surface risk monitoring module is used to compare the final predicted value of road surface collapse risk with the preset risk level range, determine the road surface collapse risk level, and adjust the road surface to be tested according to the road surface collapse risk level. The trained random forest model comprises several decision trees. Based on the geological and traffic load data of the road surface at the current moment, the trained random forest model predicts the road subsidence risk to obtain a first road subsidence risk prediction value. This includes: inputting the soil moisture value, soil void value, traffic flow, and road pressure of the road surface at the current moment into each decision tree of the trained random forest model; for each decision tree, recursively traversing the soil moisture value, soil void value, traffic flow, and road pressure using built-in feature splitting rules to determine target child nodes, and using the subsidence risk prediction value stored in the target child node as the subsidence risk prediction output value of that decision tree; wherein, the subsidence risk prediction value stored in the target child node is determined during the training of the random forest model by calculating the average of the actual subsidence risk values ​​of all training samples falling into that child node; and calculating the mean of the subsidence risk prediction output values ​​based on the subsidence risk prediction output values ​​of each decision tree, and using the mean of the subsidence risk prediction output values ​​as the first road subsidence risk prediction value.

Images