A method and system for analyzing and processing roadbed subgrade coefficient detection data

Abstract

This invention discloses a method and system for analyzing and processing roadbed subgrade coefficient detection data. The method includes: constructing a stiffness-energy response matrix; performing frequency domain modal decomposition on the stiffness and energy sequences to generate a set of frequency domain modal features; constructing a load-bearing dominant mode enhancement mechanism to generate a weighted frequency domain feature tensor; performing physical constraint inversion based on the weighted frequency domain feature tensor, and predicting the subgrade coefficient by combining roadbed physical parameters; modeling the uncertainty of the prediction results, constructing an engineering confidence interval, and completing acceptance judgment. This invention improves the accuracy and reliability of subgrade coefficient assessment through joint stiffness-energy characterization, frequency domain modal enhancement, physical constraint inversion, and uncertainty quantification.

Description

Technical Field

[0001] This invention relates to the field of railway geotechnical engineering and foundation testing data processing, specifically a method and system for analyzing and processing roadbed foundation coefficient testing data. Background Technology

[0002] The subgrade of a high-speed railway is a crucial component of the track structure system, bearing the train load and transferring it to the foundation. Its load-bearing capacity directly affects the operational safety and long-term stability of the line. The subgrade foundation coefficient, as a key parameter characterizing the deformation characteristics and bearing capacity of the foundation, is one of the core control indicators in the design, construction, and acceptance of high-speed railway subgrades. In practical engineering, the subgrade foundation coefficient is typically obtained through on-site testing methods such as plate load tests, and its calculation results are widely used in the formulation of subgrade reinforcement schemes, settlement control, and quality assessment.

[0003] However, in high-speed railway engineering practice, the detection and evaluation of subgrade foundation coefficients face numerous complex influences. On the one hand, material conditions such as subgrade filler type, gradation composition, moisture content, and cementation state exhibit significant spatial dispersion and time-varying characteristics. On the other hand, factors such as loading levels, loading rates, and the accuracy of testing equipment during on-site testing can introduce varying degrees of disturbance, causing the load-settlement curve to exhibit obvious nonlinear and fluctuating characteristics. Against this backdrop, traditional subgrade coefficient calculation methods based on the slope of a finite loading interval often fail to fully reflect the true stiffness evolution and energy dissipation characteristics of the subgrade throughout the entire stress process.

[0004] To improve the engineering applicability of test results, ground parameter inversion methods based on data analysis and intelligent algorithms have gradually emerged in recent years. For example, regression models, machine learning, or deep learning are used to fit and predict load-settlement data. These methods have improved the ability to characterize nonlinear relationships to some extent, but most still use raw time-domain or spatial-domain data as the main input, making it difficult to effectively distinguish between the dominant bearing characteristics and noise disturbance components at different scales. At the same time, they rarely unify the modeling of subgrade material state parameters with test data, which can easily lead to unstable or highly biased prediction results under complex working conditions.

[0005] Therefore, in the detection and evaluation of subgrade coefficients for high-speed railways, there is still an urgent need for an analytical processing method that can simultaneously characterize the evolution of subgrade stiffness, energy dissipation characteristics, and the influence of material state, and can quantitatively characterize the uncertainty of detection, so as to improve the stability, reliability, and engineering judgment value of the subgrade coefficient evaluation results. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for analyzing and processing roadbed subgrade coefficient test data, which can achieve stable inversion of subgrade coefficient and output a reliable range under conditions of discrete plate load test data, noise disturbance, loading stage differences, and fluctuations in state such as moisture content and gradation, thereby improving the accuracy, reliability and engineering decision-making of roadbed test results.

[0007] Compared with the prior art, the present invention has the following advantages:

[0008] 1. This invention uses the dual physical quantities of stiffness sequence and strain energy sequence to jointly characterize the bearing capacity and damage evolution, thereby reducing the instability of manual segment fitting.

[0009] 2. This invention highlights the dominant mode by using frequency domain mode decomposition and adaptive mode enhancement mechanisms, thereby improving noise immunity and generalization ability under complex working conditions.

[0010] 3. This invention ensures that the prediction results conform to the basic laws of mechanics by using physical priors and joint loss constraints, thereby reducing the risk of non-physical abnormal outputs.

[0011] 4. This invention incorporates uncertainty into the acceptance judgment by outputting a confidence interval, providing a quantifiable basis for roadbed quality control. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the overall process of a method for analyzing and processing roadbed foundation coefficient detection data according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the stiffness-energy response construction process in step 1 of an embodiment of the present invention.

[0014] Figure 3 This is a schematic diagram of the modal enhancement processing flow in step 3 of an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0016] Please see Figure 1 This invention provides a method for analyzing and processing roadbed foundation coefficient detection data, comprising the following steps:

[0017] Step 1: Perform stiffness-energy response characterization processing on the raw data of roadbed subgrade coefficient detection, and construct a stiffness-energy response matrix. For example... Figure 2 As shown, step 1 includes:

[0018] 1.1 Collect raw data for roadbed subgrade coefficient testing. Collect the load values ​​corresponding to each sampling point or loading level in the plate load test. Settlement value Loading time The roadbed condition parameters, including moisture content, are collected. Cementation index Surface roughness and packing gradation parameters This forms the original dataset:

[0019]

[0020] in, Indicates the sampling point or load level index, with a value of ; For the first Each sampling time or loading completion time (in seconds or minutes); For the corresponding load (unit: kN or kPa); For corresponding settlement (unit: mm); Moisture content (%) The cementation index (dimensionless or graded); Contact roughness (dimensionless or graded); For gradation parameters (from) Uniformity coefficient curvature coefficient (and other components). The above dataset integrates load-settlement-time response and material state parameters into the same calculation framework, enabling subsequent models to simultaneously perceive external response and internal state, avoiding systematic bias caused by inferring subgrade coefficients solely from P-S curves.

[0021] 1.2 Consistency correction and outlier handling were performed on the original sequence. To address instrument vibration, occasional jumps, and localized setbacks observed in the field testing, median filtering was used for robust load correction, and outlier identification and replacement were performed for settlement.

[0022]

[0023]

[0024] in, The corrected load; It is a median function; The width of the window (typically 1–3). This is the threshold for abnormal settlement (generally taken as 10%-20%).

[0025] Through the above processing, the impulse noise, which is highly sensitive to differential and frequency domain transformations, is confined to a local window, preventing it from being amplified into pseudo-high-frequency modes, thereby ensuring that the subsequent stiffness sequence and spectral characteristics have real structural response significance.

[0026] 1.3 Calculate the instantaneous stiffness sequence and form a stiffness evolution characterization. Under the condition of load-settlement data discretization, the load increment and settlement increment are obtained by the difference between two adjacent points, and the instantaneous stiffness sequence is calculated accordingly.

[0027]

[0028]

[0029] in, For the first Segment load increment; For the first Segment settlement increment; For the first Instantaneous stiffness of segment (unit: kN / mm or kPa / mm); For stable terms (values) mm), used in Minimal values ​​should be minimized to prevent unbounded increase in stiffness. Instantaneous stiffness sequence. By making the entire process of roadbed evolution from initial compaction and stiffness growth to potential yielding and softening explicit, it is easier to extract the dominant bearing information from multiple scales in the future.

[0030] 1.4 Calculate the cumulative strain energy sequence and form an energy dissipation-damage characterization. Treating the load-settlement relationship as a work-displacement relationship, the cumulative strain energy during the loading process is obtained using trapezoidal integration:

[0031]

[0032] in, Cumulative strain energy (unit: kN·mm or kPa·mm); For cumulative indexing; For the first Average load per segment; This represents the displacement increment. Strain energy sequence. It reflects the energy absorption and dissipation of the roadbed during the loading process, and can complement the stiffness sequence to characterize the bearing capacity and irreversible deformation / damage evolution, thereby enhancing the ability to identify complex fill materials and nonlinear working conditions.

[0033] 1.5 Construct the stiffness-energy response matrix and unify its dimensions. Combine stiffness and energy into a two-channel response vector to form the stiffness-energy response matrix:

[0034]

[0035] To avoid the single-channel dominating feature extraction process due to different units, the two channels are standardized separately:

[0036]

[0037]

[0038] in, The mean and standard deviation of the stiffness sequence; The mean and standard deviation of the energy series; These are the dimensions after dimensionless transformation. By unifying the dimensions, stiffness information and energy information become comparable on a numerical scale, ensuring that subsequent modal contribution assessment and enhancement mechanisms can simultaneously reflect the bearing significance of both types of physical quantities.

[0039] Step 2: The stiffness sequence and energy sequence in the stiffness-energy response matrix are respectively subjected to frequency domain modal decomposition to extract load-bearing related multi-scale modal features and generate a frequency domain modal feature set.

[0040] 2.1 Resampling is performed using equally spaced sequences. When sampling time intervals are unequal or loading levels are uneven, [the following steps are taken]. Interpolation resampling to equal-interval sequences:

[0041]

[0042]

[0043] in, This is the resampling time step; For interpolation operators; This represents the number of resampling points. By using equal-interval processing, the frequency meaning and modal interpretation of the Discrete Fourier Transform are ensured to hold true, avoiding the introduction of pseudo-frequency components by non-uniform sampling.

[0044] 2.2 The stiffness spectrum is obtained by performing a discrete Fourier transform on the stiffness sequence. The stiffness spectrum separates the low-frequency compaction trend from the high-frequency local abrupt changes in stiffness evolution, enabling subsequent modal enhancement to specifically highlight the characteristics of the load-bearing skeleton.

[0045]

[0046] in, Frequency mode index; The imaginary unit; The value represents the complex value of the stiffness spectrum.

[0047] 2.3 The energy spectrum is obtained by performing a discrete Fourier transform on the energy sequence. The energy spectrum reflects the distribution of energy dissipation and damage at different scales, and when combined with the stiffness spectrum, it can distinguish between the "load-bearing dominant mode" and the "energy-dissipating dominant mode".

[0048]

[0049] in, For the complex value of the energy spectrum.

[0050] 2.4 Constructing the Amplitude-Phase Frequency Domain Tensor. Amplitude describes the modal energy magnitude, while phase describes response hysteresis and coupling characteristics; together, they constitute the frequency domain structural features that can be used for identification. The complex spectrum is decomposed into amplitude and phase, and a four-channel feature is constructed:

[0051]

[0052]

[0053]

[0054] in, For amplitude operators, This is the phase operator.

[0055] Step 3: Construct a dominant mode enhancement mechanism, assigning high weights to key modes in the frequency domain modal feature set and suppressing noise modes to generate a weighted frequency domain feature tensor. For example... Figure 3 As shown, step 3 includes:

[0056] 3.1 Calculate the modal contribution intensity.

[0057]

[0058] in, Contributes intensity to the mode; These are the channel weighting coefficients. The modal energy is represented by the square of the amplitude, which can quantify the contribution of each mode to the overall response and provide a unified scale for the selection of key modes.

[0059] 3.2 Adaptively determine the threshold of key modes.

[0060]

[0061] in, Percentile parameter (70%–90%) The threshold is determined by quantiles. This allows for adaptive adjustment of the energy distribution across different road sections and with varying filler conditions, preventing the fixed threshold from becoming ineffective.

[0062] 3.3 Construct the modal weight function.

[0063]

[0064] in, , As the dominant modal weights, Weights are assigned to non-dominant modes. Higher weights are given to high-contribution modes to highlight the supporting framework; while maintaining... To preserve the details of the material state and maintain spectral continuity.

[0065] 3.4 Generate a weighted frequency domain feature tensor.

[0066]

[0067] The weighted feature tensor is more sensitive to key modes in subsequent network training, thereby improving prediction accuracy and noise resistance.

[0068] Step 4: Perform physical constraint subgrade coefficient inversion based on the weighted frequency domain feature tensor, and construct a joint input tensor in combination with subgrade physical parameters, and predict the subgrade coefficient through a depth model.

[0069] 4.1 Constructing the joint input tensor of the subgrade frequency domain and state. The weighted frequency domain stiffness-energy modal characteristics obtained in step 3 are combined with the subgrade physical properties to form the input of the depth model:

[0070]

[0071] in, For the first Weighted stiffness-energy frequency domain characteristics of each frequency mode; The number of frequency modes; These are the standardized moisture content, cementation index, roughness, and gradation parameters. This input construction method projects the frequency domain load-bearing structure and material state conditions onto the same feature space, enabling the model to simultaneously identify the "stiffness skeleton" and "environmental degradation factors," thereby avoiding the ambiguity caused by relying solely on the load-settlement curve.

[0072] 4.2 Deep Fusion Subgrade Coefficient Inversion Based on Physical Constraints. Input Tensor The predicted values ​​of the subgrade coefficients are obtained through multi-layer nonlinear mapping.

[0073]

[0074]

[0075]

[0076] in, For network parameters; For activation functions; Number of floors; To predict the subgrade coefficient, a physical prior is constructed using the subgrade reaction-settlement relationship.

[0077]

[0078]

[0079] Wherein, 'a' is the nonlinear compaction coefficient of the subgrade, reflecting the stiffness increase effect caused by the gradual interlocking and densification of filler particles during loading; 'b' is the initial equivalent elastic stiffness coefficient, characterizing the linear elastic response of the subgrade during the small settlement stage. Both are obtained by least-squares fitting of measured load-settlement data, forming the equivalent nonlinear mechanical model of the subgrade.

[0080] 4.3 This physical prior embeds the overall mechanical behavior of the roadbed, from linear elasticity to nonlinear compaction, into the model constraint space, thus subjecting the predictions of the deep network to fundamental engineering mechanics laws. Based on this prior, at the standard settlement control points... The supervision objective is constructed, and the joint loss function is defined. The loss function parameters are updated using gradient descent.

[0081]

[0082]

[0083]

[0084]

[0085] Step 5: Perform uncertainty modeling on the predicted foundation coefficient results, construct the engineering confidence interval, and compare it with the preset design value to complete the acceptance judgment.

[0086] 5.1 Uncertainty Modeling of Subgrade Coefficient. Multiple random deactivation samples are performed on the trained deep model, and the prediction statistics are calculated:

[0087]

[0088] Where M is the number of samples;

[0089] 5.2 Calculate the mean of the forecast statistic and standard deviation :

[0090]

[0091] in, This represents the expected value of the subgrade coefficient. This represents the combined dispersion caused by detection noise, material dispersion, and model structural uncertainty.

[0092] 5.3 Construct the engineering confidence interval and complete the acceptance judgment. The engineering confidence interval for the subgrade coefficient is constructed based on the normal approximation.

[0093]

[0094] 5.4 Lower bound of the confidence interval With design value When comparing, If the foundation coefficient meets the design bearing requirements at the given confidence level, then the roadbed foundation coefficient is determined to meet the design bearing requirements; otherwise, it is determined that there is a risk of failure and retesting or treatment is required.

[0095] The technical solution of the present invention will be described below with a specific example:

[0096] 1. Construction of foundation stiffness-energy response data for high-speed railway subgrade

[0097] 1.1 Acquisition of raw test data for roadbed subgrade coefficient

[0098] 1.1.1 Engineering Testing Conditions

[0099] Taking a section of high-speed railway subgrade as an example, the subgrade fill material uses Group A fill material laid out on-site. This fill material is a mixture of excavated soil from the construction site and graded crushed stone from the quarry at a volume ratio of 70:30. Its main physical properties are: particle size distribution of 4–20 mm, moisture content of 8%–10%, and compaction degree controlled above 95%. The design subgrade coefficient is required to be no less than 80 MPa / m. According to the "Standard for Acceptance of Construction Quality of High-Speed ​​Railway Subgrade Engineering", a rigid circular bearing plate with a diameter of 300 mm was used for on-site plate load test in this section, with graded loading up to 400 kPa. The settlement was recorded after each stage of stabilization. Some measured data are as follows.

[0100] Table 1. Loading Measured Data (Partial)

[0101]

[0102] Simultaneously, the moisture content of the roadbed fill material in this section was recorded as w = 8.6%, the compaction degree as K = 97%, and the uniformity coefficient of the gradation curve as Cu = 8.1, which were used as subsequent state parameters.

[0103] 1.1.2 Original Data Expression. The moisture content of the subgrade fill material in this section, w = 8.6%, compaction degree K = 97%, and uniformity coefficient of the gradation curve Cu = 8.1, were also recorded as subsequent state parameters.

[0104]

[0105] in, This represents the load applied at level i. This indicates the corresponding cumulative settlement. This is for loading the steady-state time. This dataset includes both structural response and material state.

[0106] 1.2 Stability Correction for Load-Settlement Sequence

[0107] 1.2.1 Median Filtering of Load Sequence. To eliminate instantaneous fluctuations in the hydraulic loading system during the pressure stabilization phase, median filtering with a window width h=1 is applied to the load sequence. This processing maintains the graded monotonicity of the load curve, avoiding the introduction of pseudo-stiffness abrupt changes in subsequent difference calculations.

[0108]

[0109] 1.2.2 Correcting outliers in the settlement sequence. This process maintains the graded monotonicity of the load curve, avoiding the introduction of pseudo-stiffness abrupt changes in subsequent difference calculations.

[0110]

[0111] in, This threshold is used to distinguish between actual soil compression and instantaneous jumps caused by measurement noise. Its value can be determined based on the detection accuracy and the magnitude of subgrade deformation, and is typically taken as 20% of the current average settlement.

[0112] 1.3 Constructing the instantaneous stiffness sequence of roadbed foundation

[0113] 1.3.1 Calculation of Load and Settlement Increment

[0114] The increment of adjacent loading levels is calculated for the corrected sequence.

[0115]

[0116] For example, at level 3 loading:

[0117]

[0118] 1.3.2 Instantaneous Stiffness Calculation

[0119] Differential stiffness form is used:

[0120]

[0121] in A value of 10⁻³ mm is used to avoid stiffness divergence under minor settlement. Substitute the data from level 3.

[0122]

[0123] This stiffness reflects the actual compressive strength of the roadbed within the loading range.

[0124] 1.4 Construction of cumulative strain energy sequence for roadbed

[0125] Treating the load-settlement relationship as a work-displacement curve, and integrating it using the trapezoidal method:

[0126]

[0127] For example, at level 3:

[0128]

[0129] This quantity describes the deformation energy absorbed by the subgrade during loading and can reflect the compaction and damage accumulation state.

[0130] 1.5 Constructing the joint stiffness-energy response vector

[0131] 1.5.1 Two-Physical-Quantity Response Vector

[0132] Build a state vector at each loading level

[0133]

[0134] For example, Level 3

[0135]

[0136] This vector contains both "bearing stiffness" and "energy dissipation level", which can distinguish between the elastic compaction stage and the damage-dominated stage.

[0137] 1.5.2 Dimensional Unification and Standardization

[0138] Standardize stiffness and energy separately:

[0139]

[0140] And it forms a dimensionless characteristic:

[0141]

[0142] The standardized result serves as the direct input for the frequency domain mode decomposition in step 2, ensuring that the spectral analysis is not affected by differences in units and scales.

[0143] 2. Decomposition of the frequency domain modes of the stiffness-energy response of high-speed railway subgrade foundation

[0144] 2.1 Call the dimensionless stiffness sequence from step 1.5.2 The dimensionless energy sequence in step 1.5.2 The time series in step 1.1 That is, the input for step 2 can be represented as:

[0145]

[0146] in For the number of load levels, .

[0147] 2.2 Equal Interval Processing

[0148] 2.2.1 Calculate the time difference between adjacent times to determine equal intervals.

[0149]

[0150] When satisfied When this occurs, the sequence is considered to be an equally spaced sequence. In this example, the allowable time interval fluctuation threshold is... Meanwhile, the on-site recording meets the requirements. Since the value is constant at 120 s, no interpolation is needed; the value can be determined directly.

[0151] 2.2.2 If the on-site records are not at equal intervals, then construct an equally spaced time axis using the following formula:

[0152]

[0153] And an interpolation operator is used:

[0154]

[0155] The interpolation type is set to linear interpolation.

[0156] 2.3 Constructing the stiffness spectrum and energy spectrum.

[0157] 2.3.1 Setting the frequency domain transformation method and window function. For equally spaced sequences Perform Discrete Fourier Transform:

[0158] in, For frequency mode indexing, The imaginary unit, , It represents the complex spectrum.

[0159] To suppress spectral leakage caused by finite sequence truncation, a window function is applied to the input sequence. :

[0160] The Hanning window is chosen as the window function type because the number of loading stages in the plate load test is relatively small (N=5 in this example). The Hanning window can effectively reduce the pseudo-high frequency components caused by endpoint mutations when the sample size is limited, making the spectral structure more stable.

[0161] 2.4 Constructing the amplitude-phase mode tensor. The complex spectrum is decomposed into amplitude and phase:

[0162]

[0163] in, Represents the modulus of a complex number. The phase (in rad) is used to construct the four-channel feature vector for each mode, forming a frequency domain mode feature set.

[0164]

[0165]

[0166] Where N=5, Δt=120 s, and the window function is the Hanning window, the stiffness spectrum is calculated. With energy spectrum The corresponding modal amplitude and phase set The final output is a set of frequency domain modal features:

[0167]

[0168] in The four-channel modal feature vector will be used in step 3 to identify and enhance the dominant mode.

[0169] 3. Load-bearing dominant mode enhancement processing

[0170] After obtaining the frequency domain modal feature set in step 2, the modal contribution intensity is calculated according to the method described in step 3, and key modes are screened based on the quantile threshold and given high weights to obtain the weighted frequency domain feature tensor.

[0171] The process is the same as step 3 above, and will not be repeated here.

[0172] 4. Subgrade coefficient inversion based on weighted stiffness spectrum-energy mode and physical property parameters

[0173] 4.1 Frequency Domain – Material State Joint Modeling and Subgrade Coefficient Prediction

[0174] After completing the weighted frequency domain mode enhancement processing described in step 3, the weighted frequency domain mode set of the high-speed railway subgrade under multi-scale bearing conditions can be obtained:

[0175]

[0176] In this embodiment, the number of frequency modes is taken as Each modality vector is in four-dimensional form:

[0177]

[0178] Therefore, the dimension of the frequency domain carrying capacity feature is... .

[0179] Meanwhile, the specific values ​​of the relevant parameters can be obtained from the collection and standardization of roadbed filling material and construction status parameters in step 1, as shown in Table 2.

[0180] Table 2

[0181]

[0182] The weighted frequency domain modal features are concatenated with the aforementioned material state parameters to form the input vector of the deep inversion model:

[0183]

[0184] Its dimension is This vector comprehensively characterizes the multi-scale bearing structure of the high-speed railway subgrade under the current filling condition. Input vector Mapped via a three-layer fully connected neural network:

[0185]

[0186] After training, the model outputs the predicted equivalent subgrade coefficient value for the detected cross section.

[0187] 4.2 Load-settlement physical consistency verification and uncertainty analysis

[0188] The original data from the plate load test collected in step 1.1, within the settlement range:

[0189]

[0190] The corresponding load is:

[0191]

[0192] According to the standard definition of interval stiffness:

[0193]

[0194] This value is the interval equivalent stiffness result based on the load-settlement curve in the current railway and foundation testing specifications.

[0195] To characterize the uncertainty of the model's predictions, the trained deep model was subjected to 100 random deactivation samplings.

[0196]

[0197] The foundation coefficient statistic is obtained as follows

[0198]

[0199] At a 95% confidence level ( Constructing a trustworthy interval:

[0200]

[0201] This range also reflects the combined impact of detection noise, material dispersion, and model structural uncertainty on the evaluation results of the subgrade coefficient.

[0202] 5. Engineering Judgment and Comparison of Bearing Capacity of High-Speed ​​Railway Subgrade

[0203] The design specifications for high-speed railway subgrade stipulate that the subgrade coefficient should meet the following requirements. The confidence interval for the subgrade coefficient obtained by the method of this invention is: The interval covers the design control value. This indicates that, at a given confidence level, the load-bearing capacity of the tested section is within an evaluable range. This invention not only provides a single predicted value... It also provides a reliable interval reflecting the volatility of the project, offering more comprehensive information support for the quality evaluation and risk control of high-speed railway subgrade. At the same testing point, the results are obtained based on the standard interval method. The method of this invention yields...

[0204]

[0205] Although the results of the two methods are within the same order of magnitude, this invention further reveals the variation range of the subgrade coefficient under different frequency domain modes and material states, making the assessment of the bearing capacity of high-speed railway subgrade more robust, detailed and interpretable.

[0206] This invention also provides a roadbed subgrade coefficient detection data analysis and processing system for performing the above method, the system comprising:

[0207] The data acquisition module is used to collect raw data of the roadbed foundation coefficient test. The raw data includes the load value, settlement value, loading time, and roadbed state parameters corresponding to each sampling point or loading level in the plate load test.

[0208] The stiffness-energy response construction module is used to process the raw data and construct the stiffness-energy response matrix;

[0209] The frequency domain modal decomposition module is used to perform frequency domain modal decomposition on the stiffness sequence and energy sequence in the stiffness-energy response matrix, extract load-related multi-scale modal features, and generate a frequency domain modal feature set.

[0210] The modal enhancement module is used to construct a dominant modal enhancement mechanism, which assigns high weights to key modes in the frequency domain modal feature set and suppresses noise modes to generate a weighted frequency domain feature tensor.

[0211] The subgrade coefficient inversion module is used to perform physical constraint subgrade coefficient inversion based on the weighted frequency domain feature tensor, and to construct a joint input tensor by combining the roadbed physical property parameters collected by the data acquisition module, and to predict the subgrade coefficient through a depth model.

[0212] The uncertainty quantification module is used to perform uncertainty modeling on the foundation coefficient prediction results and construct the engineering confidence interval;

[0213] The output and judgment module is used to output the predicted foundation coefficient and the project confidence interval, and compare them with the preset design value to complete the acceptance judgment.

[0214] This invention achieves high-precision and high-reliability detection and analysis of roadbed subgrade coefficients through the coordinated processing of stiffness-energy response construction, frequency domain modal decomposition, load-bearing dominant mode enhancement, physical constraint inversion, and uncertainty quantification. Compared with existing technologies, this invention has the following characteristics and effects:

[0215] Joint characterization of stiffness and energy: By constructing a stiffness-energy response matrix, the bearing capacity and damage evolution characteristics of the roadbed are characterized simultaneously, avoiding the system bias caused by relying solely on the load-settlement curve, and providing structured input for subsequent frequency domain analysis.

[0216] Frequency domain modal decomposition and adaptive enhancement: Discrete Fourier transform is performed on the stiffness-energy response matrix to extract multi-scale frequency domain modal features. Key modes are adaptively screened by calculating modal contribution intensity and quantile threshold. High weights are assigned to the load-bearing dominant modes and noise modes are suppressed, significantly improving noise resistance and generalization ability under complex working conditions.

[0217] Physical constraint deep fusion inversion: The weighted frequency domain feature tensor is concatenated with the roadbed physical parameters to construct a joint input tensor. The subgrade coefficient is predicted by a deep neural network. At the same time, a roadbed reaction force-settlement physical prior model is introduced to construct a joint loss function to ensure that the prediction results conform to the basic mechanical laws and reduce the risk of non-physical anomaly output.

[0218] Uncertainty Quantification and Engineering Judgment: By obtaining the predicted distribution of subgrade coefficients through random deactivation sampling, an engineering confidence interval is constructed. The uncertainty of detection noise, material dispersion, and model structure is incorporated into the acceptance judgment, providing a quantifiable basis for subgrade quality control and improving the stability and engineering judgment of the evaluation results.

[0219] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for analyzing and processing roadbed subgrade coefficient detection data, characterized in that, Includes the following steps: (1) The original data of the roadbed foundation coefficient detection were subjected to stiffness-energy response characterization processing to construct a stiffness-energy response matrix; (2) Perform frequency domain modal decomposition on the stiffness sequence and energy sequence in the stiffness-energy response matrix respectively, extract the load-bearing related multi-scale modal features, and generate a frequency domain modal feature set; (3) Construct a dominant mode enhancement mechanism, assign high weights to key modes in the frequency domain mode feature set and suppress noise modes to generate a weighted frequency domain feature tensor; (4) Based on the weighted frequency domain feature tensor, perform physical constraint subgrade coefficient inversion, and construct a joint input tensor in combination with subgrade physical property parameters, and predict subgrade coefficient through a depth model; (5) Perform uncertainty modeling on the predicted foundation coefficient, construct the engineering confidence interval, and compare it with the preset design value to complete the acceptance judgment.

2. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 1, characterized in that, Step (1) specifically includes: (1.1) Collect raw data of roadbed subgrade coefficient detection, including the load value corresponding to each sampling point or loading level in the plate load test. Settlement value Loading time The data includes roadbed state parameters, such as moisture content w, cementation index c, surface roughness r, and filler gradation parameter g, forming the original dataset. ； in, Indicates the sampling point or load level index, with a value of ; For the first Each sampling time or loading completion time; For the corresponding load; To correspond to settlement; Moisture content; The cementation index; For contact roughness; For gradation parameters, by Uniformity coefficient curvature coefficient constitute; (1.2) The original sequence is corrected for consistency and outliers are handled. Median filtering is used to perform robust correction of the load and outlier identification and replacement of settlement. ； ； in, The corrected load; It is a median function; The width is half the width of the window; This is the threshold for settlement anomalies; (1.3) Calculate the load increment and settlement increment by difference between two adjacent points, and calculate the instantaneous stiffness sequence accordingly. This forms a characterization of stiffness evolution; (1.4) The cumulative strain energy sequence during the loading process is calculated using the trapezoidal integral method. This forms an energy consumption-damage characterization. (1.5) Combine the stiffness evolution characterization and the energy dissipation-damage characterization into a dual-channel response vector and construct the stiffness-energy response matrix. The two channels are standardized separately to obtain dimensionless features. .

3. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 2, characterized in that, The formula for calculating the instantaneous stiffness sequence in step (1.3) is as follows: ; in, For the first Segment load increment; For the first Segment settlement increment; For the first Instantaneous stiffness of the segment; It is a stable term; The formula for calculating the cumulative strain energy sequence in step (1.4) is as follows: ;in, For cumulative strain energy; For cumulative indexing; For the first Average load per segment; This represents the displacement increment.

4. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 2, characterized in that, Step (2) specifically includes: (2.1) Extract the dimensionless stiffness sequence from the stiffness-energy response matrix. and energy sequence And resampling is performed to obtain equally spaced stiffness sequences. and equally spaced energy sequences ; (2.2) Perform a discrete Fourier transform on the equally spaced stiffness sequence to obtain the complex values ​​of the stiffness spectrum. ; (2.3) Perform a discrete Fourier transform on the equally spaced energy sequence to obtain the complex numerical value of the energy spectrum. ; (2.4) Decompose the complex spectrum into amplitude and phase, and construct an amplitude-phase frequency domain tensor: ； in, ; , For amplitude operators, For phase operators; This leads to the formation of a frequency domain modal feature set.

5. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 4, characterized in that, The steps In (2.1), when the sampling time intervals are unequal, an equally spaced time axis is constructed using an interpolation method: ; ； in, This is the resampling time step; For interpolation operators; This represents the number of resampling points.

6. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 4, characterized in that, In steps (2.2) and (2.3), a Hanning window function is applied to the input sequence before performing the discrete Fourier transform to suppress spectral leakage. ; 。 7. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 4, characterized in that, Step (3) specifically includes: (3.1) Based on the frequency domain modal feature set, calculate the modal contribution intensity: ； in Contributes intensity to the mode; This refers to the channel weighting coefficient; (3.2) Adaptively determine the critical mode threshold using the quantile method: ； in, Percentile parameter (70%–90%) For the threshold; (3.3) Construct a modal weighting function W(m) and assign high weights to the dominant modes where D(m)>T. Low weights are assigned to non-dominant modes with D(m)≤T. ,and ; (3.4) Apply the modal weighting function to the frequency domain modal feature set to generate a weighted frequency domain feature tensor. .

8. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 7, characterized in that, Step (4) specifically includes: (4.1) The weighted frequency domain feature tensor By concatenating the standardized roadbed physical parameters, a frequency domain-state joint input tensor is constructed. ,in, For the first Weighted stiffness-energy frequency domain characteristics of each frequency mode; The number of frequency modes; These are the moisture content, cementation index, roughness, and gradation parameters after standardization treatment. (4.2) A deep neural network based on physical constraints performs multi-layer nonlinear mapping on the input tensor Z to obtain the predicted value of the foundation coefficient. At the same time, a physical prior model is constructed using the roadbed reaction force-settlement relationship. , where a is the nonlinear compaction coefficient of the subgrade and b is the initial equivalent elastic stiffness coefficient, which is obtained by least squares fitting of the measured load-settlement data; (4.3) Define the joint loss function ,in The mean square error between the predicted subgrade coefficient and the actual subgrade coefficient as defined in the standard is... The fitting error between the physical prior model and the measured load-settlement data is denoted by λ, which is the equilibrium coefficient. The network parameters are updated using the gradient descent method.

9. The method for analyzing and processing roadbed subgrade coefficient detection data according to claim 8, characterized in that, Step (5) specifically includes: (5.1) Perform multiple random deactivation samplings on the trained deep model to obtain the predicted distribution of the foundation coefficient. Where M is the number of samples; (5.2) Calculate the mean of the forecast statistic and standard deviation : ; ； in, This represents the expected value of the subgrade coefficient. This represents the combined dispersion caused by detection noise, material dispersion, and model structural uncertainty. (5.3) Constructing the engineering confidence interval based on the normal approximation ,in The quantiles corresponding to the confidence level; (5.4) Lower limit of the confidence interval With design value When comparing, If the subgrade coefficient meets the design bearing capacity requirements, it is determined that the subgrade coefficient meets the design bearing capacity requirements; otherwise, it is determined that there is a risk of failure.

10. A roadbed subgrade coefficient detection data analysis and processing system, characterized in that, include: The data acquisition module is used to collect raw data of the roadbed foundation coefficient test. The raw data includes the load value, settlement value, loading time, and roadbed state parameters corresponding to each sampling point or loading level in the plate load test. The stiffness-energy response construction module is used to process the raw data and construct the stiffness-energy response matrix; The frequency domain modal decomposition module is used to perform frequency domain modal decomposition on the stiffness sequence and energy sequence in the stiffness-energy response matrix, extract load-related multi-scale modal features, and generate a frequency domain modal feature set. The modal enhancement module is used to construct a dominant modal enhancement mechanism, which assigns high weights to key modes in the frequency domain modal feature set and suppresses noise modes to generate a weighted frequency domain feature tensor. The subgrade coefficient inversion module is used to perform physical constraint subgrade coefficient inversion based on the weighted frequency domain feature tensor, and to construct a joint input tensor by combining the roadbed physical property parameters collected by the data acquisition module, and to predict the subgrade coefficient through a depth model. The uncertainty quantification module is used to perform uncertainty modeling on the foundation coefficient prediction results and construct the engineering confidence interval; The output and judgment module is used to output the predicted foundation coefficient and the project confidence interval, and compare them with the preset design value to complete the acceptance judgment.

Images