Deep learning-based crash barrier post depth inspection method, system and device

Abstract

The application discloses a kind of based on deep learning's crash barrier upright post depth inspection method, system and device, the method includes: constructing physical prior expression;Based on the physical prior expression, the detection process of the upright post is simulated under different working condition parameter combinations by numerical simulation, generates the virtual dataset containing electromagnetic response simulation signal and its corresponding upright post depth true value;The virtual dataset is input to the preset inversion model for training, and the constraint of physical prior expression is introduced in the training process to optimize model parameters, to obtain the inversion model after training is used to output depth detection result.Therefore, by introducing physical prior and physical simulation, the problems of high sample acquisition cost and strong destructive in traditional deep learning method in engineering detection are solved, and simulation data is used to provide sufficient and high-quality supervision information for the model.

Description

Technical Field

[0001] This invention relates to the fields of traffic engineering inspection technology and artificial intelligence application technology, and in particular to a method, system and device for checking the burial depth of crash barrier posts based on deep learning. Background Technology

[0002] Highway crash barriers are crucial facilities for ensuring traffic safety, and the accuracy of their post burial depth directly affects their crashworthiness. Currently, reliable testing of crash barrier post burial depth still heavily relies on destructive or semi-destructive methods, such as partial excavation. These methods have significant problems: first, they require damaging the roadbed or concrete foundation, affecting structural integrity; second, they are inefficient and difficult to conduct large-scale surveys; and third, they pose high construction safety risks and are extremely costly.

[0003] Currently, non-destructive testing technologies applicable to metal columns mainly include ground-penetrating radar and magnetic flux leakage or magnetic anomaly detection. However, due to the highly nonlinear relationship between signal response and burial depth, the influence of multiple factors, and the interference of the detection signal from the column diameter, material, corrosion state, and surrounding media (such as soil moisture content and concrete encapsulation), these methods generally suffer from insufficient accuracy in quantitative inversion of column burial depth. Furthermore, existing methods often rely on manual experience or simple empirical formulas for signal interpretation, leading to significant differences in results under different equipment, operators, and operating conditions.

[0004] Therefore, how to achieve non-destructive, high-precision, and robust column burial depth detection is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method, system, and device for inspecting the burial depth of guardrail posts based on deep learning. It introduces deep learning technology based on physical priors, solves the problem of data scarcity through physical simulation, addresses the problem of working condition differences by fusing physical parameters through a dual-branch network, and bridges the gap between simulation and reality through actual measurement calibration, thereby achieving high-precision intelligent detection of the burial depth of guardrail posts.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] In a first aspect, this application discloses a method for inspecting the burial depth of crash barrier posts based on deep learning, comprising: establishing a physical model of the electromagnetic response of a highway crash barrier post and its surrounding medium system, establishing key physical control parameters affecting the detection signal, and constructing a physical prior expression; based on the physical prior expression, simulating the detection process of the post under different combinations of working condition parameters through numerical simulation, generating a virtual dataset containing electromagnetic response simulation signals and their corresponding true values ​​of post burial depth; inputting the virtual dataset into a pre-set inversion model for training, and introducing constraints from the physical prior expression during the training process to optimize the model parameters, thereby obtaining a trained inversion model for outputting burial depth detection results.

[0008] Its effect is as follows:

[0009] By introducing physical priors and physical simulation, the problems of high sample acquisition costs and strong destructiveness faced by traditional deep learning methods in engineering inspection are solved, and simulation data is used to provide the model with sufficient and high-quality supervision information.

[0010] By adopting a dual-branch network structure, the medium parameters and structural parameters are explicitly incorporated into the model input, and a "signal-parameter" joint inversion framework is constructed, which improves the model's cross-condition transferability and engineering interpretability under different media and different column specifications from the source.

[0011] An adaptive calibration mechanism using a small amount of measured data bridges the gap between numerical simulation and real working conditions. While ensuring detection accuracy, it significantly reduces the need for on-site destructive verification, thereby significantly reducing detection costs and improving efficiency.

[0012] Secondly, this application discloses a deep learning-based system for checking the burial depth of crash barrier posts, comprising: a memory storing a computer program for implementing the deep learning-based method for checking the burial depth of crash barrier posts; and a processor for performing read and write operations on the memory to execute the deep learning-based method for checking the burial depth of crash barrier posts.

[0013] Thirdly, this application discloses a deep learning-based device for inspecting the burial depth of crash barrier posts, comprising: a signal acquisition instrument, a multi-frequency excitation transmitter, an environmental physical parameter input box, signal lines, and a deep learning model box; the deep learning model box houses the deep learning-based crash barrier post burial depth inspection system; the multi-frequency excitation transmitter emits electromagnetic waves toward the post and surrounding medium, and the signal acquisition instrument receives and acquires reflected, transmitted, or induced signals to form time-domain and frequency-domain electromagnetic response signals; the environmental physical parameter input box receives the acquired signals and inputs the physical parameters of the current detection environment; the electromagnetic response signals and physical parameter vectors are input through the signal lines into the deep learning model box, and the burial depth detection result is output through the trained inversion model. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of a deep learning-based method for checking the burial depth of crash barrier posts according to some embodiments of this application;

[0015] Figure 2 This is a simplified flowchart illustrating a deep learning-based method for checking the burial depth of crash barrier posts according to some embodiments of this application.

[0016] Figure 3 This is a structural schematic diagram of a deep learning-based anti-collision guardrail post burial depth inspection device according to some embodiments of this application.

[0017] In the picture:

[0018] 01 - Signal acquisition instrument, 02 - Multi-frequency excitation transmitter probe, 03 - Environmental physical parameter input box, 04 - Signal line, 05 - Deep learning model box. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] According to the deep learning-based method for checking the burial depth of crash barrier posts in the embodiments of this application, refer to... Figure 1 It is understood that this method can be implemented based on any electronic device with computing capabilities, such as a local computer, a cloud server, etc., including:

[0021] S1: Multiphysics response modeling and physical prior construction.

[0022] A physical model of the electromagnetic response of the highway crash barrier post and the surrounding medium system was established, and key physical control parameters affecting the detection signal were established as physical priors.

[0023] Specifically, this stage first models the non-destructive testing process from a physical mechanism perspective, treating the physical system consisting of the column and the medium as a typical electromagnetic coupling problem. By uniformly describing the interaction between the metal column and the surrounding medium under electromagnetic excitation, a physical process model is established, illustrating the propagation of the electromagnetic field in the medium, its induction in the metal conductor, and its attenuation and distortion with propagation distance and medium properties. This forms the electromagnetic physical model. Furthermore, key physical control parameters of the physical model are identified, including the column diameter, material, and the conductivity, dielectric constant, and moisture content of the surrounding medium (e.g., concrete / soil).

[0024] Then, by identifying key physical control parameters, such as signal attenuation and phase, the relationship between the amplitude, phase, and spectral characteristics of the probe signal and the column burial depth, the signal and depth, and environmental influences are analyzed. Based on the patterns of these parameter changes, they are organized into a constrained physical prior expression. In response to the differences in electromagnetic propagation mechanisms under different engineering conditions (concrete embedment, conventional soil embedment, and backfill soil embedment), three examples will be disclosed later for explanation.

[0025] The first example addresses concrete embedding conditions. Specifically:

[0026]

[0027] In the formula, The frequency of electromagnetic wave oscillation; The depth of the column; For a burial depth of The electromagnetic response signal obtained from actual measurement or simulation at the location includes amplitude and phase information; For the speed of transmission; It is the attenuation constant; The interface reflectance coefficient, , The wave impedance of the metal column is... The wave impedance of concrete, , The relative permittivity, Permeability, , Angular frequency, Let be the relative permittivity of concrete. The vacuum permittivity, The imaginary unit represents the orthogonal component of a complex number and is used to uniformly describe the polarization and loss of a medium.

[0028] Second example: Conventional soil burial conditions Specifically:

[0029]

[0030] In the formula, The attenuation coefficient is... , This is a constant term for the attenuation coefficient. frequency Power of 1 It is a dimensionless exponent. The frequency-varying coefficient of the attenuation coefficient; For phase velocity, , For the constant term of phase velocity, The frequency coefficient of the phase velocity. For frequency Power of 1 It is a dimensionless exponent; This is a dispersion path correction term. , For the partial derivative of phase velocity with respect to frequency, The frequency of electromagnetic wave oscillation; This refers to the depth of the column.

[0031] Third example: Backfill soil burial case Specifically, it includes:

[0032]

[0033] In the formula, For effective attenuation coefficient, , The intrinsic attenuation coefficient of the backfill matrix is ​​given by [the value of the backfill matrix]. The scattering attenuation coefficient; The scattered signal component; For effective wave velocity, , The intrinsic wave velocity of the backfill matrix, Porosity or volume fraction; path extension factor .

[0034] S2: Large-scale virtual sample generation based on physical simulation.

[0035] In practical engineering, the cost of actually sampling deeply buried columns is too high. In some preferred embodiments, the method introduces a virtual sample generation strategy based on physical simulation, which specifically includes the following steps:

[0036] S21. Based on the aforementioned physical prior expression, the system defines a simulation parameter space including burial depth range, medium parameters, column parameters, and detection conditions; specifically, the input engineering parameter space includes burial depth, column specifications, medium, etc.

[0037] S22. Using any electromagnetic field numerical simulation software, combine the parameters in the simulation parameter space to generate a batch of three-dimensional models, simulating the propagation of electromagnetic waves in the medium and their interaction with the column. Collect time-domain and frequency-domain response signals as part of the detection signal, and inject environmental interference that conforms to the actual measurement noise model to obtain a high-fidelity electromagnetic numerical simulation. The electromagnetic numerical simulation also requires inputting physical prior expressions as constraints; the specific simulation process is well known to those skilled in the art and will not be elaborated upon here.

[0038] S23. Extract multi-dimensional signal features related to burial depth (such as time domain peak value, arrival time, frequency domain main frequency offset, etc.) from the simulation results, and label the true value of burial depth and the working condition parameter vector for each sample.

[0039] S24. Organize the simulation samples into a dataset with paired annotations of the column burial depth corresponding to the detection signal, thereby obtaining a large-scale virtual dataset as the output. Perform quality control such as sample equalization and abnormal sample removal (based on physical prior constraints), and divide it into training set, validation set and test set, so as to output the virtual dataset in the form of a large-scale virtual dataset.

[0040] S3: Introducing a deep learning inversion model with physical priors.

[0041] In this embodiment, a deep learning response model configured as a signal-parameter joint conditional inversion framework is used for training. Specifically, the pre-configured inversion model's network structure employs a dual-branch design, which can combine... Figure 2 To understand this further, see the details below:

[0042] Based on the observed data, perform the following actions respectively: Input Dataset A: Input the observed electromagnetic signal response into the timing header; Input Dataset B: Input the physical and mechanical parameters into the parameter header.

[0043] The Time-Series Branch is the first branch of the inversion model. This branch takes the training set obtained from S2 as its input, serving as the original time-domain response signal. This branch employs a lightweight Transformer structure to extract long-range dependencies and multi-scale temporal features of the response signal, forming a depth-sensitive signal representation vector. For example, the observed electromagnetic signal response can be input, and the result output to the fusion layer.

[0044] The Parametric Branch is the second branch of the inversion model. This branch takes as input a vectorized description of the operating conditions and medium parameters (key physical control parameters identified by S2), i.e., physical and mechanical parameter inputs. This branch extracts features using a Multilayer Perceptron (MLP) to generate conditional embedding vectors, which characterize the modulation effect of the current operating condition on the electromagnetic response-burial depth mapping. The results are also output to the fusion layer.

[0045] In the fusion layer, the signal representation output from the time-series head and the conditional embedding output from the parameter head are concatenated and fused. Then, the MLP inversion prediction head outputs an estimated burial depth and can also output a confidence index. In some embodiments, this MLP inversion prediction head is specifically a neural network module based on a multi-layer perceptron (MLP), whose function is to convert the fused feature vector into an estimated value of the column burial depth and simultaneously output a confidence index or an uncertainty index.

[0046] The loss function is calculated based on the output. Specifically, in the loss function during model training, in addition to the usual burial depth regression error... In addition, constraints consistent with physical priors are added to penalize inversion results that clearly violate physical common sense (such as a mismatch between signal attenuation rate and depth), ensuring the physical consistency of the model. The total loss function is as follows:

[0047]

[0048] In the formula, As weight, For physical prior consistency constraints, This represents the depth regression error. Among them, The three operating conditions included in the foregoing embodiments are as follows: (Concrete embedding) (Conventional soil burial) and (Backfill soil is placed).

[0049] Furthermore, the parameter header or fusion layer of the inversion model integrates a hybrid expert module; the hybrid expert module includes: a gated network for calculating the weights of each expert network based on the input working condition parameter vector; and three expert networks, each corresponding to a physical prior expression under a preset working condition. It includes three operating conditions, as described above. , and The physical prior consistency constraint term is obtained by weighted summation of the outputs of each expert network based on the weights calculated by the gating network.

[0050] S4: Measured data calibration and adaptive correction of the model.

[0051] To ensure the accuracy of the prediction results, this stage is performed in a further embodiment, where a measured data calibration mechanism is introduced after the inversion model training is completed.

[0052] Specifically, this includes: inputting a small amount of high-reliability measured calibration data, such as obtaining a small number of real samples through small-scale on-site excavation or known calibration points. Based on this measured data, the model prediction bias is analyzed. Then, systematic biases caused by differences between the simulation environment and the real geological environment (such as soil heterogeneity and unknown underground clutter) are corrected, thereby analyzing the systematic biases, such as analyzing the simulated and measured signals, noise, and media, to achieve controlled fine-tuning. In this embodiment, a transfer learning or parameter fine-tuning strategy is adopted to perform controlled fine-tuning of the model parameters while keeping the network structure of the inversion model unchanged. In detail, the fine-tuning process is parameter adjustment under physical constraints ( Finally, by fine-tuning the output of the inversion model adapted to the project, it can be called the calibrated inversion model. This calibrated inversion model can be used for large-scale detection of the burial depth of highway crash barrier posts without the need for extensive destructive verification.

[0053] For example, the parameter fine-tuning of the trained inversion model includes: fusion layer weights: adjusting the fusion ratio of temporal features and conditional features to enhance the model's sensitivity to measured signal features. MLP inversion prediction head parameters: fine-tuning the weights and biases of the output layer to correct the burial depth regression bias. Conditional embedding vector mapping parameters: optimizing the MLP's encoding capability for operating condition parameters in the parameter header to improve cross-operating condition adaptability.

[0054] For example, the method for checking the burial depth of highway crash barrier posts described in the foregoing embodiments can be performed using a highway crash barrier post burial depth detection device, as referred to... Figure 3 The device, as understood, includes a signal acquisition unit, a multi-frequency excitation transmitter probe, an environmental physical parameter input box, signal lines, and a deep learning model box. Its use involves the following steps:

[0055] The first step is to use the multi-frequency excitation transmitter 02 of the electromagnetic excitation source to emit electromagnetic waves to the column and the surrounding medium.

[0056] The second step involves using signal acquisition instrument 01 to receive and collect reflected, transmitted, or induced signals, forming time-domain and frequency-domain electromagnetic response signals.

[0057] The third step is for the environmental physical parameter input box 03 to receive the signal from the signal acquisition instrument 01 and input the physical parameters of the current detection environment, including: column material, diameter, corrosion status, surrounding medium type, medium conductivity, dielectric constant, and moisture content.

[0058] The fourth step involves inputting the signal feature vector obtained in the second step and the operating parameters (physical parameters) obtained in the third step into the deep learning model box 05 via signal line 04. This input is either into the pre-trained inversion model or the calibrated inversion model described in the previous embodiment. The deep learning response model extracts the temporal features of the signal through a temporal branch and fuses the physical prior conditions through a parameter branch, ultimately outputting the confidence interval of the burial depth estimate.

[0059] It is understood that, since the inversion model in this embodiment is a lightweight model, it can be mounted in the deep learning model box to perform corresponding result prediction; the lightweighting method has been described in the foregoing method embodiments and will not be repeated here. A deep learning-based system for checking the burial depth of crash barrier posts according to an embodiment of this application includes: a memory storing a computer program for implementing the deep learning-based method for checking the burial depth of crash barrier posts; and a processor for reading and writing to the memory to execute the deep learning-based method for checking the burial depth of crash barrier posts.

[0060] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the present invention.

Claims

1. A method for checking the burial depth of crash barrier posts based on deep learning, characterized in that, include: A physical model of the electromagnetic response of the highway crash barrier post and its surrounding medium system is established, and key physical control parameters affecting the detection signal are determined to construct a physical prior expression. The physical prior expression can be any one of the following conditions: concrete embedment, conventional soil embedment, or backfill soil embedment. The physical prior expression for the concrete embedding condition is as follows: : In the formula, The frequency of electromagnetic wave oscillation; The depth of the column; For a burial depth of The electromagnetic response signal obtained from actual measurement or simulation at the location includes amplitude and phase information; For the speed of transmission; It is the attenuation constant; The interface reflectance coefficient; The physical prior expression for conventional soil burial conditions is: : In the formula, The attenuation coefficient is... , This is a constant term for the attenuation coefficient. frequency Power of 1 It is a dimensionless exponent. The frequency-varying coefficient of the attenuation coefficient; For phase velocity, , For the constant term of phase velocity, The frequency coefficient of the phase velocity. For frequency Power of 1 It is a dimensionless exponent; This is a dispersion path correction term; The physical prior expression for the backfill soil burial case is: : In the formula, For effective attenuation coefficient, The scattered signal component; For effective wave velocity, It is the path extension factor, and ; Based on the aforementioned physical prior expression, the detection process of the column under different combinations of working parameters is simulated by numerical simulation, generating a virtual dataset containing electromagnetic response simulation signals and their corresponding true values ​​of column burial depth. The virtual dataset is input into a pre-set inversion model for training. During the training process, physical prior expressions are introduced to constrain and optimize the model parameters. The trained inversion model is then used to output the burial depth detection results.

2. The method for checking the burial depth of crash barrier posts based on deep learning according to claim 1, characterized in that, The pre-set network structure of the inversion model includes: The first branch, which is a timing header, is used to receive and process electromagnetic response signals to extract their timing features. The second branch, which is the parameter header, is used to receive and process physical parameters that characterize the current detection condition in order to extract its conditional features. The fusion layer is used to fuse the temporal features with the conditional features, and to invert the estimated value of the column burial depth based on the fused features.

3. The method for checking the burial depth of crash barrier posts based on deep learning according to claim 2, characterized in that, The first branch is a lightweight Transformer structure, and the second branch performs feature extraction through a multilayer perceptron (MLP) to generate conditional embedding vectors.

4. The method for checking the burial depth of crash barrier posts based on deep learning according to claim 1, characterized in that, Also includes: A portion of the true burial depth data is obtained from field measurements. The measured data is then used to fine-tune the parameters of the trained inversion model to obtain a calibrated inversion model. Furthermore, the calibrated inversion model outputs the following burial depth detection results: The field detection signal of the column to be tested and the corresponding working parameters are input into the calibrated inversion model, and the burial depth detection result of the column to be tested is output.

5. The method for checking the burial depth of crash barrier posts based on deep learning according to claim 4, characterized in that, The step of generating a virtual dataset containing electromagnetic response simulation signals and their corresponding true values ​​of column burial depth based on the physical prior expression and simulating the detection process of the column under different combinations of working parameters through numerical simulation includes the following sub-steps: S21. Based on the aforementioned physical prior expression, define the simulation parameter space for the burial depth range, medium parameters, column parameters, and detection conditions; S22. Combine the parameters in the simulation parameter space to generate a three-dimensional model in batches. The three-dimensional model simulates the propagation of electromagnetic waves in the medium and their interaction with the column. The time domain and frequency domain response signals of the electromagnetic response signal are collected as part of the detection signal, and environmental interference is injected. S23. Extract multi-dimensional signal features related to burial depth from the simulation results, and label the true value of burial depth and the working condition parameter vector for each sample. S24. Organize the simulation samples into a dataset with paired annotations of the column burial depth corresponding to the detection signal, thereby obtaining a virtual dataset.

6. The method for checking the burial depth of crash barrier posts based on deep learning according to claim 4, characterized in that, The total loss function of the trained inversion model is: In the formula, As weight, For physical prior consistency constraints, This represents the regression error due to burial depth. in, This includes concrete embedment conditions, conventional soil embedment conditions, and backfill soil embedment conditions, and the parameter header or fusion layer of the inversion model integrates a hybrid expert module. The hybrid expert module includes: a gated network for calculating the weights of each expert network based on the input working condition parameter vector; and three expert networks, each corresponding to the physical prior expressions for the concrete embedding working condition, the conventional soil embedding working condition, and the backfill soil embedding working condition, respectively; the physical prior consistency constraint term. The result is obtained by weighted summation of the outputs of each expert network based on the weights calculated by the gating network.

7. A deep learning-based system for checking the burial depth of crash barrier posts, characterized in that, include: The memory stores a computer program for implementing the deep learning-based method for checking the burial depth of crash barrier posts as described in any of claims 1-6. The processor performs read and write operations on the memory to execute the deep learning-based method for checking the burial depth of the guardrail posts.

8. A device for checking the burial depth of crash barrier posts based on deep learning, characterized in that, include: Signal acquisition device, multi-frequency excitation transmitter probe, environmental physical parameter input box, signal line and deep learning model box; The deep learning model box is equipped with the deep learning-based anti-collision guardrail post burial depth inspection system as described in claim 7; The multi-frequency excitation transmitting probe emits electromagnetic waves toward the column and the surrounding medium. The signal acquisition instrument receives and collects reflected, transmitted, or induced signals to form time-domain and frequency-domain electromagnetic response signals. The environmental physical parameter input box receives the acquired signals and inputs the physical parameters of the current detection environment. The electromagnetic response signal and physical parameter vector are input into the deep learning model box through the signal line, and the burial depth detection result is output through the trained inversion model.

Images