Highway subgrade velocity field intelligent imaging method based on swarm intelligence enhancement decision

Abstract

The invention discloses a highway subgrade velocity field intelligent imaging method based on swarm intelligence enhancement decision, which combines a swarm intelligence optimization algorithm and deep reinforcement learning to enable the algorithm not to be gathered near a local better solution too early in the search process, enhances the search ability of a group, expands the range of a solution space exploration, and improves the imaging efficiency of a highway subgrade velocity field. Local convergence and premature phenomena are effectively avoided, so that the imaging precision is remarkably improved, and the imaging result can more accurately reflect the real structure condition of the highway subgrade; the optimal parameter combination can be automatically searched by means of deep reinforcement learning, and the optimization difficulty caused by the fact that the search result is affected by numerous parameters and different combinations is reduced; under the assistance of deep reinforcement learning, the stability of the algorithm in a multi-extremum space is improved, the multi-valued problem caused by random parameters is effectively solved, the robustness of the algorithm is further improved, and the requirements of actual engineering for efficient and high-precision imaging are met.

Description

Technical Field

[0001] The present invention relates to an imaging method, in particular to an intelligent imaging method for the speed field of highway subgrade based on swarm intelligence enhanced decision-making. Background Art

[0002] In the inspection and maintenance of highway subgrade structures, non-destructive testing methods play a crucial role. They can obtain internal information without damaging the subgrade structure, providing important basis for highway maintenance. Currently, the main non-destructive testing techniques include ground penetrating radar method, ultrasonic testing method, and electrical resistance tomography method. The ground penetrating radar method can detect underground media, but the detection depth and resolution decrease in high-conductivity media, and signal interpretation depends on experience and is prone to misjudgment. The ultrasonic testing method can detect changes in material density, but the effective depth is limited, and it is greatly affected by the shape and pores of subgrade materials. The electrical resistance tomography method forms an image by measuring the potential difference through electrode currents, but the electrode arrangement affects data collection and the resolution is low.

[0003] Different from these traditional methods, the Rayleigh wave detection method has significant advantages. Rayleigh waves can propagate along the ground surface, and their dispersion characteristics are sensitive to underlying materials, capable of accurately inferring subgrade physical parameters, with a large and stable detection depth, not affected by the complex shape and pores of subgrade materials, simple data collection, strong anti-interference ability, and high imaging resolution. However, there are still problems with the Rayleigh wave dispersion curve inversion subgrade structure imaging method based on traditional swarm intelligence optimization algorithms as follows:

[0004] I. Low imaging accuracy and poor imaging effect of subgrade structure

[0005] In the inversion of Rayleigh wave dispersion curves, traditional swarm intelligence optimization algorithms are prone to local convergence and premature problems. During algorithm search, some individuals gather around local optimal solutions prematurely, weakening the swarm search ability and unable to fully explore the solution space. Taking the particle swarm algorithm as an example, particles overly follow the current optimal particle and often fall into local optima, making it difficult to find the global optimal solution. This directly affects the imaging accuracy of the subgrade structure, resulting in a deviation between the imaging result and the actual subgrade structure and being unable to accurately reflect the real situation.

[0006] II. Insufficient robustness and low stability of the subgrade structure imaging algorithm

[0007] In the inversion of Rayleigh wave dispersion curves, traditional swarm intelligence optimization algorithms face severe challenges. The algorithm itself has numerous parameters, and a slight difference in parameter settings will greatly affect the algorithm performance. Moreover, the inversion problem has the characteristic of multiple extreme values, and different parameter combinations will produce different search results, making it extremely difficult to find the optimal parameter combination. The algorithm is also extremely prone to falling into local extrema and difficult to obtain the global optimal solution, which makes the optimization speed slow, the imaging resolution and speed poor, and it is difficult to meet the urgent needs of actual engineering for high-efficiency and high-precision imaging. Summary of the Invention

[0008] To solve the defects existing in the prior art, the present invention discloses an intelligent imaging method for the speed field of a highway subgrade based on swarm intelligence enhanced decision-making, and its technical solution is as follows:

[0009] An intelligent imaging method for the speed field of a highway subgrade based on swarm intelligence enhanced decision-making, characterized in that:

[0010] S1. Obtain the vibration field signal of the internal structure of the highway subgrade to obtain an integrated vibration field time-domain diagram;

[0011] S2. Invert the one-dimensional subgrade speed field structure below the measuring point No. 1;

[0012] S3. Repeat step S2 to obtain the corresponding one-dimensional speed field structure curves of measuring points 2 - measuring point N, and convert the N one-dimensional speed field structure curves into a two-dimensional subgrade speed field structure diagram below the measuring line 1 through two-dimensional linear interpolation. Among them, the number of measuring points needs to satisfy N≥12, mainly based on the requirements of the spatial sampling theorem and inversion stability. According to the Nyquist criterion, insufficient measuring point density (N<12) will lead to spatial aliasing, making it difficult to distinguish small-scale geological anomalies, and the one-dimensional inversion error is amplified when interpolating between sparse measuring points, generating false anomalies or masking the true structure. Although having more than 12 measuring points (such as 24) can improve the imaging accuracy of complex strata, the cost and efficiency need to be weighed. Therefore, it is recommended to adjust dynamically according to the exploration depth;

[0013] S4. Repeat step S3 to obtain the two-dimensional subgrade speed field structure diagrams corresponding to measuring lines 2 to measuring line M. Among them, measuring line M≥4. Based on the requirements of the spatial sampling theorem, the number of measuring lines less than 4 will lead to spatial aliasing and is not conducive to three-dimensional interpolation;

[0014] S5. Integrate the two-dimensional subgrade speed field structure diagrams of the M measuring lines described in S4, establish a three-dimensional subgrade speed field by three-dimensional interpolation method, and use the specific geological exploration report and geological map data of the highway test section to locally correct the speed field, so as to obtain the final three-dimensional speed field model.

[0015] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program controls the device where the non-volatile storage medium is located to execute the above method when running.

[0016] The present invention also discloses a terminal device, characterized in that the terminal device includes: a processor, a memory, a communication interface and a bus; the processor, the memory and the communication interface are connected through the bus and complete communication with each other; the memory stores executable program code; the processor runs the program corresponding to the executable program code by reading the executable program code stored in the memory to be used to execute the above method.

[0017] Beneficial effects

[0018] Utilize the decision-making advantage of deep reinforcement learning and apply it to the multi-parameter swarm intelligence optimization iteration of the subgrade structure; by combining the swarm intelligence optimization algorithm and deep reinforcement learning, deep reinforcement learning can continuously improve the inversion process based on previous results and real-time feedback; this enables the algorithm not to prematurely converge near local optimal solutions during the search process, enhances the search ability of the swarm, expands the scope of the exploration solution space, effectively avoids local convergence and premature phenomena, thereby significantly improving the imaging accuracy and enabling the imaging results to more accurately reflect the true structural conditions of the highway subgrade.

[0019] Traditional algorithms face optimization difficulties brought about by multiple parameters and multiple extreme values, resulting in poor imaging effects. The deep reinforcement swarm intelligence optimization algorithm of the present invention adopts an intelligent parameter adjustment method. With the help of deep reinforcement learning, it can automatically find the optimal parameter combination, greatly reducing the optimization difficulty caused by numerous parameters and different combinations affecting the search results; at the same time, with the assistance of deep reinforcement learning, the stability of the algorithm in the multi-extreme value space is improved, effectively overcoming the multi-value problem caused by random parameters, thereby enhancing the robustness of the algorithm and meeting the requirements of actual engineering for efficient and high-precision imaging. Brief description of the drawings

[0020] Figure 1 Schematic diagram of the layout of sensors for subgrade structure detection;

[0021] Figure 2 Flowchart of the deep reinforcement swarm intelligence optimization algorithm for inverting Rayleigh wave dispersion curves;

[0022] Figure 3 Schematic diagram of the principle of calculating the dispersion energy spectrum by the phase shift method;

[0023] Figure 4 Block diagram of the overall network model of deep reinforcement learning;

[0024] Figure 5 Schematic diagram of the structures of the deep decision network and the target deep decision network;

[0025] Figure 6 Structure diagram of interpolating a two-dimensional velocity field from a one-dimensional velocity field. Detailed implementation manners

[0026] The present invention proposes an intelligent optimization method for inverting Rayleigh wave dispersion curves of highway subgrades, and the overall design idea is as Figure 1As shown in the figure, (1) Obtain the vibration field signal of the internal structure of the highway subgrade, set appropriate acquisition parameters for multi-channel surface wave data acquisition, and perform data processing such as signal denoising on the acquired data; (2) Extract the actual dispersion curve corresponding to the vibration field information, set an appropriate spatial window length, perform wave field transformation on the surface wave data acquired within this window, convert it from the time-space domain to the frequency-velocity domain, and extract the dispersion curve based on the energy amplitude and trend of the obtained dispersion spectrum. Establish an initial one-dimensional velocity field model of the subgrade below the measuring point, and continuously optimize the subgrade velocity field model by the method of swarm intelligence combined with deep reinforcement decision-making with the actual dispersion curve as the standard, so as to obtain the one-dimensional velocity field structure below the center measuring point of the spatial window; (3) Repeat step 2 to obtain the one-dimensional velocity field structures below all measuring points on the measuring line, and integrate and splice them, and interpolate to obtain the two-dimensional subgrade velocity field structure below the measuring line; (4) Obtain the two-dimensional subgrade velocity field structures below each measuring line by the same method, and restore the three-dimensional velocity field structure model of the subgrade through interpolation.

[0027] S1. Obtain the vibration field signal of the internal structure of the highway subgrade to obtain the integrated vibration field time-domain diagram.

[0028] 12 vibration sensors are linearly arranged on both sides of each one-way single lane, with a sensor spacing of 2 m, as specifically shown in Figure 1 the figure. At the same time, use a coupling agent (such as coupling glue or double-sided tape) to ensure good coupling between the geophone and the ground to reduce energy loss and interference.

[0029] Use a multi-channel vibration signal acquisition system to acquire vibration signals, set the sampling frequency to 20 kHz, the acquisition mode to trigger acquisition, and the sampling time to 2 seconds. After the signal acquisition is completed, integrate the signals of the sensor array to obtain the corresponding integrated vibration field time-domain diagram.

[0030] S2. Invert the one-dimensional subgrade velocity field structure below the No. 1 measuring point.

[0031] The present invention proposes an optimization method for inverting the subgrade velocity field of a highway based on swarm intelligence reinforcement decision-making. By introducing deep reinforcement learning to replace the particle update process in the traditional particle swarm algorithm, the intelligent optimization of the search process is realized. The improved overall design block diagram is as shown in Figure 2 the figure.

[0032] S2.1 Obtain the dispersion curve corresponding to the measuring point

[0033] The present invention adopts the phase shift method, sets a reasonable spatial window length for the target measuring point, performs wave field transformation on the surface wave data acquired within this window, and obtains the dispersion energy spectrum. Through the principle of the maximum energy spectrum amplitude, the measured Rayleigh wave dispersion curve c is extracted from the dispersion energy spectrum obs .

[0034] S2.2 Establish the initial multi-parameter model of the subgrade structure under the measuring point

[0035] Based on the fast vector transfer algorithm for calculating the Rayleigh wave dispersion curve in layered media by axisymmetric cylindrical Rayleigh wave, set the basic structure of the inversion model of the subgrade structure under the measuring point. The parameters include the number of strata d, the shear wave velocity of the strata The longitudinal wave velocity of the strata The density of the strata ρ = [ρ1, ρ2, ···, ρ d T , and the thickness of the strata H = [H1, H2, ···, H d T , as well as the search range of each type of parameter (V smin , V smax ), (V pmin , V pmax ), (ρ min , ρ max ), (H min , H max ).

[0036] S2.3 Take the multi-parameters of the subgrade structure as population particles and initialize them

[0037] Combined with the geological structure of the expressway in Shanxi Province, taking the velocity field set as 6 layers as an example, particle initialization is carried out. Set the population particle i represents the i-th particle. The initial population size M is 30, the number of strata parameter d = 6, and each layer has four parameters V S , V P , ρ, H. Therefore, the particle dimension D = 4xd is 24.

[0038] According to the subgrade characteristics of the Shanxi plain area, combined with the engineering geology handbook and engineering experience, the setting of the inversion search range of the shear wave velocity, longitudinal wave velocity, density and stratum thickness of each layer is shown in the following table

[0039]

[0040] Use the chaos optimization strategy to initialize the particle swarm according to the known prior information of the highway subgrade to form the first-generation particles.

[0041] S2.4 Adopt the fast vector transfer algorithm to perform forward simulation on the multi-parameters of the subgrade structure to obtain the simulated dispersion curve under the corresponding measuring point

[0042] ​​The multi-parameter particle swarm initialized in S2.3 is used to calculate the wave field propagation by the fast vector transfer algorithm. This algorithm processes the solution of the wave equation through vectorized transfer and numerically simulates the Rayleigh wave propagation in the simulation area to generate Rayleigh wave field data at different frequencies. By processing the wave field data, the corresponding dispersion curve c is extracted. cal 。

[0043] In S2.5, the simulated dispersion curve generated in S2.4 and the measured dispersion curve in S2.1 are used to construct the fitness function for particle update.

[0044] The present invention uses the mean square error function of the simulated dispersion curve c cal and the measured dispersion curve c obs as the fitness function for particle swarm update, and its expression is shown in Equation (2-1).

[0045]

[0046] The simulated dispersion curve c generated in S2.4 cal and the measured dispersion curve c in S2.1 obs are substituted into Equation (2-1) to obtain the fitness value, where m is the index variable of the dispersion curve data sample. M is the total number of frequency sampling points in the dispersion curve. For example, when the frequency range is 10 Hz - 100 Hz and the step size is 1 Hz, M = 91.

[0047] In S2.6, the velocity field particle swarm is updated using deep reinforcement learning

[0048] As Figure 4 shown, the deep reinforcement learning network consists of a deep decision network, a target deep decision network, a search strategy model, and a reinforcement learning module.

[0049] In S2.6.1, the deep decision network and the target deep decision network are designed

[0050] The deep decision network and the target deep decision network are the same network. The deep decision network copies the training parameters to the target deep decision network, and the target deep decision network is used to estimate the target value and use the target value as part of the label of the deep decision network.

[0051] The deep decision network and the target deep decision network adopt the same form of deep neural network (DNN), including three hidden layers with widths h1 = 10h in , h3 = 10h out and (see Figure 5 ), where h in (h outare the widths of the input (output) layer respectively. The hidden layer uses the activation function tanh, while the output layer uses a simple linear link, where the input h in is the state of the particle. According to the initial multi-parameter model of the subgrade structure, the present invention intends to use 24-dimensional particles, and the output h out is the probability distribution of each discrete action for each dimension of data, with a width of 24 * 9.

[0052] S2.6.2 Design the search strategy model

[0053] The search strategy model takes the value range of each parameter in each particle as the overall search area. For a certain dimension parameter X i,j (t) of the current particle, define the action of "moving in the increasing direction of parameter X i,j (t)", that is, let the value of X i,j (t) move a set step size ΔX i,j (t) in the increasing direction within the particle search range. The step size ΔX i,j (t) is different for different dimension parameters of the particle. Specifically, according to the population particles its parameter types can be divided into four categories V s / (m·s -1 ), V p / (m·s -1 ), ρ / (g·cm -3 ), H / (m). According to the inversion search range of different parameters, determine the step size ΔV s = 1, ΔV p = 1, Δρ = 0.1, ΔH = 0.1.

[0054] Similarly, for each parameter, define such an action a i,j (t) of moving in the increasing direction. The action space of each parameter type consists of the following nine possible actions:

[0055] [-4ΔV s , -3ΔV s , -2ΔV s , -1ΔV s , 0, ΔV s , 2ΔV s , 3ΔV s , 4ΔV s

[0056] [-4ΔV p , -3ΔV p , -2ΔV p , -1ΔV p , 0, ΔV p , 2ΔV p , 3ΔV​p , 4ΔV p

[0057] [-4Δρ, -3Δρ, -2Δρ, -1Δρ, 0, Δρ, 2Δρ, 3Δρ, 4Δρ]

[0058] [-4ΔH, -3ΔH, -2ΔH, -1ΔH, 0, ΔH, 2ΔH, 3ΔH, 4ΔH]

[0059] Therefore, a i,j (t) represents the discrete action of the j - dimensional data of the i - th particle at the t - th iteration. Each action represents a multiple of the step size by which the parameter of this dimension is increased or decreased based on the current value.

[0060] The target depth decision network outputs the probability distribution of each discrete action. The particle selects an action a i,j (t) according to this probability distribution, and then updates its state (i.e., position) according to this action. The update formula for the j - dimensional position of the i - th particle at the t - th iteration is shown in Equation (2 - 2),

[0061] X i,j (t + 1) = X i,j (t)+α i,j (t) (2 - 2)

[0062] S2.6.3 Design the reward function and value function

[0063] The reinforcement learning module is used to define whether the current action is positive or negative and feedback an immediate reward function value during the interaction process; the specific formula for the reward function value is:

[0064] R i (t) = ln(1 + |F i (t)-F i (t + 1)|) (2 - 3)

[0065] where F i (t), F i (t + 1) obtain the fitness values of the t - th generation particles and the (t + 1) - th generation particles through Equation (2 - 1) in Steps S3.3 and S3.4. Using the above - mentioned reward function can make the particles continuously update towards the minimum fitness value while avoiding the phenomenon of too small rewards.

[0066] Design the value function. When each particle is updated, the reward function value of the current action is obtained through the fast vector transfer algorithm. When the particle is at the current position X i (t), the position X i (t + 1) after taking the action a i (t) and the reward function value R t+1 ​Independent of the historical position, only related to the current position and action. That is, at position X i (t), take action a i (t) of the long-term expectation Q * (X i (t), a i (t)) The return as the value function can be expressed as:

[0067]

[0068] where γ is the discount factor of the long-term return, a i (t) is the action taken at the next moment. In order to enable the learning of Q to have predictive ability, use a deep decision network to fit Q * (X i (t), a i (t)).

[0069] S2.6.4 Training the deep reinforcement learning network

[0070] A. According to the search strategy model, using the measured dispersion curve as the ultimate goal of the contemporary particle iteration, generate the next generation of particles by the actions output by the target deep decision network with the initialized particles, and obtain the reward value under the corresponding actions;

[0071] B. Input the particles in stage A and their corresponding fitness values into the deep decision network to predict the value of the corresponding execution actions in stage A;

[0072] C. Copy the parameters of the deep decision network to the target deep decision network, and input the next generation of particles generated in step C into the target deep decision network to obtain the maximum value function value. Add this value to the reward function value of the corresponding action in step A as the target value label.

[0073] D. Optimize the deep decision network using the predicted value in step B and the target value label generated in step C;

[0074] E. Regard the next generation of particles obtained in step A as the current particles, and loop through steps A to D until the optimal fitness value is reached to complete the network optimization.

[0075] S2.7 Output the last generation of particles that reach the optimal fitness value to obtain the one-dimensional velocity field structure below the target measurement point.

[0076] When the fitness value of a particle in a generation of particles output by the deep reinforcement learning network reaches the convergence condition (the fitness value is less than 5% and the change in five consecutive generations is less than 1%), stop the iteration and output this particle as the optimal particle to obtain

[0077]

[0078] wherein is the optimal shear wave velocity of the j-th layer, is the optimal P-wave velocity of the j-th layer, ρ j is the optimal density of the j-th layer, H j is the optimal thickness of the j-th layer. The optimal one-dimensional velocity field structure is shown in the form of a curve through one-dimensional interpolation as: V s (z), V p (z), ρ(z), where V s , V p , ρ are multi-parameters of the subgrade structure, and z is the depth.

[0079] S3. Repeat step S2 to obtain the corresponding one-dimensional velocity field structure curves for measuring points 2 - 12, and convert the 12 one-dimensional velocity field structure curves into a two-dimensional subgrade velocity field map under survey line 1 through two-dimensional linear interpolation, as shown in Figure 6 .

[0080] S4. Repeat step S3 to obtain the two-dimensional subgrade velocity field maps corresponding to survey lines 2 to 4.

[0081] S5. Integrate the two-dimensional subgrade velocity field maps of the 4 survey lines described in S4, establish a three-dimensional subgrade velocity field using three-dimensional interpolation, and perform local correction on the velocity field using specific geological exploration reports, geological maps, etc. of the highway test section, so as to obtain the final three-dimensional velocity field model.

[0082] The present invention proposes an intelligent imaging method for highway subgrade structure based on swarm intelligence enhanced decision-making. First, ground vibration signals are collected by using a vibration sensor array, and then signal filtering is performed to remove noise. Then, the phase shift method is used to obtain wave velocity information at different frequencies, thereby extracting the dispersion curve. The core innovation lies in the use of the swarm intelligence enhanced decision-making optimization algorithm in the inversion process, which combines the swarm intelligence optimization algorithm and deep reinforcement learning. The swarm intelligence method effectively explores the solution space, while deep reinforcement learning enhances the optimization by learning previous results and improving the inversion process according to real-time feedback. This combination not only improves the accuracy of the inversion but also speeds up the convergence rate, ultimately providing a more accurate highway subgrade velocity field structure. This method solves the computational problem and improves the reliability of subgrade condition assessment, which is a major progress in this field.

[0083] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent imaging method for the speed field of highway subgrade based on swarm intelligence enhanced decision-making, characterized by: S1. Obtain the vibration field signal of the internal structure of the highway subgrade to obtain the integrated vibration field time-domain diagram; S2. Invert the one-dimensional subgrade speed field structure under the measuring point No. 1; S3. Repeat step S2 to obtain the corresponding one-dimensional speed field structure curves of measuring points 2 - measuring point N, and convert the N-way one-dimensional speed field structure curves into the two-dimensional subgrade speed field diagram under the measuring line 1 through two-dimensional linear interpolation, where the measuring point N ≥ 12; S4. Repeat step S3 to obtain the two-dimensional subgrade speed field diagrams corresponding to measuring lines 2 to measuring line M, where the measuring line M ≥ 4; S5. Integrate the two-dimensional subgrade speed field diagrams of the M measuring lines described in S4, establish a three-dimensional subgrade speed field by three-dimensional interpolation method, and use the specific geological exploration report and geological map data of the highway test section to locally correct the speed field, so as to obtain the final three-dimensional speed field model.

2. The intelligent imaging method for the speed field of highway subgrade based on swarm intelligence enhanced decision-making according to claim 1, characterized in that, The said step 1 includes the following contents: A plurality of vibration sensors are linearly arranged on both sides of each one-way single lane, and a couplant is used to ensure the coupling of the geophone with the ground; a multi-channel vibration signal acquisition system is used to collect vibration signals, the sampling frequency is set, the acquisition mode is trigger acquisition, and the sampling time is set; after the signal acquisition is completed, the signals of the sensor array are integrated to obtain the corresponding integrated vibration field time-domain diagram.

3. The intelligent imaging method of the highway subgrade speed field based on swarm intelligence enhanced decision-making according to claim 1, characterized in that, The said step 2 includes the following contents: S2.1: Obtain the dispersion curve corresponding to the measuring point; S2.2: Establish the initial multi-parameter model of the subgrade structure under the measuring point; S2.3: Take the multi-parameters of the subgrade structure as population particles and initialize them; S2.4: Adopt the fast vector transfer algorithm to perform forward simulation on the multi-parameters of the subgrade structure to obtain the simulated dispersion curve under the corresponding measuring point; S2.5: Use the simulated dispersion curve generated by S2.4 and the measured dispersion curve of S2.1 to construct the fitness function for particle update; S2.6: Adopt deep reinforcement learning to update the speed field particle swarm; S2.7: Output the last generation of particles that reach the optimal fitness value to obtain the one-dimensional speed field structure under the target measuring point.

4. The intelligent imaging method for the speed field of highway subgrade based on swarm intelligence enhanced decision-making according to claim 3, characterized in that, The said step S2.2 of establishing the initial multi-parameter model of the subgrade structure under the measuring point includes the following contents: Fast vector transfer algorithm for calculating Rayleigh wave dispersion curves in layered media based on axisymmetric cylindrical Rayleigh waves. Set the basic structure of the subgrade structure inversion model under the measuring point. The parameters include the number of strata d, the shear wave velocity of the strata The longitudinal wave velocity V of the strata P = The formation density ρ = [ρ1, ρ2, ···, ρ d T , the formation thickness H = [H1, H2, ···, H d T , and the search range of each type of parameter (V smin , V smax ), (V pmin , V pmax ), (ρ min , ρ max ), (H min , H max ); where ρ1, ρ2, ···, ρ d , H1, H2, ···, H d respectively represent the longitudinal wave velocity, shear wave velocity, density and thickness of the 1st, 2nd ··· dth layers in the formation structure below the measuring point.​​ 5. The intelligent imaging method for the speed field of highway subgrade based on swarm intelligence enhanced decision-making according to claim 3, characterized in that In step S2.4, the fast vector transfer algorithm is used to perform forward simulation on multiple parameters of the subgrade structure, and the simulated dispersion curves at corresponding measuring points are obtained, including the following: The multi-parameter particle swarm initialized in S2.3 is used to calculate the wave field propagation by the fast vector transfer algorithm. This algorithm processes the solution of the wave equation through vectorized transfer and numerically simulates the Rayleigh wave propagation in the simulation area to generate Rayleigh wave field data at different frequencies. By processing the wave field data, the corresponding dispersion curve c is extracted. cal ; Step S2.5 uses the simulated dispersion curve generated by S2.4 and the measured dispersion curve of S2.1 to construct the fitness function for particle update; Using the simulated dispersion curve c cal and the measured dispersion curve c obs the mean square error function is used as the fitness function for particle swarm update, and its expression is shown in Equation (2-1): where c obs (m) is the measured dispersion curve obtained in S2.1, c cal (m) is the simulated dispersion curve generated in S2.4, m is the index variable of the dispersion curve data sample, and M is the total number of samples of the dispersion curve data sample.

6. The intelligent imaging method of the highway subgrade velocity field based on swarm intelligence enhanced decision-making according to claim 3, characterized in that The said step S2.6 of adopting deep reinforcement learning to update the speed field particle swarm includes the following contents: S2.6.1: Design the deep decision network and the target deep decision network; S2.6.2: Design the search strategy model; S2.6.3: Design the reward function and the value function; S2.6.4: Train the deep reinforcement learning network.

7. The intelligent imaging method of the highway subgrade velocity field based on swarm intelligence enhanced decision-making according to claim 3, characterized in that, The said step S2.6.4 of training the deep reinforcement learning network includes the following contents: A. According to the search strategy model, taking the measured dispersion curve as the ultimate goal of the contemporary particle iteration, use the initialized particles to generate the next generation of particles through the actions output by the target deep decision network, and obtain the reward value under the corresponding actions; B. Input the particles and their corresponding fitness values in the A-step stage into the deep decision network to predict the value of the corresponding actions to be performed in the A-step stage; C. Copy the parameters of the deep decision network to the target deep decision network, and input the next-generation particles generated in step C into the target deep decision network to obtain the maximum value function value; add this value to the reward function value of the corresponding action in step A as the target value label; D. Optimize the deep decision network using the predicted value in step B and the target value label generated in step C; E. Regard the next-generation particles obtained in step A as the current particles, and loop through steps A to D until the optimal fitness value is reached to complete the network optimization.

8. The intelligent imaging method of the highway subgrade velocity field based on swarm intelligence enhanced decision-making according to claim 3, characterized in that The last generation of particles that reach the optimal fitness value in step S2.7 are output to obtain the one-dimensional velocity field structure below the target measurement point, including the following content: Stop the iteration when the fitness value of a particle in a generation of particles output by the deep reinforcement learning network reaches the convergence condition, and output the particle as the optimal particle to obtain ρ1, ρ2, ···, ρ6, H1, H2, ···, H6], where is the optimal shear wave velocity of the j-th layer, is the optimal P-wave velocity of the j-th layer, ρ j is the optimal density of the j-th layer, H j is the optimal thickness of the j-th layer; display the optimal one-dimensional velocity field structure in the form of a curve through one-dimensional interpolation as: V s (z), V p (z), ρ(z), where V s ,V p ,ρ are the multi-parameters of the subgrade structure, and z is the depth.

9. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the method according to any one of claims 1 to 8.

10. A terminal device, characterized in that, The terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor runs the program corresponding to the executable program code by reading the executable program code stored in the memory to be used to execute the method according to any one of claims 1-8 above.

Images