A Simulation-Measurement Fusion-Driven Method for Identifying Hidden Defects in Rail Transit

Abstract

This application belongs to the field of intelligent operation and maintenance technology for rail transit, specifically involving a simulation-measurement fusion-driven method for identifying hidden defects in rail transit. The method includes: firstly, establishing a dynamic model of the rail transit structure; simulating different defect states through parameter changes to generate a precisely labeled vibration simulation dataset as the source domain; and using this dataset to train a pre-trained model. Simultaneously, collecting measured vibration response data from the field and confirming its actual defect information, constructing a target domain. The source and target domains are input into a transfer learning framework, and through adversarial training between a feature extractor and a domain discriminator, domain-invariant features with consistent distributions between the source and target domains are optimized and generated. Based on these features, the pre-trained model is optimized and adapted to obtain the final identification model. This application achieves high-precision, low-cost transfer of mechanistic models to real-world scenarios, enabling accurate qualitative and quantitative integrated diagnosis of hidden defects such as fastener failure and track bed delamination.

Description

Technical Field

[0001] This application belongs to the field of intelligent operation and maintenance technology of rail transit, specifically involving a simulation-measurement fusion-driven method for identifying hidden defects in rail transit. Background Technology

[0002] Hidden defects in rail transit infrastructure (such as fastener failure and track bed delamination) are difficult to detect but seriously threaten operational safety. Therefore, accurate and efficient intelligent diagnostic technology is crucial.

[0003] Currently, deep learning-based identification methods are the mainstream approach to achieving this goal. However, methods that rely on large amounts of real-world labeled data for model training face serious challenges: acquiring real-world data covering various disease types and severity is costly and time-consuming, resulting in a severe shortage of training samples, which leads to insufficient model generalization ability and makes it difficult to apply to complex and ever-changing real-world operating environments.

[0004] To overcome the bottleneck of scarce experimental data, a technical approach based on numerical simulation to generate training data to drive deep learning models has emerged. This method synthesizes a large amount of labeled simulation data through physical mechanism models, effectively alleviating the sample shortage problem. However, this method has inherent drawbacks: due to model simplification and environmental idealization, there are significant differences in probability distribution between simulation data and real experimental data (i.e., domain differences). This leads to a significant decrease in the generalization ability of models trained on simulation data in real experimental scenarios, making it difficult to meet the diagnostic accuracy requirements of engineering applications.

[0005] Therefore, how to overcome the domain difference barrier between simulation and measured data and achieve efficient and accurate transfer of simulation data-driven pre-trained models to measured scenarios has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This application proposes a simulation-measurement fusion-driven scheme for identifying hidden defects in rail transit, aiming to solve the problems of model migration failure and unsatisfactory diagnostic accuracy caused by domain differences between simulation data and measured data in the prior art.

[0007] This application provides a simulation-measurement fusion-driven method for identifying hidden defects in rail transit, including: A dynamic model for simulating the vibration of rail transit structures was established. By changing the model parameters, different working conditions and hidden defects were simulated, and a simulation dataset of rail transit structure vibration with defect category labels and defect severity labels was generated as the source domain. A model is trained using the simulation dataset of the source domain to obtain a pre-trained model that outputs the disease category and severity. Collect measured vibration response data of the on-site rail transit structure, confirm the corresponding actual type and degree of damage, and construct a measured dataset as the target domain; The source domain and the target domain are input into a transfer learning framework, which performs adversarial training between a feature extractor and a domain discriminator to optimize the feature extractor and generate domain-invariant features with consistent distribution between the source domain and the target domain. Based on the domain-invariant features, the pre-trained model is optimized and adapted to obtain an identification model that can be applied to the measured data of the target domain. The identification model is used to process the measured data of vibration response of rail transit structures collected on site and output the identification results of the category and degree of hidden defects.

[0008] In some embodiments of this application, the dynamic model is a dynamic model of a vehicle-track-tunnel-soil coupled system.

[0009] In some embodiments of this application, the hidden defects include one or more of the following: fastener failure, track bed detachment, voids behind the lining wall, and voids inside the concrete.

[0010] In some embodiments of this application, the pre-trained model is a multi-task deep learning model; The multi-task deep learning model includes a shared feature extraction layer, a disease classification subnetwork, and a disease severity regression subnetwork. The shared feature extraction layer is used to extract general features from the input vibration data; the disease classification sub-network is used to output disease categories based on the general features; The disease severity regression subnetwork is used to output disease severity values ​​based on the general features.

[0011] In some embodiments of this application, the shared feature extraction layer is a convolutional neural network (CNN) or a long short-term memory network (LSTM).

[0012] In some embodiments of this application, when training the multi-task deep learning model, the model parameters are optimized by minimizing the weighted sum of the classification cross-entropy loss and the regression mean squared error loss.

[0013] In some embodiments of this application, when collecting the measured vibration response data, the monitoring equipment used is an accelerometer or a vibration fiber optic cable.

[0014] In some embodiments of this application, confirming the corresponding true disease category and severity includes: The actual disease type and severity corresponding to the measured data are confirmed by ground-penetrating radar, elastic waves, array ultrasonic waves, or manual inspection.

[0015] In some embodiments of this application, the transfer learning framework includes a feature extractor, a domain discriminator, and a task predictor, wherein: The feature extractor is used to perform feature mapping on the input source domain data and target domain data; The domain discriminator is used to determine the source domain of the extracted features and calculate the domain discriminant loss based on the determination result; The task predictor is used to predict the type and severity of diseases based on source domain features and to calculate the task prediction loss. The adversarial training aims to minimize the prediction loss of the predictor while maximizing the discrimination error of the domain discriminator, and jointly optimizes the feature extractor, domain discriminator, and task predictor. Through the optimization, the feature extractor is forced to learn domain-invariant features that are consistent with the distribution of the source and target domains.

[0016] In some embodiments of this application, the task predictor combines task classification and regression functions, and includes: A classification module whose network structure is the same as that of the disease classification subnetwork in the pre-trained model; A regression module whose network structure is the same as that of the disease severity regression subnetwork in the pre-trained model.

[0017] In summary, the simulation-measurement fusion-driven method for identifying hidden defects in rail transit proposed in the embodiments of this application firstly employs a transfer learning framework based on adversarial training to force the feature extractor to filter out specific interferences related to load randomness in both simulation and measured data. This extracts cross-domain consistent features (domain-invariant features) useful for defect identification. These features mathematically minimize the feature distribution differences between simulation and measured data, enabling the optimal task predictor trained on simulation data to maintain near-optimal performance on measured data. This allows the simulation-pre-trained model to be applied effectively in real-world scenarios. The high-precision diagnosis fundamentally solves the core problems of "model transfer failure and unsatisfactory diagnostic accuracy" pointed out in the background technology. Furthermore, this application transfers the qualitative and quantitative integrated diagnostic capabilities obtained by the multi-task model to the real-world domain with high fidelity by utilizing simulation data in the pre-training stage. This enables the final model to synchronously output the disease category and degree in real-world scenarios, providing more comprehensive diagnostic information. Finally, the above technical features, through the closed-loop synergy of "simulation data generation - multi-task pre-training - adversarial domain adaptation", jointly improve the accuracy, robustness, and reliability of the entire intelligent identification system. Attached Figure Description

[0018] The features and advantages of this application will become clearer with reference to the accompanying drawings, which are illustrative and should not be construed as limiting the application in any way. In the drawings: Figure 1 This is a technical framework diagram of a simulation-measurement fusion-driven method for identifying hidden defects in rail transit, based on some embodiments of this application. Figure 2 This is a schematic diagram illustrating the principle of a transfer learning framework based on a multi-domain adversarial adaptive network, according to some embodiments of this application; Figure 3 This is a flowchart illustrating a simulation-measurement fusion-driven method for identifying hidden defects in rail transit, based on some embodiments of this application. Detailed Implementation

[0019] In the following detailed description, numerous specific details of this application are illustrated by example to provide a thorough understanding of the relevant disclosure. However, it will be apparent to those skilled in the art that this application can be practiced without these details. It should be understood that the terms “system,” “apparatus,” “unit,” and / or “module” used in this application are one way of distinguishing different parts, elements, sections, or components at different levels in a sequential arrangement. However, these terms may be replaced with other expressions if other expressions can achieve the same purpose.

[0020] It should be understood that when a device, unit, or module is referred to as being "on," "connected to," or "coupled to" another device, unit, or module, it may be directly connected to or coupled to, or communicate with, other devices, units, or modules, or there may be intermediate devices, units, or modules present, unless the context explicitly indicates otherwise. For example, the term "and / or" as used herein includes any one and all combinations of one or more of the relevant listed items.

[0021] The terminology used in this application is for the purpose of describing specific embodiments only and is not intended to limit the scope of this application. As shown in the specification and claims of this application, unless the context clearly indicates otherwise, words such as "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate that explicitly identified features, integrals, steps, operations, elements, and / or components are included, and such expressions do not constitute an exclusive list, and other features, integrals, steps, operations, elements, and / or components may also be included.

[0022] Referring to the following description and accompanying drawings, these and other features and characteristics, operating methods, functions of related structural elements, combinations of parts, and economics of manufacture of this application can be better understood, wherein the description and drawings form part of the specification. However, it is clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. It is understood that the drawings are not drawn to scale.

[0023] Various structural diagrams are used in this application to illustrate various variations of the embodiments according to this application. It should be understood that the preceding or following structures are not intended to limit this application. The scope of protection of this application is determined by the claims.

[0024] As described in the background section, urban rail transit tunnels and track structures are prone to hidden defects such as fastener failure, track bed delamination, and voids behind the lining during long-term service, posing a serious threat to traffic safety. Currently, deep learning-based identification methods are the mainstream approach for intelligent diagnosis of these hidden defects. However, their development faces a core bottleneck: the scarcity of real defect samples makes data-driven model training difficult, while numerical simulation methods suffer from distribution differences (domain differences) between simulation and measured data due to model simplification, resulting in insufficient generalization ability of simulation models in real-world scenarios.

[0025] To address the aforementioned challenges, this application proposes a simulation-measurement fusion-driven method for identifying hidden defects in rail transit. The overall technical framework of the method is as follows: Figure 1 As shown, its core achieves a complete technical loop from physical mechanism-driven to cross-domain knowledge transfer through the synergy of four stages: high-fidelity dynamics simulation, multi-task deep learning pre-training, experimental data construction, and multi-domain adversarial adaptive transfer learning.

[0026] Specifically, the method first establishes a coupled dynamic model of vehicle-track-tunnel-soil, and simulates various working conditions and hidden defects by changing train operation and structural parameters, generating a vibration simulation dataset with accurate category and degree labels, providing a massive and complete sample source for model training; then, a multi-task deep learning network including a shared feature extraction layer and a specific task layer (classification and regression sub-networks) is constructed, and trained using the simulation dataset to form a pre-trained model that can synchronously output defect categories and degrees; subsequently, measured vibration response data are collected through on-site monitoring equipment, and the real defect labels are confirmed using ground penetrating radar and other means to construct a small-scale measured dataset.

[0027] Finally, a transfer learning framework is built based on a multi-domain adversarial adaptive network (its principle is as follows). Figure 2As shown, the simulation set is set as the source domain and the actual test set is set as the target domain. Through adversarial training between the feature extractor and the domain discriminator, the network is forced to generate domain-invariant features, thereby achieving high-precision and lossless transfer of the diagnostic knowledge of the pre-trained model to the actual test scenario. This completely solves the core problem of poor generalization ability of the simulation model in actual test applications.

[0028] Figure 3 This is a flowchart illustrating a simulation-measurement fusion-driven method for identifying hidden defects in rail transit, based on some embodiments of this application. Figure 3 As shown, the method specifically includes: S310. Establish a dynamic model for simulating the vibration of rail transit structures. By changing the model parameters, simulate different working conditions and hidden defects, and generate a rail transit structure vibration simulation dataset with defect category labels and defect severity labels as the source domain.

[0029] Specifically, using the established dynamic model of the rail transit structure, the entire process of a train running in a tunnel is simulated, and a vibration simulation dataset for the rail transit structure is constructed. During the simulation, various operating conditions are simulated by changing the model parameters: different train speeds, axle loads, and track irregularities are set as influencing factors; different types of hidden defects are also categorized within the model. These include fastener failure, track bed detachment, voids behind the lining, and internal voids in the concrete, with different severity levels set for each type of defect. .

[0030] Acceleration time history data of rails, track bed, or tunnel walls obtained from simulation calculations are collected as samples. and the corresponding {disease category} Disease severity } as a tag, construct containing Simulation dataset of group data .

[0031] S320, A model is trained using the simulation dataset of the source domain to obtain a pre-trained model that outputs the disease category and degree.

[0032] First, a diagnostic model for the categories and severity of hidden diseases is established. To simultaneously obtain qualitative and quantitative information about the diseases, a neural network model based on multi-task deep learning is built.

[0033] The network includes: Shared feature extraction layer: Utilizes convolutional neural networks (CNN) or long short-term memory networks (LSTM) to extract deep, general features from the input vibration data; Task-specific layer: It is divided into two parallel branches. One is a disease classification sub-network, which is used to predict disease categories; the other is a disease severity prediction sub-network, which is used to regress and predict disease severity values.

[0034] The simulation dataset was then constructed using S310. The model underwent supervised training. The loss function consisted of a weighted average of the classification cross-entropy loss and the regression mean squared error loss. Through training, the model learned the nonlinear mapping relationship between vibration signals and disease states, forming a pre-trained model. .

[0035] S330: Collect measured vibration response data of the on-site rail transit structure, confirm the corresponding actual fault type and degree, and construct the measured dataset as the target domain.

[0036] This step involves obtaining measured structural vibration data through on-site testing, which serves as the target domain. The corresponding hidden defects and their severity labels are then confirmed using methods such as ground-penetrating radar, elastic wave analysis, or manual inspection.

[0037] Specifically, acceleration sensors or vibration optical fibers are deployed on actual operating lines or test tracks to collect structural vibration responses when trains pass by. Meanwhile, in order to obtain accurate labels, high-precision detection methods are used for calibration.

[0038] Compare the measured vibration data with their corresponding actual disease categories and degree Perform matching to construct the actual test dataset. Due to the difficulty of obtaining data on-site, the size of the actual test dataset is limited. Typically much smaller than the size of the simulation dataset. .

[0039] S340, the source domain and the target domain are input into a transfer learning framework. This framework is trained adversarially with a feature extractor and a domain discriminator to optimize the feature extractor so as to generate domain-invariant features with consistent distribution between the source domain and the target domain.

[0040] S350, Based on the domain-invariant features, the pre-trained model is optimized and adapted to obtain an identification model that can be applied to the measured data of the target domain. The identification model is used to process the measured data of vibration response of rail transit structures collected on site and output the identification results of the category and degree of hidden defects.

[0041] To address the issue of poor performance when directly applying simulation models to measured data, this application introduces transfer learning to build a framework based on a multi-domain adversarial adaptive network, thereby constructing an intelligent identification model for hidden disease states driven by simulation-measurement fusion. Specifically: First, define the domain: simulation dataset As the source domain The dataset is large in volume and rich in labels, and will be used in the actual test dataset. As the target domain The amount of data is small.

[0042] Define the network structure: Including feature extractor Domain discriminator And the task classifier / regressor defined in s320 .

[0043] Next, we will conduct combat training: Source domain data and target domain data Input Feature Extractor The feature vectors are obtained respectively. and Domain discriminator The aim is to distinguish whether the input features come from the simulation domain or the measured domain. Feature Extractor This approach attempts to "deceive" the domain discriminator, preventing it from distinguishing the feature source and thus generating "domain-invariant features." Simultaneously, it optimizes the task classifier / regressor using labeled source domain data and partially labeled target domain data. The accuracy of the prediction.

[0044] Through the aforementioned adversarial training, alignment of the source and target domains in the feature space was achieved. At this point, the diagnostic logic learned from the simulation data (pre-trained model)... The knowledge acquired was successfully transferred and adapted to the distribution of measured data. In practical applications, simply inputting the newly acquired measured vibration data into the trained model will directly output accurate diagnostic results on the type and severity of hidden defects.

[0045] In summary, the simulation-measurement fusion-driven method for identifying hidden defects in rail transit proposed in the embodiments of this application firstly employs a transfer learning framework based on adversarial training to force the feature extractor to filter out specific interferences related to load randomness in both simulation and measured data. This extracts cross-domain consistent features (domain-invariant features) useful for defect identification. These features mathematically minimize the feature distribution differences between simulation and measured data, enabling the optimal task predictor trained on simulation data to maintain near-optimal performance on measured data. This allows the simulation-pre-trained model to be applied effectively in real-world scenarios. The high-precision diagnosis fundamentally solves the core problems of "model transfer failure and unsatisfactory diagnostic accuracy" pointed out in the background technology. Furthermore, this application transfers the qualitative and quantitative integrated diagnostic capabilities obtained by the multi-task model to the real-world domain with high fidelity by utilizing simulation data in the pre-training stage. This enables the final model to synchronously output the disease category and degree in real-world scenarios, providing more comprehensive diagnostic information. Finally, the above technical features, through the closed-loop synergy of "simulation data generation - multi-task pre-training - adversarial domain adaptation", jointly improve the accuracy, robustness, and reliability of the entire intelligent identification system.

[0046] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding descriptions in the foregoing device embodiments, and will not be repeated here.

[0047] Although the subject matter described herein is provided in the general context of execution on a computer system in conjunction with an operating system and applications, those skilled in the art will recognize that other implementations can also be executed in conjunction with other types of program modules. Generally, program modules include routines, programs, components, data structures, and other types of structures that perform specific tasks or implement specific abstract data types. Those skilled in the art will understand that the subject matter described herein can be practiced using other computer system configurations, including handheld devices, multiprocessor systems, microprocessor-based or programmable consumer electronics, minicomputers, mainframes, etc., and can also be used in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In a distributed computing environment, program modules may reside on both local and remote memory storage devices.

[0048] Those skilled in the art will recognize that the units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0049] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A simulation-measurement fusion-driven method for identifying hidden defects in rail transit, characterized in that, include: A dynamic model for simulating the vibration of rail transit structures was established. By changing the model parameters, different working conditions and hidden defects were simulated, and a simulation dataset of rail transit structure vibration with defect category labels and defect severity labels was generated as the source domain. A model is trained using the simulation dataset of the source domain to obtain a pre-trained model that outputs the disease category and severity. Collect measured vibration response data of the on-site rail transit structure, confirm the corresponding actual type and degree of damage, and construct a measured dataset as the target domain; The source domain and the target domain are input into a transfer learning framework, which performs adversarial training between a feature extractor and a domain discriminator to optimize the feature extractor to generate domain-invariant features with consistent distribution between the source and target domains. Based on the domain-invariant features, the pre-trained model is optimized and adapted to obtain an identification model that can be applied to the measured data of the target domain. The identification model is used to process the measured data of vibration response of rail transit structures collected on site and output the identification results of the category and degree of hidden defects.

2. The method according to claim 1, characterized in that: The dynamic model is a dynamic model of a vehicle-track-tunnel-soil coupled system.

3. The method according to claim 1, characterized in that: The hidden defects include one or more of the following: fastener failure, track bed detachment, voids behind the lining wall, and voids inside the concrete.

4. The method according to claim 1, characterized in that: The pre-trained model is a multi-task deep learning model; The multi-task deep learning model includes a shared feature extraction layer, a disease classification subnetwork, and a disease severity regression subnetwork. The shared feature extraction layer is used to extract general features from the input vibration data; The disease classification sub-network is used to output disease categories based on the general features; The disease severity regression subnetwork is used to output disease severity values ​​based on the general features.

5. The method according to claim 4, characterized in that: The shared feature extraction layer is a convolutional neural network (CNN) or a long short-term memory network (LSTM).

6. The method according to claim 4, characterized in that: When training the multi-task deep learning model, the model parameters are optimized by minimizing the weighted sum of the classification cross-entropy loss and the regression mean squared error loss.

7. The method according to claim 1, characterized in that: When collecting the measured data of the vibration response, the monitoring equipment used is an accelerometer or a vibration fiber optic cable.

8. The method according to claim 1, characterized in that, The confirmation of the corresponding actual disease category and severity includes: The actual disease type and severity corresponding to the measured data are confirmed by ground-penetrating radar, elastic waves, array ultrasonic waves, or manual inspection.

9. The method according to claim 1, characterized in that, The transfer learning framework includes a feature extractor, a domain discriminator, and a task predictor, wherein: The feature extractor is used to perform feature mapping on the input source domain data and target domain data; The domain discriminator is used to determine the source domain of the extracted features and calculate the domain discriminant loss based on the determination result; The task predictor is used to predict the type and severity of diseases based on source domain features and to calculate the task prediction loss. The adversarial training aims to minimize the prediction loss of the predictor while maximizing the discrimination error of the domain discriminator, and jointly optimizes the feature extractor, domain discriminator, and task predictor. Through the optimization, the feature extractor is forced to learn domain-invariant features that are consistent with the distribution of the source and target domains.

10. The method according to claim 9, characterized in that... : The task predictor combines task classification and regression functions, and includes: A classification module whose network structure is the same as that of the disease classification subnetwork in the pre-trained model; A regression module whose network structure is the same as that of the disease severity regression subnetwork in the pre-trained model.

Images