A method and system for predicting subgrade settlement

Abstract

The present application relates to the technical field of road construction, and particularly relates to a subgrade settlement prediction method and system. A subgrade settlement prediction system comprises a subgrade excitation module, a response characteristic spectrum construction module, a first subgrade settlement prediction module and a second subgrade settlement prediction module. The present application excites the subgrade by means of an exciter, collects multi-dimensional response signals such as ground sound, resistivity and geomagnetic signals, extracts nonlinear characteristic indexes reflecting the change of the subgrade structure, such as the third harmonic energy ratio and the Lyapunov index, maps the high-dimensional features to a low-dimensional space by means of UMAP dimension reduction to generate an evolution trajectory graph, and then identifies potential settlement risks through abnormal detection of the evolution trajectory graph, so that the micro changes of the subgrade can be found in time in the settlement incubation stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road construction, in particular to a subgrade settlement prediction method and system. BACKGROUND

[0002] As a common engineering problem, subgrade settlement has a direct impact on the stability and safety of the road. Therefore, accurate prediction of subgrade settlement is an indispensable part of the road construction and maintenance process. Existing subgrade settlement monitoring methods mainly rely on traditional sensor data collection, geological survey and physical model establishment. Although these methods have improved the accuracy of settlement early warning to some extent, they still face many challenges. On the one hand, traditional methods often rely on frequent data collection and manual analysis, making it difficult to achieve real-time monitoring and automated decision-making. On the other hand, existing technologies focus more on local area settlement detection, lacking a global and long-term monitoring and early warning mechanism for the incubation stage of settlement. In addition, subgrade settlement often occurs gradually, and traditional methods often fail to identify potential settlement risks in time when dealing with dynamic changes and complex environmental settlement risks. SUMMARY

[0003] The present application stimulates the subgrade by a vibration exciter, collects multi-dimensional response signals such as ground sound, resistivity and geomagnetic field, extracts nonlinear feature indicators reflecting changes in subgrade structure such as third harmonic energy ratio and Lyapunov exponent, maps high-dimensional features to low-dimensional space by UMAP dimensionality reduction to generate an evolution trajectory graph, and identifies potential settlement risks through anomaly detection of the evolution trajectory graph, enabling timely detection of microscopic changes in the subgrade during the incubation stage of settlement.

[0004] The present application provides a subgrade settlement prediction method, comprising:

[0005] At the subgrade monitoring time point, start the vibration exciter and perform a frequency sweep operation for a preset sweep time to determine the main response frequency, then set the vibration exciter based on the main response frequency and continuously perform a single-frequency excitation operation, and collect excitation response signals from the sensor group for a preset collection time;

[0006] Perform short-time Fourier transform on the excitation response signals to determine the response feature spectrum corresponding to the excitation response signals, which includes third harmonic energy ratio, Lyapunov exponent, peak delay time and WVD energy perturbation rate;

[0007] The UMAP dimension reduction operation is performed on the response feature spectrum to obtain a plurality of embedded feature points, the horizontal coordinate of the embedded feature point is a timestamp, and the vertical coordinate is a data value after UMAP dimension reduction, all embedded feature points are added to the corresponding evolution trajectory diagram, and then the evolution trajectory diagram is subjected to abnormality identification, if the identification result is abnormal, it is considered that there is a risk of subgrade settlement, otherwise it is considered that there is no risk of subgrade settlement.

[0008] As a preferred aspect, the exciter is started, a frequency sweep operation of a preset sweep time is performed to determine the main response frequency, specifically including the following steps:

[0009] A linear frequency sweep signal is set to the exciter, the exciter performs an excitation operation, and an excitation response signal of a preset sweep time is collected, and the excitation response signal includes a plurality of sub-channel excitation signals, each sub-channel excitation signal corresponds to one sensor in the sensor group;

[0010] The Fourier transform is performed on each sub-channel excitation signal to obtain a sub-channel excitation spectrum corresponding to each sub-channel excitation signal, and then a peak detection algorithm is used to determine an energy peak value corresponding to each sub-channel excitation spectrum, and the frequency corresponding to the energy peak value is recorded as a response frequency, and then a weighted sum operation is performed on all response frequencies and corresponding energies to obtain a main response frequency, and the energy corresponding to the response frequency is the square of the energy peak value corresponding to the response frequency.

[0011] As a preferred aspect, the short-time Fourier transform is performed on the excitation response signal to determine the response feature spectrum corresponding to the excitation response signal, specifically including the following steps:

[0012] The following operations are performed on each sub-channel excitation signal in the excitation response signal: the sub-channel excitation signal is divided into a plurality of sub-channel excitation signal segments according to a short-time window, the third harmonic energy ratio extraction operation, Lyapunov index extraction operation, peak delay time extraction operation and WVD energy perturbation rate extraction operation are performed on each sub-channel excitation signal segment, and the outputs after the third harmonic energy ratio extraction operation, Lyapunov index extraction operation, peak delay time extraction operation and WVD energy perturbation rate extraction operation of the sub-channel excitation signal segment are spliced into a response feature, and the response feature is labeled by a timestamp;

[0013] All response features under the same timestamp are spliced to construct a comprehensive response feature, and all comprehensive response features labeled with timestamps form a response feature spectrum.

[0014] As a preferred aspect, the UMAP dimension reduction and evolution trajectory mapping operations are performed on the response feature spectrum to obtain a plurality of embedded feature points, specifically including the following operations:

[0015] Performing the UMAP dimension reduction operation on each comprehensive response feature in the response feature spectrum, and constructing an embedded feature point with the timestamp corresponding to the comprehensive response feature as the horizontal coordinate and the output of the comprehensive response feature after performing the UMAP dimension reduction operation as the vertical coordinate;

[0016] The UMAP dimension reduction operation specifically includes the following contents:

[0017] Step S1: Calculate the similarity between any two comprehensive response features, and select the k nearest comprehensive response features of each comprehensive response feature using the k nearest neighbor algorithm, and construct a similar edge between the comprehensive response feature and its corresponding nearest comprehensive response feature, the weight of each similar edge being the similarity between the two comprehensive response features, all comprehensive response features and similar edges forming a high-dimensional KNN graph;

[0018] Step S2: Randomly initialize a low-dimensional data point for each comprehensive response feature in a low-dimensional space, calculate the similarity between any two low-dimensional data points, and select the k nearest low-dimensional data points of each low-dimensional data point using the k nearest neighbor algorithm, and construct a similar edge between the low-dimensional data point and its corresponding nearest low-dimensional data point, the weight of each similar edge being the similarity between the two low-dimensional data points, all low-dimensional data points and similar edges forming a low-dimensional KNN graph;

[0019] Step S3: Construct a cost function including the cross-entropy loss between the high-dimensional KNN graph and the low-dimensional KNN graph, and optimize the cost function using the gradient descent algorithm to adjust the low-dimensional data points in the low-dimensional space;

[0020] Step S4: Repeat steps S2-S3 until the cost function converges, and the low-dimensional data point corresponding to the comprehensive response feature is taken as the output of the comprehensive response feature after performing the UMAP dimension reduction operation.

[0021] As a preferred aspect, anomaly identification is performed on the evolution trajectory graph, specifically including the following steps:

[0022] At the current roadbed monitoring time point, unsupervised clustering is performed on all embedded feature points in the evolution trajectory graph, and the cluster including the most embedded feature points after performing the unsupervised clustering is recorded as the target cluster, the ratio between the number of embedded feature points in the target cluster and the total number of all embedded feature points is recorded as the evolution stability value, and it is determined whether the evolution stability value is higher than the stability threshold, if the evolution stability value is higher than the stability threshold, the identification result is output as normal, and if the evolution stability value is not higher than the stability threshold, the identification result is output as abnormal.

[0023] As a preferred aspect, it also includes performing roadbed settlement prediction through machine learning, specifically including the following contents:

[0024] The response feature spectrum is sent into a subgrade settlement prediction model for processing, and a subgrade settlement prediction result is output, the subgrade settlement prediction result including existence of a subgrade settlement risk and nonexistence of the subgrade settlement risk, the subgrade settlement prediction model being established based on an LSTM model, and the subgrade settlement prediction model further including a full connection operation.

[0025] The subgrade settlement prediction model is trained, specifically including the following contents: a plurality of subgrade settlement prediction training samples are obtained, the subgrade settlement prediction training samples including the response feature spectrum, the subgrade settlement prediction training samples are labeled by the subgrade settlement prediction result, all the labeled subgrade settlement prediction training samples are combined to form a subgrade settlement prediction training set, and the subgrade settlement prediction model is trained by the subgrade settlement prediction training set, and the training target is the labeled subgrade settlement prediction result.

[0026] The application further provides a subgrade settlement prediction system, including:

[0027] The subgrade excitation module is configured to start the exciter at a subgrade monitoring time point, perform a frequency sweep operation for a preset sweep time, determine a main response frequency, and then set the exciter based on the main response frequency, and continuously perform a single-frequency excitation operation, and collect excitation response signals for a preset collection time by the sensor group.

[0028] The response feature spectrum construction module is configured to perform a short-time Fourier transform on the excitation response signals to determine a response feature spectrum corresponding to the excitation response signals, the response feature spectrum including a third harmonic energy ratio, a Lyapunov exponent, a peak delay time, and a WVD energy perturbation rate.

[0029] The first subgrade settlement prediction module is configured to perform a UMAP dimension reduction operation on the response feature spectrum to obtain a plurality of embedded feature points, the horizontal coordinate of the embedded feature point being a timestamp and the vertical coordinate being a data value after the UMAP dimension reduction, and then add all the embedded feature points to a corresponding evolution trajectory graph, and perform abnormality identification on the evolution trajectory graph, and if the identification result is abnormal, it is considered that there is a subgrade settlement risk, otherwise, it is considered that there is no subgrade settlement risk.

[0030] The second subgrade settlement prediction module is configured to send the response feature spectrum into a subgrade settlement prediction model for processing, and output a subgrade settlement prediction result, the subgrade settlement prediction result including existence of a subgrade settlement risk and nonexistence of the subgrade settlement risk, the subgrade settlement prediction model being established based on an LSTM model, and the subgrade settlement prediction model further including a full connection operation.

[0031] The application has the following advantages:

[0032] The application stimulates the roadbed through the exciter, collects multi-dimensional response signals such as ground sound, resistivity, geomagnetic and the like, extracts nonlinear characteristic indexes reflecting the roadbed structure change such as third harmonic energy ratio, Lyapunov index and the like, maps the high-dimensional features to a low-dimensional space through UMAP dimension reduction to generate an evolution trajectory graph, and identifies the potential settlement risk through the abnormal detection of the evolution trajectory graph, so that the micro changes of the roadbed can be found in time in the settlement gestation period. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 A structural schematic diagram of a roadbed settlement prediction system used for the embodiment of the application. DETAILED DESCRIPTION

[0034] In order to enable personnel in the technical field to better understand the technical solutions in the application, the technical solutions in the embodiment of the application will be clearly and completely described below in combination with the drawings in the embodiment of the application.

[0035] Embodiment 1, a roadbed settlement prediction method, comprising:

[0036] At the roadbed monitoring time point, the exciter is started, and a frequency sweep operation of a preset sweep frequency time is performed, the preset sweep frequency time is generally 3-5s, and the main response frequency is determined, the main response frequency is the most significant vibration frequency of the roadbed structure under the action of external excitation, and reflects the dynamic characteristics of the roadbed, especially when subjected to periodic or continuous excitation, the response of the roadbed to the excitation is the most intense frequency band, and then the exciter is set based on the main response frequency, and a single-frequency excitation operation is continuously performed, the excitation response signals of a preset collection time are collected through the sensor group, the preset collection time is generally 5min, and the preset collection time single-frequency excitation operation means that a fixed frequency (i.e. the main response frequency) excitation operation is applied through the exciter, and it needs to be noted that when the main response frequency corresponding to the roadbed is close to the frequency applied by the exciter, the roadbed will have a more significant reaction, and in the settlement prediction, this sensitive change can provide an early warning signal about the stability or potential problems of the roadbed; the sensor group here generally includes a ground sound sensor (which can measure the characteristics of the ground sound wave after the excitation signal is applied), a soil resistivity sensor (which is used to measure the resistivity of the foundation) and a geomagnetic sensor (which is used to measure the magnetic field data of the roadbed) and the like;

[0037] In actual operation, a vibration exciter is generally arranged at a preset distance along the roadbed. The vibration exciter can be an electromagnetic vibration exciter, which can convert an excitation signal into a mechanical vibration and apply the mechanical vibration to the surface of the roadbed by using electromagnetic force. The roadbed part corresponding to each vibration exciter is analyzed, and therefore, an edge processor is arranged at the position corresponding to each vibration exciter to analyze the excitation response signal corresponding to the vibration exciter. In this embodiment, the excitation response signal corresponding to any vibration exciter is analyzed.

[0038] The short-time Fourier transform is performed on the excitation response signal to determine a response characteristic spectrum corresponding to the excitation response signal. The response characteristic spectrum includes characteristic information such as a third harmonic energy ratio, a Lyapunov index, a peak delay time, and a WVD energy perturbation rate, which can reflect the material structure change of the roadbed and further reflect the settlement risk of the roadbed.

[0039] The UMAP dimension reduction operation is performed on the response characteristic spectrum to obtain a plurality of embedded feature points. The abscissa of the embedded feature point is a timestamp, and the ordinate is a data value after UMAP dimension reduction. All embedded feature points are added to the corresponding evolution trajectory graph. The data value after UMAP dimension reduction can also be two-dimensional, and the corresponding embedded feature point is a point in a three-dimensional space. Abnormality identification is performed on the evolution trajectory graph. The abnormality identification refers to analyzing whether the embedded feature points in the evolution trajectory graph conform to the normal change trend. If the identification result is abnormal, it is considered that there is a roadbed settlement risk, otherwise, it is considered that there is no roadbed settlement risk. It should be noted that when the roadbed is subjected to external force, disturbance, or environmental change, it will exhibit a certain response mode. For example, after the soil body or structure is excited, the vibration, temperature, electromagnetic field, and other characteristic signals will exhibit characteristics such as frequency, amplitude, and phase. As the settlement proceeds, the physical properties of the soil body, such as elasticity, pore structure, humidity, and soil pressure, change. These changes will cause significant changes in the nonlinear response characteristics of the soil body and structure. The differences between the normal evolution trajectory and the abnormal evolution trajectory in the evolution trajectory graph mapped by these nonlinear response characteristics can be identified. The appearance of the abnormal trajectory usually indicates the potential occurrence of settlement.

[0040] The application excites the roadbed by using a vibration exciter, collects multi-dimensional response signals such as ground sound, resistivity, and geomagnetic field, extracts nonlinear characteristic indexes reflecting the structure change of the roadbed, such as the third harmonic energy ratio and the Lyapunov index, maps the high-dimensional characteristics to a low-dimensional space by UMAP dimension reduction to generate an evolution trajectory graph, and identifies the potential settlement risk by abnormality detection of the evolution trajectory graph, which can discover the microscopic changes of the roadbed in the settlement incubation stage in a timely manner.

[0041] Start the exciter, perform a frequency sweep operation for a preset sweep time to determine the main response frequency, specifically including the following steps:

[0042] A linear frequency sweep signal is set for the exciter, which is a continuous sine wave signal with a frequency gradually increasing from low to high at a preset rate (e.g. 0.1 HZ / s), and the frequency range is generally 0.1 HZ-30HZ. The exciter performs an excitation operation, and the excitation response signal is collected for a preset sweep time, and the excitation response signal includes several sub-channel excitation signals, each corresponding to a sensor in the sensor group;

[0043] Perform Fourier transform on each sub-channel excitation signal to obtain the corresponding sub-channel excitation spectrum, and then use a peak detection algorithm to determine the energy peak value corresponding to each sub-channel excitation spectrum. The peak detection algorithm can determine the maximum amplitude by finding the local maximum value or the significant gradient change, which is the energy peak value. The frequency corresponding to the energy peak value is recorded as the response frequency. Perform a weighted sum operation on all response frequencies and corresponding energies to obtain the main response frequency. The energy corresponding to the response frequency is the square of the energy peak value corresponding to the response frequency.

[0044] Perform short-time Fourier transform on the excitation response signal to determine the response feature spectrum corresponding to the excitation response signal, specifically including the following steps:

[0045] For each sub-channel excitation signal in the excitation response signal, perform the following operations: divide the sub-channel excitation signal into several sub-channel excitation signal segments according to the short-time window, which is generally 3-5s. Then perform third-order harmonic energy ratio extraction, Lyapunov index extraction, peak delay time extraction, and WVD energy perturbation rate extraction on each sub-channel excitation signal segment, and concatenate the outputs of the sub-channel excitation signal segments after the above operations to obtain the response feature. Label the response feature with a timestamp, which is generally the median value of the time axis corresponding to the sub-channel excitation signal segment. The third-order harmonic energy ratio extraction, Lyapunov index extraction, peak delay time extraction, and WVD energy perturbation rate extraction are generally implemented through functions in numpy or scipy libraries;

[0046] Concatenate all response features under the same timestamp to construct a comprehensive response feature, and all labeled comprehensive response features form the response feature spectrum.

[0047] Perform UMAP dimensionality reduction and evolutionary trajectory mapping operations on the response feature spectrum to obtain several embedded feature points, specifically including the following operations:

[0048] For each comprehensive response feature in the response feature spectrum, a UMAP dimension reduction operation is performed, and an embedding feature point is constructed by taking the timestamp corresponding to the comprehensive response feature as the horizontal coordinate and taking the output of the comprehensive response feature after the UMAP dimension reduction operation as the vertical coordinate.

[0049] The UMAP dimension reduction operation specifically includes the following contents:

[0050] Step S1: Calculate the similarity between any two comprehensive response features, and use the k-nearest neighbor algorithm to select the k nearest neighbors of each comprehensive response feature, and construct a similar edge between the comprehensive response feature and its corresponding nearest neighbor comprehensive response feature, the weight of each similar edge is the similarity between the two comprehensive response features, all comprehensive response features and similar edges form a high-dimensional KNN graph.

[0051] Step S2: In the low-dimensional space, randomly initialize a low-dimensional data point for each comprehensive response feature, calculate the similarity between any two low-dimensional data points, and use the k-nearest neighbor algorithm to select the k nearest neighbors of each low-dimensional data point, and construct a similar edge between the low-dimensional data point and its corresponding nearest neighbor low-dimensional data point, the weight of each similar edge is the similarity between the two low-dimensional data points, all low-dimensional data points and similar edges form a low-dimensional KNN graph.

[0052] Step S3: Construct a cost function, the cost function includes the cross-loss entropy between the high-dimensional KNN graph and the low-dimensional KNN graph, use the gradient descent algorithm to optimize the cost function, and adjust the low-dimensional data points in the low-dimensional space.

[0053] Step S4: Repeat steps S2-S3 until the cost function converges, and take the low-dimensional data point corresponding to the comprehensive response feature as the output of the comprehensive response feature after the UMAP dimension reduction operation.

[0054] Abnormal identification is performed on the evolution trajectory graph, specifically including the following steps:

[0055] At the current roadbed monitoring time point, unsupervised clustering is performed on all embedded feature points in the evolution trajectory diagram, the DBSCAN algorithm can be used, and the clustering cluster including the most embedded feature points after performing unsupervised clustering is recorded as a target clustering cluster. The ratio between the number of embedded feature points in the target clustering cluster and the total number of all embedded feature points is recorded as an evolution stability value. Then, it is determined whether the evolution stability value is higher than the stability threshold. If the evolution stability value is higher than the stability threshold, it means that the normal change trend is met, and the recognition result is output as normal. If the evolution stability value is not higher than the stability threshold, it means that the normal change trend is not met, and the recognition result is output as abnormal. The evolution stability value can reflect whether the current monitoring data meets the expected stable mode. If the evolution stability value is high, it means that most data points are concentrated in the target cluster, indicating that the response change of the roadbed meets the normal evolution trend. If the stability value is low, it may indicate an abnormal evolution process, such as a sudden change in the structure of the soil body or an abnormal change in the settlement.

[0056] In order to further improve the accuracy of roadbed settlement prediction, the application also considers that when the edge processor analyzes, if there is a risk of roadbed settlement, the edge processor can send the response feature spectrum to the cloud server for machine learning processing to further predict the roadbed settlement. Specifically, the following contents are included:

[0057] The response feature spectrum is sent to the roadbed settlement prediction model for processing to output the roadbed settlement prediction result. The roadbed settlement prediction result includes the existence of roadbed settlement risk and the non-existence of roadbed settlement risk. The roadbed settlement prediction model is constructed based on the LSTM model. In order to output the roadbed settlement prediction result, the roadbed settlement prediction model further includes a full connection operation.

[0058] The roadbed settlement prediction model is trained, and the specific contents include the following: a plurality of roadbed settlement prediction training samples are obtained, the roadbed settlement prediction training sample includes a response characteristic spectrum constructed according to an actual roadbed settlement experiment, it should be noted that an operator generally simulates a path settlement experiment through simulation software or actual operation, and the response characteristic spectrum is constructed from the acquired excitation response signal before the roadbed settlement occurs, the roadbed settlement prediction result is labeled on the roadbed settlement prediction training sample, and the labeled roadbed settlement prediction result is generally a roadbed settlement risk, because the response characteristic spectrum is constructed from the acquired excitation response signal before the roadbed settlement occurs, and the value is generally 1, all the labeled roadbed settlement prediction training samples form a roadbed settlement prediction training set, the roadbed settlement prediction model is trained through the roadbed settlement prediction training set, the training target is the labeled roadbed settlement prediction result, whether the training condition is met is judged, the training condition is generally that the accuracy of the roadbed settlement prediction model meets the expectation, if the training condition is met, the trained roadbed settlement prediction model is output, otherwise, the roadbed settlement prediction model is continuously trained through the roadbed settlement prediction training set.

[0059] Embodiment 2, a roadbed settlement prediction system, see Figure 1 , comprising:

[0060] The roadbed excitation module is used to start the exciter at the roadbed monitoring time point, and perform a frequency sweep operation for a preset sweep time, the preset sweep time is generally 3-5s, and the main response frequency is determined, the main response frequency is the most significant vibration frequency of the roadbed structure under the action of external excitation, and reflects the dynamic characteristics of the roadbed, especially when subjected to periodic or continuous excitation, the response of the roadbed to the excitation is the most intense frequency band, and the exciter is set based on the main response frequency, and a single-frequency excitation operation is continuously performed, the excitation response signal is collected by the sensor group for a preset collection time, and the preset collection time is generally 5min. The preset collection time single-frequency excitation operation refers to an excitation operation of a fixed frequency (i.e. the main response frequency) through the exciter, and it should be noted that when the main response frequency of the roadbed corresponds to the frequency applied by the exciter, the roadbed will have a more significant response. In the settlement prediction, this sensitive change can provide us with early warning signals about the stability or potential problems of the roadbed; the sensor group here generally includes a ground sound sensor (which can measure the characteristics of the ground sound wave after the excitation signal is applied), a soil resistivity sensor (which is used to measure the resistivity of the foundation), and a geomagnetic sensor (which is used to measure the magnetic field data of the roadbed), etc.

[0061] The response feature spectrum construction module is configured to perform a short-time Fourier transform on the excitation response signal, and determine a response feature spectrum corresponding to the excitation response signal, wherein the response feature spectrum includes characteristic information such as a third harmonic energy ratio, a Lyapunov index, a peak delay time, and a WVD energy perturbation rate, and can reflect changes in the material structure of the roadbed, and further can reflect the settlement risk of the roadbed.

[0062] The first roadbed settlement prediction module is configured to perform a UMAP dimension reduction operation on the response feature spectrum, to obtain a plurality of embedded feature points, the horizontal coordinate of the embedded feature point being a timestamp, and the vertical coordinate being a data value after UMAP dimension reduction, and then add all the embedded feature points to the corresponding evolution trajectory graph. The data value after UMAP dimension reduction can also be two-dimensional, so the corresponding embedded feature point is a point in a three-dimensional space. Then, the evolution trajectory graph is subjected to anomaly identification. The anomaly identification refers to analyzing whether the embedded feature points in the evolution trajectory graph conform to the normal change trend. If the identification result is abnormal, it is considered that there is a roadbed settlement risk, otherwise it is considered that there is no roadbed settlement risk. It should be noted that when the roadbed is subjected to external force, disturbance or environmental change, it will exhibit a certain response mode. For example, after the soil body or structure is excited, the characteristic signals such as vibration, temperature and electromagnetic field will exhibit characteristics such as frequency, amplitude and phase. With the settlement, the physical properties of the soil body such as elasticity, pore structure, humidity and soil pressure change, which will cause significant changes in the nonlinear response characteristics of the soil body and structure. The differences between the normal evolution trajectory and the abnormal evolution trajectory in the evolution trajectory graph mapped by these nonlinear response characteristics can be identified. The appearance of the abnormal trajectory usually indicates the potential occurrence of settlement.

[0063] The second roadbed settlement prediction module is configured to send the response feature spectrum into a roadbed settlement prediction model for processing, to output a roadbed settlement prediction result, wherein the roadbed settlement prediction result includes the existence of a roadbed settlement risk and the non-existence of a roadbed settlement risk. The roadbed settlement prediction model is constructed based on an LSTM model. In order to output the roadbed settlement prediction result, the roadbed settlement prediction model further includes a full connection operation.

[0064] It should be understood that those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall fall within the protection scope of the appended claims of the present application. The parts not described in detail in the specification belong to the prior art known to those skilled in the art.

Claims

1. A method of predicting subgrade settlement, characterized by, Comprise: At the roadbed monitoring time point, the exciter is started, and a frequency sweep operation of a preset sweep time is performed to determine a main response frequency, and then the exciter is set based on the main response frequency, and a single-frequency excitation operation is continuously performed, and an excitation response signal of a preset collection time is collected through the sensor group; Performing short-time Fourier transform on the excitation response signal to determine the response characteristic spectrum corresponding to the excitation response signal, the response characteristic spectrum includes third harmonic energy ratio, Lyapunov index, peak delay time and WVD energy disturbance rate; Performing UMAP dimension reduction operation on the response characteristic spectrum to obtain a plurality of embedded feature points, the horizontal coordinate of the embedded feature point is the timestamp, and the vertical coordinate is the data value after UMAP dimension reduction, then all embedded feature points are added to the corresponding evolution trajectory diagram, and then the evolution trajectory diagram is identified for abnormality, if the identification result is abnormal, it is considered that there is a risk of roadbed settlement, otherwise it is considered that there is no risk of roadbed settlement; Performing short-time Fourier transform on the excitation response signal to determine the response characteristic spectrum corresponding to the excitation response signal, comprising the following steps: Perform the following operations on each sub-channel excitation signal in the excitation response signal: divide the sub-channel excitation signal into a plurality of sub-channel excitation signal segments according to the short-time window, and then perform third harmonic energy ratio extraction operation, Lyapunov index extraction operation, peak delay time extraction operation and WVD energy disturbance rate extraction operation on each sub-channel excitation signal segment, and then splice the outputs of the third harmonic energy ratio extraction operation, Lyapunov index extraction operation, peak delay time extraction operation and WVD energy disturbance rate extraction operation on the sub-channel excitation signal segment to obtain the response feature, and label the response feature with the timestamp; Splice all response features under the same timestamp to construct a comprehensive response feature, and all labeled comprehensive response features with timestamps form a response characteristic spectrum; The UMAP dimension reduction operation specifically includes the following contents: Step S1: Calculate the similarity between any two comprehensive response features, and use the k-nearest neighbor algorithm to select the k nearest neighbors of each comprehensive response feature, and construct a similar edge between the comprehensive response feature and its corresponding nearest neighbor comprehensive response feature, the weight of each similar edge is the similarity between the two comprehensive response features, and all comprehensive response features and similar edges form a high-dimensional KNN graph; Step S2: Randomly initialize a low-dimensional data point for each comprehensive response feature in a low-dimensional space, calculate the similarity between any two low-dimensional data points, and use the k-nearest neighbor algorithm to select the k nearest neighbors of each low-dimensional data point, and construct a similar edge between the low-dimensional data point and its corresponding nearest neighbor low-dimensional data point, the weight of each similar edge is the similarity between the two low-dimensional data points, and all low-dimensional data points and similar edges form a low-dimensional KNN graph; Step S3: Construct a cost function, the cost function includes the cross loss entropy between the high-dimensional KNN graph and the low-dimensional KNN graph, and use the gradient descent algorithm to optimize the cost function to adjust the low-dimensional data points in the low-dimensional space; Step S4: repeating steps S2-S3 until the cost function converges, and taking the low-dimensional data points corresponding to the integrated response features as the output of the UMAP dimensionality reduction operation performed on the integrated response features.

2. The method of predicting subgrade settlement according to claim 1, wherein, The exciter is started, and a frequency sweep operation of a preset sweep time is performed to determine the main response frequency, specifically including the following steps: A linear frequency sweep signal is set for the exciter, the exciter performs excitation operation, and excitation response signals of a preset sweep time are collected, and the excitation response signals include a plurality of sub-channel excitation signals, each sub-channel excitation signal corresponding to one sensor in the sensor group; For each sub-channel excitation signal, a Fourier transform is performed to obtain a sub-channel excitation spectrum corresponding to each sub-channel excitation signal, and then a peak detection algorithm is used to determine an energy peak value corresponding to each sub-channel excitation spectrum, and the frequency corresponding to the energy peak value is recorded as a response frequency, and then a weighted sum operation is performed on all response frequencies and corresponding energies to obtain the main response frequency, and the energy corresponding to the response frequency is the square of the energy peak value corresponding to the response frequency.

3. The method of predicting subgrade settlement according to claim 2, wherein, Abnormal identification is performed on the evolution trajectory diagram, specifically including the following steps: At the current roadbed monitoring time point, unsupervised clustering is performed on all embedded feature points in the evolution trajectory diagram, and the clustering cluster including the most embedded feature points after the unsupervised clustering is performed is recorded as the target clustering cluster, the ratio between the number of embedded feature points in the target clustering cluster and the total number of all embedded feature points is recorded as the evolution stability value, and then it is judged whether the evolution stability value is higher than the stability threshold, if the evolution stability value is higher than the stability threshold, the identification result is output as normal, and if the evolution stability value is not higher than the stability threshold, the identification result is output as abnormal.

4. The method of predicting subgrade settlement according to claim 3, wherein, It also includes performing roadbed settlement prediction through machine learning, specifically including the following contents: The response feature spectrum is sent into the roadbed settlement prediction model for processing, and a roadbed settlement prediction result is output, the roadbed settlement prediction result including existing roadbed settlement risk and non-existing roadbed settlement risk, the roadbed settlement prediction model being constructed based on an LSTM model, and the roadbed settlement prediction model further including a full connection operation; The roadbed settlement prediction model is trained, specifically including the following contents: a plurality of roadbed settlement prediction training samples are obtained, the roadbed settlement prediction training samples including response feature spectra, the roadbed settlement prediction training samples are labeled through roadbed settlement prediction results, all labeled roadbed settlement prediction training samples are combined to form a roadbed settlement prediction training set, and the roadbed settlement prediction model is trained through the roadbed settlement prediction training set, and the training target is the labeled roadbed settlement prediction result.

5. A system for predicting subgrade settlement, the system comprising: The system applies the roadbed settlement prediction method of any one of claims 1-4, including: A roadbed excitation module is configured to start the exciter at a roadbed monitoring time point, perform a frequency sweep operation of a preset sweep time to determine a main response frequency, and then set the exciter based on the main response frequency to continuously perform a single-frequency excitation operation and collect excitation response signals of a preset collection time through the sensor group. The response feature spectrum construction module is configured to perform a short-time Fourier transform on the excitation response signal, determine a response feature spectrum corresponding to the excitation response signal, and include a third harmonic energy ratio, a Lyapunov index, a peak delay time, and a WVD energy perturbation rate in the response feature spectrum. The first roadbed settlement prediction module is configured to perform a UMAP dimension reduction operation on the response feature spectrum, obtain a plurality of embedded feature points, the horizontal coordinate of the embedded feature point being a timestamp and the vertical coordinate being a data value after the UMAP dimension reduction, add all the embedded feature points to a corresponding evolution trajectory graph, and perform abnormality identification on the evolution trajectory graph. If the identification result is abnormal, it is considered that there is a risk of roadbed settlement, otherwise it is considered that there is no risk of roadbed settlement. The second roadbed settlement prediction module is configured to send the response feature spectrum to a roadbed settlement prediction model for processing, output a roadbed settlement prediction result, and include a risk of roadbed settlement and no risk of roadbed settlement. The roadbed settlement prediction model is constructed based on an LSTM model, and further includes a full connection operation.

Images