Intelligent Prediction Method, Equipment, and Medium for Subgrade Defects Based on Time-Series Augmented Data

Abstract

This invention discloses a method, device, and medium for intelligent prediction of roadbed defects based on time-series augmented data, relating to the field of intelligent operation and maintenance of transportation infrastructure. The method first preprocesses the raw monitoring data to obtain a standardized multi-dimensional time-series dataset. Addressing the scarcity of defect samples, a time-series generative model is used to learn the temporal evolution patterns of defects and generate synthetic samples, constructing a class-balanced augmented dataset. Subsequently, a collaborative optimization framework is used to simultaneously search for the globally optimal feature subset and the globally optimal hyperparameter combination, and this is used to train a target prediction model. Finally, the target prediction model is used to analyze real-time data and output the defect prediction results. This invention effectively solves the problem of early identification and prediction of roadbed defects with small sample sizes through time-series data augmentation and feature-hyperparameter collaborative optimization, improving prediction accuracy and timeliness.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology of transportation infrastructure, and in particular relates to a method, equipment and medium for intelligent prediction of roadbed defects based on time-series enhanced data. Background Technology

[0002] The core of preventative maintenance of transportation infrastructure lies in the early and accurate warning of progressive defects such as frost heave and mudslides. Traditional detection methods, which rely on manual labor or specialized equipment, have inherent limitations in terms of timeliness, continuity, and economy, and cannot meet this need.

[0003] In recent years, continuous monitoring using sensors mounted on in-service vehicles has opened up new avenues for acquiring high-frequency, low-cost track condition data, theoretically laying the data foundation for real-time, dynamic defect prediction. However, from this new data source to achieving reliable early warning, there are two fundamental obstacles in the technological path—caused by the characteristics of the data itself and not adequately addressed by existing technologies—making early prediction difficult to achieve:

[0004] (1) Existing data augmentation techniques cannot handle the "dynamic evolution" characteristics of time-series data, making the small sample problem a dead end for prediction models. The extreme scarcity of disease samples is the primary challenge for model training. Although data augmentation is a common approach to solving the small sample problem, existing mainstream techniques (such as SMOTE (Synthetic Minority Oversampling) and its variants) are all based on a key assumption: data samples are static and independently identically distributed. They generate new samples by interpolating or perturbing the nearest neighbors in the sample feature space. However, roadbed condition monitoring data is a typical time series, and its value lies not only in the value of a single data point, but also in the state transition and evolution patterns shown by the data points in time sequence—this is the key to identifying the budding and development trends of diseases. Applying static augmentation techniques to time-series data may generate samples that are reasonable in the feature space, but fragmented in the time dimension, completely losing the continuity, causality, and consistency of physical laws necessary to characterize the disease development process. Training a model with such "distorted" augmented data is tantamount to teaching the model to learn incorrect physical processes, rendering the reliability and practicality of its predictions meaningless. Therefore, the lack of a method to generate "time-dynamically accurate" augmented samples for time-series data is the primary obstacle to achieving effective early predictions under small sample conditions.

[0005] (2) The existing feature optimization, hyperparameter tuning, and model training processes are fragmented, making it difficult to obtain the optimal configuration for the final prediction task. Feature selection is a key step in extracting effective information from high-dimensional noisy data; at the same time, the performance of deep learning models is highly dependent on hyperparameter settings. Although methods such as genetic algorithms and Bayesian optimization can be used for feature selection and hyperparameter optimization respectively, the existing process usually treats them and "model training" as multiple independent and sequential stages. This fragmented paradigm has an essential flaw: the feature subset or hyperparameter combination selected on the surrogate task cannot guarantee that it will also be optimal on the final target prediction model. More importantly, there are complex interactions between features, hyperparameters, and model structure. Isolated optimization cannot capture this synergistic effect and is prone to getting trapped in local optima, resulting in the prediction model's potential not being fully stimulated. Therefore, how to deeply coordinate and jointly optimize feature selection, hyperparameter optimization, and the training of the final prediction model is a key obstacle to improving prediction accuracy.

[0006] Some existing intelligent attempts (for example, the patent document with publication number CN117975236A uses radar data to image and then identifies it using a target detection model, or the patent document with publication number CN119669869A uses vibration data and performs multi-task learning through pseudo-label enhancement) have improved detection capabilities in specific scenarios, but their core technologies have not directly addressed the two obstacles mentioned above: the former did not utilize the temporal nature of the data, and the pseudo-label enhancement used by the latter did not solve the problem of maintaining the fidelity of temporal data generation. Moreover, their feature processing and model structure are still fixed pipeline-style and have not achieved collaborative optimization.

[0007] In summary, to achieve early prediction of roadbed defects based on continuous monitoring data, two major technical bottlenecks must be overcome: first, a time-series data augmentation method that can preserve the temporal evolution pattern must be created to fundamentally address the small sample size problem; second, a collaborative framework must be designed to simultaneously achieve joint optimization of feature selection, hyperparameter tuning, and prediction model training, thereby striking a balance between search efficiency and global optimality and fully tapping the potential of the data. This is precisely what is urgently needed in this field and is the technical solution that this invention aims to provide. Summary of the Invention

[0008] To address the aforementioned deficiencies in the existing technology, the present invention aims to provide a method, device, and medium for intelligent prediction of roadbed defects based on time-series enhanced data, in order to solve at least one of the following technical problems:

[0009] When traditional static data augmentation techniques are applied to roadbed time-series monitoring data, they disrupt the inherent time dependence and state evolution laws, resulting in invalid augmented samples and distorted model predictions.

[0010] In existing technology processes, feature selection, hyperparameter optimization, and prediction model training are disconnected, leading to inconsistent optimization objectives, a tendency to get stuck in local optima, and an inability to achieve global optimal synergy. This results in bottlenecks that limit the model's prediction accuracy and generalization ability.

[0011] This invention solves the above-mentioned technical problems through the following technical solution: a method for intelligent prediction of roadbed defects based on time-series enhanced data, comprising:

[0012] The raw monitoring data of the roadbed area are preprocessed to obtain a standardized multi-dimensional time-series dataset;

[0013] For disease categories with insufficient sample numbers in the multi-dimensional time-series dataset, a time-series generative model is used for augmentation. The time-series generative model learns the temporal evolution law of the original disease samples through a training process that includes a time-series dynamic learning stage, and generates synthetic disease samples that conform to the law, so as to construct a class-balanced augmented dataset.

[0014] Based on the augmented dataset, a globally optimal feature subset and a globally optimal hyperparameter combination are simultaneously searched using a collaborative optimization framework, and a target prediction model is trained based on the globally optimal feature subset and the globally optimal hyperparameter combination. The collaborative optimization framework achieves the simultaneous search through an iterative closed-loop process of joint optimization and evaluation feedback; wherein each iteration includes:

[0015] Based on the current optimization state, a candidate feature subset is proposed, and its corresponding optimal hyperparameter combination is determined to form a joint configuration.

[0016] A prediction model for evaluation is trained based on the joint configuration, and its performance evaluation results on the validation set are obtained.

[0017] The optimization state is updated based on the performance evaluation results for use in the next iteration;

[0018] The target prediction model is used to process the real-time acquired monitoring data and output the prediction results of roadbed area diseases.

[0019] To address the problem of existing static data augmentation techniques disrupting temporal dynamics, this invention introduces a temporal generative model. Through a dedicated "temporal dynamic learning stage," the generator is forced to learn the state transitions and evolutionary patterns of the original disease sequence over time. The resulting synthetic disease samples not only expand the sample size but, more importantly, maintain the temporal continuity, causality, and physical consistency necessary for representing disease development. This fundamentally solves the problem of sample scarcity preventing models from effectively learning early disease patterns, providing a high-quality, high-fidelity data foundation for the accurate training of subsequent prediction models.

[0020] To address the issue of isolated optimization of features, hyperparameters, and model performance easily getting trapped in local optima, this invention designs a collaborative optimization framework that achieves simultaneous optimization of all three through an iterative closed-loop process of "joint optimization and evaluation feedback." Within this framework, the performance evaluation of any candidate feature subset is performed under its corresponding optimal hyperparameter configuration (obtained by the inner optimizer), ensuring the fairness and guidance of the evaluation results. This deep collaborative optimization mechanism can automatically explore and lock in the optimal feature-hyperparameter joint configuration for the ultimate prediction task, fully tapping the potential of data and the model, thereby significantly improving the accuracy and robustness of the final prediction model.

[0021] Based on the innovations at the data and model layers, this invention enables early identification and risk warning of progressive defects such as frost heave and mud spillage in real-world operational scenarios, leveraging high-frequency, low-cost daily operational vehicle monitoring data. This method not only boasts high prediction accuracy but also provides a sufficient warning window, offering a crucial time buffer for maintenance departments to make scientific decisions and proactively intervene. This helps to fundamentally avoid track structure deterioration, increased operational safety risks, and high emergency repair costs caused by defect development, resulting in significant economic and social safety benefits.

[0022] Furthermore, using preset fixed-length distance segments as spatial analysis units, the raw monitoring data falling within each distance segment are processed to construct multi-dimensional time series samples corresponding to each distance segment, thereby forming the standardized multi-dimensional time series dataset.

[0023] By segmenting monitoring data into pre-defined fixed-length distance segments as spatial analysis units, the continuous, unstructured line space can be discretized and standardized, establishing a unified spatial reference system for the entire analysis process. This not only solves the spatial alignment problem caused by the discontinuity of the original monitoring data due to the running mileage, but also makes each distance segment an independent and comparable analysis object. This lays a crucial structured data foundation for subsequent feature extraction, time series construction, and accurate location and prediction of defects.

[0024] Further, the raw monitoring data falling within each of the aforementioned distance segments is processed, including, in sequence:

[0025] For each original monitoring data point falling within each distance segment, identify and correct any abnormal data points.

[0026] For each distance segment, the monitoring data is aligned and resampled in the time dimension to form a multi-dimensional time series sample with a unified time scale corresponding to the distance segment;

[0027] Based on the local extreme values ​​of key indicators in the monitoring data, combined with the known physical parameters and motion state parameters of the carrier, the dimensionless energy characteristics are calculated through a physical model, and the dimensionless energy characteristics are added to the multidimensional time series samples.

[0028] The monitoring data within each distance segment undergoes a three-step standardization process: anomaly correction, time warping, and physical feature generation. This systematically completes the entire feature engineering process, transforming raw sensor data into high-quality, comparable, and physically meaningful model input. Specifically, anomaly correction ensures data reliability, time warping eliminates timescale differences between different trains and time periods, and the dimensionless energy features generated based on the physical model fundamentally eliminate interference from individual vehicle attributes (such as vehicle weight and speed), extracting physical driving indicators that are only related to the intensity of track irregularity excitation. These three steps work together to transform the original high-dimensional, heterogeneous, and noisy monitoring data into a standardized feature sequence that can be directly input into the prediction model and has a clear physical interpretation. This is the core preprocessing step that ensures the model's prediction accuracy and generalization ability.

[0029] Furthermore, the calculation of dimensionless energy characteristics through a physical model specifically includes:

[0030] Based on the local maxima of suspension displacement identified in the monitoring data, and combined with the suspension spring stiffness, mass, and operating speed of the carrier, the dimensionless energy corresponding to each bogie side position of the carrier is calculated as the dimensionless energy characteristic according to the following model:

[0031] ;

[0032] Where Ed represents the dimensionless energy of a certain side of the bogie; SD i This represents the i-th local maximum of suspension displacement identified within a specific distance and time period; k represents the suspension spring stiffness coefficient of the carrier; M represents the mass of the carrier; V represents the average operating speed of the carrier during the current time period when it passes through the specific distance; and n represents the number of local maxima of suspension displacement.

[0033] By calculating the dimensionless energy characteristics using the aforementioned physical model, the dynamic response of the vehicle (carrier)-track interaction is refined into a pure track state excitation index decoupled from individual vehicle attributes (vehicle weight, vehicle speed). This index, by incorporating suspension spring stiffness, vehicle mass, and operating speed, standardizes the original suspension displacement response into dimensionless energy, effectively eliminating systematic biases in data collected from different vehicle models, loads, and operating speeds. This makes monitoring data from different vehicles and operating conditions directly comparable, successfully transforming the "carrier-specific" signals from vehicle sensors into physical characteristics that objectively and consistently reflect the cumulative excitation intensity of local track irregularities. This provides the model with a more sensitive and robust key input for the development of track defects.

[0034] Furthermore, the temporal generative model is a temporal generative adversarial network, whose network structure includes an encoder, a decoder, a generator, and a discriminator; the training process including the temporal dynamic learning stage includes, in sequence:

[0035] Representation learning phase: The encoder and decoder are trained with the goal of minimizing the difference between the input real sequence and the reconstructed sequence, and learn the compression and reconstruction mapping of time series data;

[0036] Temporal dynamic learning stage: The parameters of the encoder and decoder are fixed, and the generator is trained to minimize the error of its prediction of the next temporal state based on historical information, so that the generator learns the state transition dynamics and evolution law of the original disease sample in the time dimension.

[0037] Adversarial refining stage: Fix the parameters of the encoder and decoder, and jointly train the generator and discriminator to improve the authenticity and diversity of the generated sequences in the overall distribution.

[0038] By employing a temporal generative adversarial network (GAN) comprising an encoder, decoder, generator, and discriminator, and designing a three-stage training process of "representation learning - temporal dynamic learning - adversarial refinement," a data augmentation mechanism capable of learning and reproducing the intrinsic laws governing the temporal evolution of diseases was constructed. This mechanism first learns to compress and reconstruct real disease sequences to grasp their data distribution. Subsequently, under the condition of a fixed encoder and decoder, a dedicated temporal dynamic learning stage forces the generator to learn the dependencies and transitions between sequence states. Finally, adversarial training further enhances the overall realism of the synthesized sequences. The resulting synthetic disease samples not only fill the gap in the number of positive class samples but, more importantly, maintain a temporal dynamic pattern and physical evolution logic consistent with the development of real diseases. This provides high-quality, high-fidelity training data for subsequent prediction models, fundamentally solving the problem that the scarcity of samples prevents models from effectively learning early, progressive disease patterns.

[0039] Furthermore, the collaborative optimization framework adopts a nested collaborative optimization structure, wherein the outer first optimizer searches in the feature subset space and the inner second optimizer searches in the hyperparameter space.

[0040] In a single optimization iteration of the first optimizer, a complete closed-loop process is executed, specifically including:

[0041] The first optimizer generates or updates a subset of candidate features based on its current optimization state;

[0042] For the candidate feature subset, a corresponding data subset is extracted from the augmented dataset and used as input for model evaluation;

[0043] With the goal of optimizing the performance index of the prediction model used for evaluation on the data subset, the second optimizer is initiated to perform multiple rounds of search in the hyperparameter space to determine the locally optimal hyperparameter combination for the candidate feature subset, and the corresponding locally optimal performance index value is recorded; the locally optimal performance index value serves as the performance evaluation feedback of the candidate feature subset.

[0044] The first optimizer updates its optimization state based on the performance evaluation feedback of the candidate feature subset for subsequent iterations;

[0045] At the end of the entire collaborative optimization process, the output global optimal feature subset is the candidate feature subset with the best performance evaluation feedback in all iterations, and the output global optimal hyperparameter combination is the local optimal hyperparameter combination corresponding to the global optimal feature subset.

[0046] By employing a nested collaborative optimization structure, a deep joint and automatic optimization of feature selection and hyperparameter tuning is achieved, ensuring the global optimality of the final model configuration for the prediction task. Within this framework, each candidate feature subset proposed by the outer first optimizer triggers an independent, fine-tuned hyperparameter search for that subset by the inner second optimizer, with the optimal performance obtained from the search serving as the sole evaluation feedback for that candidate feature subset. This nested mechanism of "per feature subset - dedicated hyperparameter evaluation" fundamentally avoids evaluation bias caused by improper hyperparameter settings masking the true potential of feature subsets, forcing the optimization process to simultaneously explore the complex synergistic effects between features and hyperparameters. Ultimately, this method outputs not just a feature subset or a set of hyperparameters, but a globally validated optimal joint configuration with a high degree of feature-hyperparameter matching, thereby maximizing the performance potential of the prediction model and significantly improving the accuracy and robustness of disease identification.

[0047] Furthermore, the prediction method also includes a dynamic resource allocation step, specifically including:

[0048] While the optimization process of the first optimizer is in the first stage, a first number of second optimizer search rounds are allocated to each candidate feature subset.

[0049] After the optimization process enters the second stage, the candidate feature subsets are sorted according to the performance evaluation feedback, and only the candidate feature subsets ranked in the top P% are assigned a second number of second optimizer search rounds, which is greater than the first number.

[0050] By introducing a two-stage dynamic resource allocation strategy, efficient and intelligent allocation of computing resources is achieved while ensuring global search capabilities, significantly improving the overall efficiency of the collaborative optimization process. In the initial optimization phase, a low-cost, broad-coverage preliminary screening of all candidate feature subsets helps to quickly explore the solution space and avoid prematurely falling into local optima. After identifying the most promising subsets, the second stage focuses more computing resources on fine-tuning their hyperparameters. This adaptive strategy of "casting a wide net in the early stages and focusing on key targets later" frees up limited computing resources from a large number of mediocre candidates, allowing for in-depth exploration of the most promising solutions. This achieves an optimal balance between the breadth and depth of the search, and between exploration and utilization, enabling faster and more accurate convergence to the globally optimal joint configuration with the same or less computational cost.

[0051] Furthermore, the features included in the globally optimal feature subset are selected from the following feature groups:

[0052] Vehicle dynamic response characteristics include the vertical acceleration of the bogie, the standard deviation of the vertical acceleration, the suspension displacement, the suspension displacement difference, and the mean suspension displacement;

[0053] Track geometry features, including track curvature and track profile;

[0054] Vehicle operating characteristics, including vehicle speed, engine torque, and braking pressure;

[0055] And, dimensionless energy characteristics calculated based on physical models.

[0056] The globally optimal feature subset integrates multi-source heterogeneous information to construct a key feature system capable of comprehensively, multi-dimensionally, and multi-physically characterizing the interaction state of "vehicle-track-load". This system not only covers core observational information such as the vehicle's direct dynamic response, track geometry, and real-time operating status, but also transforms the suspension displacement response into a physical index that is decoupled from individual vehicle attributes and purely reflects the intensity of local track excitation by introducing dimensionless energy features derived from physical models. These features complement and corroborate each other, comprehensively capturing weak precursor signals and evolutionary patterns in the occurrence and development of frost heave from multiple perspectives, including dynamics, geometry, kinematics, and energy. This provides subsequent deep learning models with high-quality input that is information-rich, physically meaningful, and highly sensitive to early-stage defects, serving as the cornerstone for ensuring high-precision and interpretable predictions.

[0057] Based on the same concept, the present invention also provides an electronic device, including a memory, a processor, and a computer program or instructions stored in the memory, wherein the processor executes the computer program or instructions to implement the intelligent prediction method for roadbed defects based on time-series enhanced data as described above.

[0058] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implements the intelligent prediction method for roadbed defects based on time-series enhanced data as described above.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] This invention introduces a generative model with a time-series dynamic learning stage to generate synthetic samples that not only fill the quantitative gaps but also strictly maintain the temporal and physical logic of disease evolution. This provides a high-quality, high-fidelity training data foundation for the model and fundamentally solves the problem of small sample training under time-series data.

[0061] This invention designs a nested collaborative optimization framework that integrates feature selection, hyperparameter tuning, and model training. Through a closed-loop feedback mechanism, it synchronously searches for and locks the optimal feature-hyperparameter joint configuration for the final prediction task, achieving deep collaboration and global optimization in the model building process and fully releasing the model's performance potential.

[0062] The feature subset ultimately selected by this invention integrates multi-dimensional information such as vehicle dynamic response, track geometry, operating status, and dimensionless energy based on a physical model. It can comprehensively capture early signs of disease from multiple perspectives, including dynamics and energy, ensuring the accuracy and interpretability of the prediction results.

[0063] This invention can accurately identify high-risk road sections several days in advance based on low-cost daily monitoring data, providing sufficient early warning time and reliable decision-making basis for preventive maintenance, thereby effectively avoiding safety accidents and reducing the total life cycle maintenance cost. Attached Figure Description

[0064] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart of the intelligent prediction method for roadbed defects based on time-series enhanced data in an embodiment of the present invention;

[0066] Figure 2 This is an architecture diagram of the temporal generative adversarial network in an embodiment of the present invention;

[0067] Figure 3 This is a flowchart of the collaborative optimization iteration in an embodiment of the present invention. Detailed Implementation

[0068] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of 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.

[0069] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0070] Example 1

[0071] This invention aims to provide an efficient, accurate, and non-disruptive method for predicting roadbed frost heave (a type of roadbed defect). The core of this method lies in integrating a time-series data augmentation technique capable of simulating the defect's evolution process, thus addressing the bottleneck problem of predictive model failure caused by the scarcity of historical defect samples. Simultaneously, it achieves joint intelligent optimization of features and hyperparameters through a nested collaborative optimization framework of genetic algorithms (first optimizer) and Bayesian optimization (second optimizer). This method can be integrated with existing sensing systems in in-service vehicles to provide early warnings of roadbed frost heave defects, offering decision support for preventative maintenance.

[0072] Figure 1This is an overall flowchart of the intelligent prediction method for roadbed defects in an embodiment of the present invention, as shown below. Figure 1 As shown, the method specifically includes the following steps:

[0073] Step S1: Obtain the original monitoring data of the roadbed area.

[0074] The raw monitoring data includes vehicle dynamic response data, track geometry data, and vehicle operating status data. Vehicle dynamic response data includes the vertical acceleration of the bogie and the suspension displacement. Track geometry data includes track curvature and track profile. Vehicle operating status data includes vehicle speed, engine torque, and braking pressure.

[0075] The vertical acceleration of the bogies includes the vertical acceleration of the left side of the front bogie, the right side of the front bogie, the left side of the rear bogie, and the right side of the rear bogie. These vertical accelerations are collected by acceleration sensors installed at corresponding locations on the operating vehicle. The suspension displacements include the suspension displacements of the left side of the front bogie, the right side of the front bogie, the left side of the rear bogie, and the right side of the rear bogie. These suspension displacements are collected by laser displacement sensors installed at corresponding locations on the operating vehicle.

[0076] Based on the original acceleration and suspension displacement data, the standard deviation of vertical acceleration, the difference in suspension displacement, and the mean of suspension displacement for each bogie can be further calculated.

[0077] Track geometry data is collected through a high-precision measurement system, including track curvature, vertical profiles of the left and right rails, and high-precision GPS correction data that synchronizes spatial positioning and mileage.

[0078] The system collects real-time operating parameters such as vehicle speed, engine speed and torque, throttle opening, and brake pressure via a vehicle bus. It can also interface with a traffic operation management system to obtain traffic load and traffic flow data.

[0079] All data is recorded in 1-meter units and synchronized in time and space based on mileage and high-precision GPS.

[0080] Step S2: Preprocess the original monitoring data of the roadbed area to obtain a standardized multi-dimensional time series dataset.

[0081] This embodiment constructs a standardized time series using fixed distance segments as units. Specifically, 20 meters is used as a preset fixed-length spatial analysis unit, dividing the entire track into continuous independent analysis objects. All raw monitoring data falling within each spatial analysis unit are systematically processed to construct standardized, structured, multi-dimensional time series samples corresponding to each distance segment, ultimately forming a dataset usable by the model. The specific preprocessing process includes:

[0082] Step S2.1: Outlier identification and correction.

[0083] For a specific monitoring parameter (such as the vertical acceleration of the left side of the front bogie), let its entire set be D, and the total amount of data be N. First, divide D into K subsets {D1, D2, ..., D...} in 20-meter intervals. K}, each subset D i This corresponds to all readings of the monitored parameter within a specific distance range. For each subset D... i Perform the following operations:

[0084] Calculate the first quartile Q1, the third quartile Q3, and the interquartile range IQR = Q3 - Q1 of the data distribution; set the lower limit of the normal range as Q1 - 1.5 × IQR and the upper limit as Q3 + 1.5 × IQR.

[0085] Pair of subset D i All data points are evaluated: if a value exceeds the normal range mentioned above, it is considered an outlier.

[0086] Correction strategy: Statistical subset D i The percentage of outliers in the total number of outliers (i.e., outlier percentage) is used to determine whether the outlier percentage exceeds a preset threshold (e.g., 10%).

[0087] If it is (≤10%), then use subset D. i Replace these outliers one by one with the arithmetic mean of all normal data points;

[0088] If not (>10%), then subset D is considered to be... i Since the overall quality is unreliable, the principle of spatial proximity is instead used, utilizing the previous adjacent segment (D) of distance segment i. i-1 ) and the next adjacent segment (D) i+1 For normal data with the same parameter, calculate their respective robust central tendency values ​​(e.g., median Med). i-1 and Med i+1 Then, using the center mileage of these three distance segments as the independent variable and the corresponding central trend value as the dependent variable, linear interpolation is performed to calculate the estimated value Med at the center mileage of the current distance segment i. i Finally, the estimated value Med was used. i Replacement subset D i All data points (or directly as the daily representative value of this parameter for that distance segment).

[0089] The above-mentioned anomaly handling process is executed independently for all monitored parameters (including all acceleration, displacement, track geometry and operating status parameters) to ensure the reliability of input data.

[0090] Step S2.2: Time dimension regularization and sequence construction.

[0091] To form sequences with a uniform time scale, the following method is used:

[0092] Daily Feature Value Extraction: Using "day" as the basic time unit. For each 20-meter distance segment, for each monitoring parameter, extract the maximum value among the sampled values ​​of that monitoring parameter recorded by all trains passing through that distance segment on that day, as the daily representative value of that monitoring parameter for that distance segment on that day. This step aims to capture the most unfavorable or most significant conditions that the location may experience on that day.

[0093] Single-parameter time series formation: Arrange the daily representative values ​​of each monitoring parameter in chronological order according to their corresponding dates to form a single-parameter time series for that monitoring parameter within that distance range, denoted as . , where T is the sequence length (total number of days).

[0094] Missing Day Data Handling: If no trains pass through a certain distance segment on a certain day, the daily representative value cannot be directly obtained. A linear interpolation method based on temporal proximity is used to supplement it: using the daily representative values ​​of the same parameter for the three days prior to and following the day, the interpolation value for that day is calculated according to the following formula. :

[0095] (1)

[0096] Where t represents the day with missing data. and These are the daily representative values ​​for the first 3 days and the last 3 days, respectively.

[0097] Multi-dimensional time series sample synthesis: The single-parameter time series of all monitored parameters (e.g., multiple accelerations, displacements, track geometry, vehicle speeds, etc.) are spliced ​​and aligned according to the feature dimensions to form a multi-dimensional time series sample representing the evolution of the 20-meter distance segment over time. Each distance segment corresponds to one sample, and each sample is a data structure with dimensions (T, F), where T is the time length (number of days) and F is the total number of features (number of parameters).

[0098] Step S2.3: Calculate and supplement the physical characteristic – dimensionless energy.

[0099] To eliminate the system bias caused by differences in suspension stiffness, mass, and operating speed among different monitoring vehicles, and to extract a physical index that more fundamentally reflects the intensity of track irregularities, a dimensionless energy characteristic is introduced.

[0100] For each 20-meter distance segment and each designated monitoring day, perform the following calculations for the suspension displacement data of a specific side of the vehicle (such as the suspension displacement on the left side of the front bogie):

[0101] A sliding window method (e.g., window size 0.5 meters, step size 0.1 meters) was used to traverse all suspension displacement data for that side on the monitoring day, identifying all local maxima SD. i (That is, the value at this point is greater than all values ​​in its left and right adjacent windows); Calculate the local maximum SD for each point according to classical mechanics formulas. i Corresponding suspension spring potential energy :

[0102] (2)

[0103] In the formula, k is the suspension spring stiffness coefficient of the vehicle, which is provided by the train manufacturer. In this embodiment, it is taken as 50kN / m.

[0104] Introducing the mass M of the monitoring vehicle and its average speed V over the distance segment on that day, the dimensionless energy for that day and that lateral position is calculated using the following formula:

[0105] (3)

[0106] In the formula, n is the total number of local maxima of suspension displacement identified in that distance segment on that day.

[0107] The above calculations were repeated for the four positions: left side of the front bogie, right side of the front bogie, left side of the rear bogie, and right side of the rear bogie, respectively, resulting in four independent dimensionless energy features. These four new features were then added as new columns to the multidimensional time series samples corresponding to the distance segments and dates.

[0108] Through the systematic processing of the above three sub-steps, the original high-frequency, heterogeneous, and noisy monitoring data is transformed into a standardized multi-dimensional time series dataset with 20-meter distance segments as spatial units, days as time units, unified feature dimensions, and new features with clear physical meanings. This lays a high-quality data foundation for subsequent model training and prediction.

[0109] Step S3: Label the samples in the multi-dimensional time series dataset.

[0110] This step aims to build a labeled training dataset for subsequent supervised learning models. The labels are assigned strictly based on field-verified records of mudslide and frost defects provided by the historical maintenance department.

[0111] Positive sample (Label 1) labeling rules: If a 20-meter distance segment has been clearly confirmed to have experienced frost heave and mud spillage in the historical records (the exact location and date of the defect are known), then the distance segment is associated with the defect event.

[0112] Negative sample (label 0) labeling rules: If a 20-meter section has never been reported to have frost heave or mud leakage defects in a historical record of up to four years (or a sufficiently long observation period), and its various monitoring parameters have been within a stable fluctuation range for a long time, then the section is considered a healthy section and is labeled as a negative sample.

[0113] To transform defect events into fixed-length time-series samples usable by the model, a preset model input time window length T must be followed (consistent with the total length of the single-parameter time series for each distance segment constructed in step S2.2). For each distance segment marked as having a defect, the day before the defect occurred is used as the time reference point. Starting from the reference point, a continuous T-day data sequence is extracted in the historical direction (i.e., in reverse chronological order). This data is the standardized multi-dimensional time series containing all feature dimensions constructed for this distance segment in step S2.2. This extracted multi-dimensional sequence of length T constitutes a positive sample, which contains the complete state evolution process of the road segment within T days before the defect occurred.

[0114] Each 20-meter distance segment (whether a positive or negative sample) ultimately corresponds to an independent data sample. Each sample is a three-dimensional tensor (or can be viewed as a two-dimensional matrix) with dimensions (T, F), where: T is the time step, i.e. the fixed length (number of days) of the model input sequence; F is the total number of feature dimensions, which includes all preprocessed monitoring parameters and newly added features such as the calculated dimensionless energy.

[0115] The collection of all samples constitutes the initial labeled time-series dataset. Due to the scarcity of real-world defects, the number of positive samples in this dataset is extremely scarce, creating a severe imbalance with the number of negative samples. This is the core problem that subsequent steps need to address through data augmentation techniques.

[0116] The time window length T is a key hyperparameter, and its optimal value can be determined through subsequent collaborative optimization frameworks (e.g., experimentally verified to be 56 days). During the initial annotation construction, T can be set to a reasonable preset value (e.g., 60 days), and before the final model training, it can be uniformly adjusted and data resampled through optimization steps to match the optimal T value.

[0117] Step S4: For disease categories with insufficient sample size in the multi-dimensional time series dataset, use a time series generative model for enhancement.

[0118] To address the severe class imbalance caused by the scarcity of positive samples and to ensure that the generated synthetic sequences maintain a temporal regularity consistent with the actual disease evolution, this embodiment introduces and trains a temporal generative adversarial network (GAN) as a temporal generative model. The training and synthesis process of this model is specifically designed for a small number of real positive sample sequences (denoted as X). minor(e.g., 32 sequences).

[0119] like Figure 2 As shown, the constructed temporal generative adversarial network comprises four components: an encoder, a decoder, a generator, and a discriminator. All components are implemented using a single-layer bidirectional long short-term memory (Bi-LSTM) network with 128 hidden units. The latent vector dimension (i.e., latent space dimension) of the encoder output is set to 64. All components use the Adam optimizer for parameter updates.

[0120] The encoder is used to compress the input high-dimensional time series into a low-dimensional latent space; the decoder is used to reconstruct the time series from the latent space representation into the original feature space; the generator is used to autoregressively generate new latent state sequences in the latent space based on historical latent states and random noise; and the discriminator is used to distinguish whether the input sequence comes from the real data distribution or the generator.

[0121] All components (encoder, decoder, generator, and discriminator) have a hidden layer unit count of 128, a latent space dimension of 64, and are updated using the Adam optimizer.

[0122] In this embodiment, the training process of the temporal generative adversarial network is divided into three consecutive stages, which gradually guide the network to learn data distribution, temporal dynamics and overall realism.

[0123] Representation learning phase: The goal of this phase is to train the autoencoder (encoder and decoder) to learn how to represent the real disease sequence X. minor High-fidelity compression and reconstruction are performed. The real sequence X is input into the encoder to obtain a latent representation, which is then reconstructed into a sequence by the decoder. The parameters of the encoder and decoder are optimized by minimizing the reconstruction loss, which is defined as the expected L1 norm error between the input real sequence and the reconstructed sequence.

[0124] This stage enables the encoder and decoder to effectively perform latent space mapping and reconstruction of the sequence.

[0125] Temporal dynamic learning stage: This stage aims to enable the generator to learn the temporal evolution patterns of the original disease sequence. The parameters of the encoder and decoder trained in the representation learning stage are fixed. The operation flow is as follows:

[0126] The real sequence X is encoded into a latent sequence S=[s1,s2,...,s] using an encoder. T For each latent sequence S, the generator is trained autoregressively. Specifically, at each time step t (from 2 to T), the preceding true latent states are... (i.e., the true latent state up to step t-1) and random noise z t Input a generator; require the generator's output. Approaching the current true potential state s t (i.e., the true latent state at the current time step t). The generator parameters are optimized by minimizing the supervised loss, which is defined as the expected sum of the squared L2 norms of the prediction errors (differences between the current true latent state and the generator output) at each time step. This supervised loss forces the generator to learn the state transition dynamics of the true sequence in the latent space, thereby grasping the temporal causal laws of disease development.

[0127] Adversarial Refinement Stage: This stage aims to enhance the overall realism and diversity of the generated sequences, making them difficult to distinguish from real sequences in terms of data distribution. The parameters of the encoder and decoder are fixed, and the generator and discriminator are jointly trained.

[0128] The discriminator receives two types of inputs: one is the real sequence X; the other is the latent sequence generated by the generator based on autoregression of random noise and then decoded by the decoder.

[0129] The discriminator is trained to maximize its ability to distinguish between real sequences X and generated sequences, while the generator is trained to minimize the probability that its generated sequences will be detected by the discriminator. This process is achieved by optimizing the generative adversarial loss, which aims to balance the authenticity, dynamic consistency, and diversity of sequences. Specifically, the generative adversarial loss consists of two parts: the expected logarithm of the probability that the discriminator classifies a real sequence as "real," and the expected logarithm of the probability that the discriminator classifies a generated sequence as "false." During training, the discriminator aims to maximize this loss value, while the generator aims to minimize it, i.e., minimize the expected logarithm of the probability that the discriminator classifies its generated sequences as "false."

[0130] Using the trained generator, generate synthetic positive samples through the following steps:

[0131] Independently sample random noise z and an initial latent state (which can be sampled from a simple prior distribution such as the standard normal distribution); using a generator, starting with the random noise and the initial latent state, autoregressively generate a specified number (e.g., thousands) of synthetic latent sequences of length T. ; these synthetic latent sequences The pre-trained decoder is input and decoded back into the original feature space to obtain the feature sequence of the synthesized positive sample. Subsequently, samples were randomly selected from the massive negative samples. A considerable number of normal sequences. Finally, the resulting synthetic positive samples... By merging with randomly selected negative samples, a class-balanced augmented dataset is formed.

[0132] Augmented datasets not only solve the problem of imbalance in quantity, but more importantly, their positive class samples contain temporal evolution patterns that conform to physical laws and are learned from real disease data, providing a high-quality training foundation for subsequent prediction models.

[0133] Step S5: Based on the augmented dataset, simultaneously search for the globally optimal feature subset and the globally optimal hyperparameter combination through a collaborative optimization framework, and train the target prediction model based on the globally optimal feature subset and the globally optimal hyperparameter combination.

[0134] This step aims to simultaneously search for the globally optimal feature subset and globally optimal hyperparameter combination for constructing the prediction model using a nested collaborative optimization framework of genetic algorithms and Bayesian optimization, and obtain the target prediction model based on the globally optimal feature subset and globally optimal hyperparameter combination. Its specific implementation process includes:

[0135] Step S5.1: Framework initialization and parameter settings.

[0136] Outer optimizer settings: The genetic algorithm is used as the first outer optimizer to search in the discrete feature space composed of multiple initial features obtained in step S2; it uses binary encoding, the chromosome length is the number of initial features, and the gene bit is "1" to indicate that the corresponding feature is selected; the population size is randomly initialized to the number of initial features to ensure that the initial probability of the feature being selected is equal.

[0137] Genetic operators: roulette wheel selection is used, and an elite retention strategy is set (the top 10% of individuals in fitness are retained in each generation); the crossover operator is single-point crossover (probability 0.5); the mutation operator is uniform mutation (probability 0.2).

[0138] Termination condition: The maximum number of iterations is set to 30 generations.

[0139] Inner optimizer settings: The Bayesian optimization algorithm is used as the second optimizer to search in a continuous hyperparameter space. The search space specifically includes the number of LSTM layer units, Dropout rate, number of fully connected layer units, batch size, learning rate, class weights, etc.

[0140] Optimization objective: Minimize the false prediction rate Err of the evaluation prediction model on the validation set for mudslides and siltation. mud Both the evaluation prediction model and the target prediction model are LSTM models.

[0141] The core of the algorithm is to use Tree Parsons Estimation (TPE) as a surrogate model and the expected improvement function as the acquisition function.

[0142] Step S5.2: Collaborative optimization iterative process.

[0143] like Figure 3As shown, in a single iteration of the genetic algorithm, the following closed-loop process is performed for each individual in its population (i.e., a subset of candidate features):

[0144] Step S5.21: Genetic algorithm generates or updates a subset of candidate features.

[0145] Step S5.22: Extract the feature data corresponding to the candidate feature subset from the augmented dataset to form a data subset for evaluation.

[0146] Step S5.23: With the goal of optimizing the performance index of the prediction model used for evaluation on this subset of data, initiate an independent Bayesian optimization process; after multiple rounds of search, the Bayesian optimization algorithm outputs the optimal performance index value it finds (i.e., the lowest error prediction rate of mudslide and siltation, Err). mud The optimal performance index value is then used as the performance evaluation feedback for this candidate feature subset (which is regarded as the fitness of the individual in the genetic algorithm).

[0147] Step S5.24: The genetic algorithm receives performance evaluation feedback from all individuals in the current iteration.

[0148] Step S5.25: Determine if the genetic algorithm has reached the termination condition. If the termination condition has not been reached, the genetic algorithm updates its population state through selection, crossover, and mutation operations based on the performance evaluation feedback of all individuals, and proceeds to step S5.21 to start the next iteration; if the termination condition is reached (e.g., completing 30 iterations of evolution), the nested collaborative optimization framework outputs:

[0149] Globally optimal feature subset: The subset of candidate features that provides the best performance evaluation feedback across all iterations.

[0150] Global optimal hyperparameter combination: This refers to the optimal hyperparameter combination recorded in the inner optimization of the above global optimal feature subset (e.g., LSTM unit number of 256, Dropout rate of 0.496, etc.).

[0151] To improve optimization efficiency, a two-stage dynamic resource allocation strategy is adopted:

[0152] Phase 1 (Extensive Exploration): In the early stages of the genetic algorithm optimization process (e.g., the first ten generations), a small number of search rounds (e.g., 5 times) are allocated to the Bayesian optimization process corresponding to each candidate feature subset to conduct a fast and low-cost preliminary screening.

[0153] The second stage (deep utilization): When the optimization process enters the later stage, the system sorts all candidate feature subsets based on performance evaluation feedback, and only increases the number of Bayesian optimization process search rounds (e.g., to 20 times) for the top 10% elite individuals (corresponding to the elite retention strategy of the genetic algorithm) to achieve fine-tuning and realize adaptive focus of computing resources.

[0154] Step S5.3: Train the target prediction model based on the optimal configuration.

[0155] Based on the globally optimal feature subset output in step S5.24, extract all corresponding feature column data from the augmented dataset to form the final dataset used for model training and validation.

[0156] The final Long Short-Term Memory (LSTM) network model (i.e., the target prediction model) is constructed directly using the globally optimal hyperparameter combination output in step S5.24. Its input dimension is (T, F*), where T is the optimized time window length (e.g., 56 days), and F* is the dimension of the optimal feature subset (e.g., changed from the initial 30 to 17).

[0157] The prepared final dataset is divided into a training set and a validation set (e.g., 7:3). Using the training set, the Long Short-Term Memory (LSTM) network model is trained sufficiently (e.g., 50 epochs) with a determined optimal batch size, learning rate, etc., until the model's performance on the validation set converges and stabilizes. After training, a target prediction model that can be used for real-time prediction is obtained. This model takes the aforementioned optimal feature subset as input and can output the probability of mudslides occurring within a specific future time period (e.g., 14 days).

[0158] Step S6: Process the real-time acquired monitoring data using the target prediction model and output the disease prediction results for the roadbed area.

[0159] The real-time vehicle monitoring data is preprocessed in strict accordance with the process in step S2 (including anomaly correction, time warping and dimensionless energy calculation) to form a standardized time series with a length of 56 days (optimized historical time window).

[0160] The processed, standardized time series data is input into a pre-trained target prediction model, which outputs the probability of mudslides and frost in each 20-meter distance segment within the next 14 days (optimized prediction time intervals). A risk threshold of 0.5 is set, and distance segments with probabilities higher than this threshold are classified as "high-risk segments." The system automatically generates an early warning report containing the precise location (GPS coordinates) of the risk segment, the risk probability value, and the prediction time window, and promptly pushes it to the maintenance management department, providing direct evidence for preventative maintenance decisions.

[0161] Through the above specific implementation methods, this invention achieves the following: by utilizing generative enhancement that maintains temporal dynamics, it fundamentally solves the small sample dilemma in disease prediction; through nested collaborative optimization of genetic algorithm and Bayesian optimization, it automatically obtains the globally optimal feature and model configuration for the prediction task; and finally achieves high-precision, advanced (e.g., 14 days in advance) disease early warning based on low-cost operating vehicle data, with an error prediction rate as low as 8.81%, significantly improving the intelligence level and economic benefits of roadbed maintenance.

[0162] Example 2

[0163] This invention also provides an electronic device, which includes a memory, a processor, and a computer program or instructions stored in the memory. The processor executes the computer program or instructions to implement the intelligent prediction method for roadbed defects based on time-series enhanced data in this invention.

[0164] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from a storage portion into random access memory (RAM). The processor can be a multi-core processor or may contain multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit, graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for device operation are also stored in RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0165] The processor and memory described above are used together to execute programs / instructions stored in the memory. When the program / instructions are executed by the computer, they can implement the methods, steps, or functions described in the above embodiments.

[0166] Although not shown, embodiments of the present invention also provide a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implements the intelligent prediction method for roadbed defects based on time-series enhanced data in embodiments of the present invention.

[0167] Readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0168] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of roadbed defects based on time-series enhanced data, characterized in that, The prediction method includes: The raw monitoring data of the roadbed area are preprocessed to obtain a standardized multi-dimensional time-series dataset; For disease categories with insufficient sample numbers in the multi-dimensional time-series dataset, a time-series generative model is used for augmentation. The time-series generative model learns the temporal evolution law of the original disease samples through a training process that includes a time-series dynamic learning stage, and generates synthetic disease samples that conform to the law, so as to construct a class-balanced augmented dataset. Based on the augmented dataset, a globally optimal feature subset and a globally optimal hyperparameter combination are simultaneously searched using a collaborative optimization framework, and a target prediction model is trained based on the globally optimal feature subset and the globally optimal hyperparameter combination. The collaborative optimization framework achieves the simultaneous search through an iterative closed-loop process of joint optimization and evaluation feedback; wherein each iteration includes: Based on the current optimization state, a candidate feature subset is proposed, and its corresponding optimal hyperparameter combination is determined to form a joint configuration. A prediction model for evaluation is trained based on the joint configuration, and its performance evaluation results on the validation set are obtained. The optimization state is updated based on the performance evaluation results for use in the next iteration; The target prediction model is used to process the real-time acquired monitoring data and output the prediction results of roadbed area diseases. The temporal generative model is a temporal generative adversarial network, whose network structure includes an encoder, a decoder, a generator, and a discriminator; the training process, which includes a temporal dynamic learning phase, sequentially includes: Representation learning phase: The encoder and decoder are trained with the goal of minimizing the difference between the input real sequence and the reconstructed sequence, and learn the compression and reconstruction mapping of time series data; Temporal dynamic learning stage: The parameters of the encoder and decoder are fixed, and the generator is trained to minimize the error of its prediction of the next temporal state based on historical information, so that the generator learns the state transition dynamics and evolution law of the original disease sample in the time dimension. Adversarial refining stage: Fix the parameters of the encoder and decoder, and jointly train the generator and discriminator to improve the authenticity and diversity of the generated sequences in the overall distribution.

2. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 1, characterized in that, Using preset fixed-length distance segments as spatial analysis units, the raw monitoring data falling within each distance segment are processed to construct multi-dimensional time series samples corresponding to each distance segment, thereby forming the standardized multi-dimensional time series dataset.

3. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 2, characterized in that, The raw monitoring data falling within each of the aforementioned distance segments are processed in the following manner: For each original monitoring data point falling within each distance segment, identify and correct any abnormal data points. For each distance segment, the monitoring data is aligned and resampled in the time dimension to form a multi-dimensional time series sample with a unified time scale corresponding to the distance segment; Based on the local extreme values ​​of key indicators in the monitoring data, combined with the known physical parameters and motion state parameters of the carrier, the dimensionless energy characteristics are calculated through a physical model, and the dimensionless energy characteristics are added to the multidimensional time series samples.

4. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 3, characterized in that, The calculation of dimensionless energy characteristics through a physical model specifically includes: Based on the local maxima of suspension displacement identified in the monitoring data, and combined with the suspension spring stiffness, mass, and operating speed of the carrier, the dimensionless energy corresponding to each bogie side position of the carrier is calculated as the dimensionless energy characteristic according to the following model: ； Where Ed represents the dimensionless energy of a certain side of the bogie; SD i This represents the i-th local maximum of suspension displacement identified within the distance segment and time period; k represents the suspension spring stiffness of the carrier; M represents the mass of the carrier; V represents the average operating speed of the carrier during the current time period when it passes through the distance segment; n represents the number of local maxima of suspension displacement; the distance segment refers to a spatial analysis unit divided according to a preset fixed length.

5. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 1, characterized in that, The collaborative optimization framework adopts a nested collaborative optimization structure, in which the outer first optimizer searches in the feature subset space and the inner second optimizer searches in the hyperparameter space. In a single optimization iteration of the first optimizer, a complete closed-loop process is executed, specifically including: The first optimizer generates or updates a subset of candidate features based on its current optimization state; For the candidate feature subset, a corresponding data subset is extracted from the augmented dataset and used as input for model evaluation; With the goal of optimizing the performance index of the prediction model used for evaluation on the data subset, the second optimizer is initiated to perform multiple rounds of search in the hyperparameter space to determine the locally optimal hyperparameter combination for the candidate feature subset, and the corresponding locally optimal performance index value is recorded; the locally optimal performance index value serves as the performance evaluation feedback of the candidate feature subset. The first optimizer updates its optimization state based on the performance evaluation feedback of the candidate feature subset for subsequent iterations; At the end of the entire collaborative optimization process, the output global optimal feature subset is the candidate feature subset with the best performance evaluation feedback in all iterations, and the output global optimal hyperparameter combination is the local optimal hyperparameter combination corresponding to the global optimal feature subset.

6. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 5, characterized in that, The prediction method further includes a dynamic resource allocation step, specifically including: While the optimization process of the first optimizer is in the first stage, a first number of second optimizer search rounds are allocated to each candidate feature subset. After the optimization process enters the second stage, the candidate feature subsets are sorted according to the performance evaluation feedback, and only the candidate feature subsets ranked in the top P% are assigned a second number of second optimizer search rounds, which is greater than the first number.

7. The intelligent prediction method for roadbed defects based on time-series enhanced data according to claim 1, characterized in that, The features included in the globally optimal feature subset are selected from the following feature groups: Vehicle dynamic response characteristics include the vertical acceleration of the bogie, the standard deviation of the vertical acceleration, the suspension displacement, the suspension displacement difference, and the mean suspension displacement; Track geometry features, including track curvature and track profile; Vehicle operating characteristics, including vehicle speed, engine torque, and braking pressure; And, dimensionless energy characteristics calculated based on physical models.

8. An electronic device comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The processor executes the computer program or instructions to implement the intelligent prediction method for roadbed defects based on time-series enhanced data as described in any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by the processor, they implement the intelligent prediction method for roadbed defects based on time-series enhanced data as described in any one of claims 1 to 7.

Images