Rail transit structure vibration data feature enhancement method under random load interference

Abstract

This application belongs to the field of intelligent operation and maintenance technology for rail transit, specifically relating to a method for enhancing the features of structural vibration data under random load disturbances in rail transit. The method includes: acquiring vehicle-induced vibration acceleration data and its refined geometric attributes to construct a labeled training dataset; constructing and training a feature enhancement network composed of a data augmentation module, a Transformer encoder, a projection head, and a contrastive learning module; using network training to bring similar defect samples closer together and push away features of dissimilar defect samples; and using intra-class similarity and contour coefficients to evaluate and select the optimal Transformer encoder as the feature enhancement encoder. The encoder provided in this application can effectively suppress random load disturbances, enhance the separability of essential defect features in vibration data, and provide a high-quality data foundation for subsequent accurate and intelligent identification.

Description

Technical Field

[0001] This application belongs to the field of intelligent operation and maintenance technology for rail transit, specifically relating to a method for enhancing the characteristics of structural vibration data under random load disturbances in rail transit. Background Technology

[0002] Structural vibration monitoring is a key technology for ensuring the safe operation of rail transit. In engineering practice, methods such as signal denoising, time-frequency analysis, or conventional machine learning models are commonly used to try to extract effective structural fault characteristics from vehicle-induced vibration acceleration data.

[0003] However, due to the randomness of factors such as train axle load and speed, the vibration signal of rail transit exhibits strong non-stationary and time-varying characteristics. This fundamental problem makes the subtle characteristics of early defects easily submerged in strong noise, resulting in an extremely low effective signal-to-noise ratio.

[0004] When using data-driven methods for intelligent diagnosis, it was found that random load disturbances present a deeper technical obstacle in the feature space: the sample features of the same type of defect become diffused due to load changes, and the feature boundaries of different types of defects are blurred. Existing data-driven intelligent diagnosis technologies for rail transit are insufficient in their ability to enhance subtle defect features in vibration data, making it difficult to effectively address this distributional dilemma. Summary of the Invention

[0005] This application proposes a method for enhancing the features of structural vibration data under random load disturbances in rail transit, aiming to solve the problem that existing data-driven intelligent diagnostic schemes for rail transit have poor diagnostic effects due to insufficient ability to enhance subtle defect features in vibration data.

[0006] This application provides a method for enhancing the characteristics of structural vibration data under random load disturbances in rail transit, including: Obtain vehicle-induced vibration acceleration data of rail transit structures and the refined geometric attributes of corresponding defects at their locations; Based on the vehicle-induced vibration acceleration data and the refined geometric attributes of the disease, a training dataset is constructed, wherein each sample is labeled with the corresponding refined geometric attribute of the disease. A feature enhancement network is constructed and trained. The network includes a data augmentation module, a Transformer encoder, a projection head, and a contrastive learning module. The training of the network includes: data augmentation of sample data by the data augmentation module, feature extraction by the Transformer encoder, mapping to a low-dimensional space by the projection head, construction of positive and negative sample pairs by the contrastive learning module based on the refined geometric attribute labels of the disease, and optimization of the contrastive loss function to reduce the feature distance of positive sample pairs and increase the feature distance of negative sample pairs. Multiple trained Transformer encoders are evaluated using intra-class similarity and contour coefficients, and the optimal feature enhancement encoder is selected based on the evaluation results. The feature enhancement encoder is used to enhance the features of the vehicle-induced vibration acceleration data to improve the separability of the defect features.

[0007] In some embodiments of this application, obtaining vehicle-induced vibration acceleration data of the rail transit structure includes: The vehicle-induced vibration acceleration data are obtained by monitoring with accelerometers or fiber optic vibration sensors.

[0008] In some embodiments of this application, after obtaining the vehicle-induced vibration acceleration data, the method further includes: Outlier removal is performed on the vibration acceleration data, and noise reduction preprocessing is performed using wavelet thresholding or empirical mode decomposition methods.

[0009] In some embodiments of this application, the refined geometric attributes of the disease include structural state and geometric attributes; The structural states include normal state, fastener failure, track slab cracking, track slab detachment, and tunnel wall detachment. The geometric attributes include the length, width, height, and area of ​​the lesion.

[0010] In some embodiments of this application, the refined geometric attributes of the disease are represented by a three-level hierarchical labeling system, which consists of: structural state, geometric attributes, and attribute values.

[0011] In some embodiments of this application, constructing the training dataset includes: The constructed labeling system includes m structural states, n geometric attributes, and p attribute values ​​to cover all fine-grained operating conditions that require diagnosis. For each fine-grained working condition uniquely determined by structural state, geometric properties, and attribute values, q vehicle-induced vibration acceleration data are collected to provide sufficient samples for the model to learn the stability characteristics of this type of working condition. The sampling frequency and duration of the vehicle-induced vibration acceleration data must meet the requirement of fully preserving the key characteristics of the vehicle-induced dynamic response of the rail transit structure.

[0012] In some embodiments of this application, the data augmentation of the sample data includes: Add Gaussian noise to the vehicle-induced vibration acceleration data; and / or The amplitude of the vehicle-induced vibration acceleration data is scaled.

[0013] In some embodiments of this application, constructing positive and negative sample pairs based on the refined geometric attribute labels of the disease includes: Samples with the same refined geometric attribute labels for diseases are constructed into positive sample pairs; Sample features with different refined geometric attribute labels for diseases are constructed into negative sample pairs.

[0014] In some embodiments of this application, the step of reducing the feature distance of positive sample pairs and increasing the feature distance of negative sample pairs by optimizing the contrastive loss function includes: The parameters of the feature enhancement network are adjusted to reduce the distance between sample features with the same refined geometric attribute labels for diseases in the feature space, and to increase the distance between sample features with different refined geometric attribute labels for diseases in the feature space.

[0015] In some embodiments of this application, evaluating the trained multiple Transformer encoders using intra-class similarity and contour coefficients includes: The overall score (Score) for each candidate Transformer encoder is calculated using the following formula:

[0016] in, Within-class similarity, representing the mean cosine similarity among feature vectors of all samples under the same disease attribute label. The silhouette coefficient is used to comprehensively evaluate the density of samples within a cluster and the separation of samples between clusters in the feature space.

[0017] In summary, the structural vibration data feature enhancement method for rail transit under random load interference provided in the embodiments of this application constructs a three-level hierarchical labeling system of "structural state-geometric attributes-attribute values," providing fine-grained structured supervision signals for supervised comparative learning. This enables in-depth mining and utilization of prior knowledge of defects, laying the foundation for constructing a high-quality feature space. By adopting a supervised comparative learning framework and using "reducing the distance between positive sample pairs and increasing the distance between negative sample pairs" as the core optimization objective, the model is driven to focus on learning essential features sensitive to defect attributes, effectively suppressing noise interference caused by random loads and achieving significant enhancement of weak defect features at the feature level. By using intra-class similarity and silhouette coefficient as model evaluation indicators, quantitative evaluation of the feature enhancement effect and precise selection of the optimal encoder are achieved, ensuring the reliability and superiority of the obtained feature enhancement model. Ultimately, the synergistic effect of the above technical features jointly achieves stable enhancement and effective extraction of weak defect features under strong noise background, significantly improving the effective signal-to-noise ratio of vibration data and providing a reliable data foundation for accurate identification of rail transit structural states. 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 the overall technical flow of a method for enhancing the features of structural vibration data under random load disturbance of rail transit based on supervised contrastive learning, as shown in some embodiments of this application. Figure 2 This is a flowchart illustrating a method for enhancing structural vibration data features under random load disturbances in rail transit, according to some embodiments of this application. Figure 3 This is a network structure of a feature enhancement network shown in 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 slabs, under the coupled effects of long-term cyclic loads and complex environments, are prone to hidden defects such as fastener failure, track slab cracking / detachment, and tunnel wall detachment, seriously threatening operational safety. Intelligent diagnosis based on vehicle-induced vibration response data is an effective means of identifying these defects. However, the development of this technology faces a fundamental bottleneck: the randomness of rail transit loads due to factors such as train axle load and speed leads to vibration signals exhibiting strong non-stationary and time-varying characteristics. This makes it easy for the subtle features of early defects to be submerged in strong noise, resulting in an extremely low effective signal-to-noise ratio. Existing data-driven diagnostic methods lack the ability to enhance subtle defect features in vibration data, making it difficult to effectively address the distribution dilemma caused by random loads, which manifests in the feature space as "diffuse distribution of samples within a class and blurred feature boundaries between classes."

[0025] To address the aforementioned challenges, this application proposes a method for enhancing the features of structural vibration data under random load disturbances from rail transit based on supervised contrastive learning. The overall technical flow of the method is as follows: Figure 1 As shown, its core lies in constructing a refined attribute dataset of diseases—vibration acceleration—and using a feature enhancement network that integrates data augmentation, Transformer feature extraction, and supervised contrastive learning to bring the features of similar disease samples closer together and push away the features of dissimilar disease samples, thereby effectively suppressing random load interference, highlighting the essential attributes of diseases, and generating highly separable feature representations.

[0026] Specifically, the method first collects vehicle-induced vibration acceleration data using sensors such as accelerometers or vibration fiber optics, and combines this with on-site detection technology to obtain the refined geometric attributes (such as type and size) of the corresponding damage points, constructing a training dataset with fine-grained labels. Next, the core of the method is to construct a feature enhancement network. This network first adds noise and amplitude scaling to the input samples to simulate load randomness, then uses a Transformer encoder to extract deep features and maps them to a low-dimensional space through a projection head. Finally, positive and negative sample pairs are constructed based on the damage attribute labels. A contrastive learning module drives the network to continuously reduce the feature distance between sample pairs with the same attributes and increase the feature distance between sample pairs with different attributes by optimizing the contrastive loss function. After training, the Transformer encoder selected by intra-class similarity and contour coefficient evaluation is used as the final feature enhancement encoder. This encoder can generate high-quality features with intra-class aggregation and inter-class separation for new vibration data, laying the foundation for accurate subsequent state identification.

[0027] Figure 2 This is a flowchart illustrating a method for enhancing structural vibration data features under random load disturbances in rail transit, according to some embodiments of this application. Figure 2 As shown, the method specifically includes: S210: Obtain vehicle-induced vibration acceleration data of rail transit structures and the refined geometric attributes of corresponding defects at their locations.

[0028] Specifically, high-frequency accelerometers or distributed vibration fiber optic sensors are used to conduct long-term real-time monitoring of the operational rail transit structure. Outlier analysis is employed to remove outliers caused by sensor malfunctions or electromagnetic interference. Environmental white noise is removed using wavelet thresholding or empirical mode decomposition to obtain vehicle-induced vibration data of the rail transit structure. .

[0029] By combining manual on-site inspections with array ultrasonic detectors and impact echo meters, the actual structural condition of the monitoring points can be obtained. and the refined geometric attributes of diseases .

[0030] The structural states include: normal state, fastener failure, track slab cracking, track slab detachment, and tunnel wall detachment. The refined geometric attributes include quantitative indicators such as the length, width, height (depth), and area of ​​the defects.

[0031] S220, Based on the vehicle-induced vibration acceleration data and the refined geometric attributes of the disease, a training dataset is constructed, wherein each sample is labeled with the corresponding refined geometric attribute of the disease.

[0032] Based on the vehicle-induced vibration acceleration data obtained from S210 and the refined geometric attributes of the aforementioned defects, a refined defect attribute-vibration acceleration dataset is constructed, where each sample is labeled with its corresponding refined geometric attribute of the defect. Specifically: The tags adopt a three-level hierarchical structure of "structural state - geometric attribute - attribute value". The first level is the structural state (m types); the second level is the geometric attribute (n attributes of interest for each state); and the third level is the attribute value level (discretizing continuous attribute values ​​into p levels). The total number of tag categories is N = m * n * p, and the total data volume is m * n * p * q.

[0033] Each label category collects q sample data points. Each sample is a time series continuously collected for t seconds at a preset sampling frequency of k Hz, containing k*t data points. All samples are matched one-to-one with labels via timestamps and spatial locations.

[0034] S230, Construct and train a feature enhancement network. The network includes a data augmentation module, a Transformer encoder, a projection head, and a contrastive learning module. Training the network includes: augmenting the sample data through the data augmentation module; extracting features through the Transformer encoder; mapping the features to a low-dimensional space through the projection head; constructing positive and negative sample pairs based on the refined geometric attribute labels of the disease through the contrastive learning module; and reducing the feature distance of positive sample pairs and increasing the feature distance of negative sample pairs by optimizing the contrastive loss function.

[0035] Specifically, a feature enhancement network based on supervised contrastive learning and Transformer is established. Figure 3 This refers to the network structure of the feature enhancement network. For example... Figure 3 As shown, the feature enhancement network includes an enhancement module, a Transformer encoder, a projection head, and a contrastive learning module, wherein: The data augmentation module is used to augment the sample data to simulate random load fluctuations, including adding Gaussian noise to the vehicle-induced vibration acceleration data and / or scaling the amplitude of the vehicle-induced vibration acceleration data.

[0036] The Transformer encoder is used to extract features from the augmented vehicle-induced vibration acceleration data. The projection head is used to map the features to a low-dimensional space; The contrastive learning module is used to construct positive and negative sample pairs based on the refined geometric attribute labels of diseases, and to reduce the feature distance of positive sample pairs and increase the feature distance of negative sample pairs by optimizing the contrastive loss function. Constructing positive and negative samples involves building positive sample pairs from sample features with the same refined geometric attribute labels of diseases, and building negative sample pairs from sample features with different refined geometric attribute labels of diseases.

[0037] The training process of the feature enhancement network is as follows: Data augmentation: To simulate signal fluctuations under different random loads, the input samples are augmented. An enhancement transformation is performed. An enhanced view is generated by adding Gaussian noise (to simulate changes in the ambient noise floor) and amplitude scaling (to simulate energy changes caused by different axle loads or vehicle speeds).

[0038] After data augmentation, the samples were divided into positive pairs. and negative sample pairs Positive sample pairs contain vibration acceleration data with the same disease attribute, while negative sample pairs contain vibration acceleration data with different disease attributes. Through training and learning, the positive samples are continuously improved. Increase the distance between negative sample pairs. The distance between them makes the characteristics of vibration acceleration data with the same defect attributes tend to be consistent, while the characteristics of vibration acceleration data with different defect attributes tend to be different, thereby eliminating the differences caused by random loads of rail transit and amplifying the differences caused by defect attributes.

[0039] After training, the projection head is removed, and the optimized Transformer encoder is retained as the feature augmentation encoder D. The final features obtained are:

[0040] S240, the trained Transformer encoders are evaluated using intra-class similarity and contour coefficient, and the optimal feature enhancement encoder is selected based on the evaluation results. The feature enhancement encoder is used to enhance the features of the vehicle-induced vibration acceleration data to improve the separability of the defect features.

[0041] To ensure the encoder can effectively separate different disease features, a clustering evaluation metric is used to assess the feature space distribution: intra-class similarity. Calculate the mean cosine similarity of sample feature vectors under the same disease attribute label. The larger the value, the more clustered the features of the same disease are under different loads, and the more thoroughly random load interference is eliminated; silhouette coefficient The density within a cluster and the separation between clusters are evaluated comprehensively. The closer this value is to 1, the clearer the characteristic boundaries of different disease attributes are.

[0042] A grid search method was used to adjust hyperparameters such as the number of Transformer layers, the number of attention heads, and the temperature parameter in contrastive learning. A comprehensive evaluation score was defined.

[0043] Calculate the comprehensive score under different parameter combinations, and select the model with the highest score as the optimal feature enhancement encoder.

[0044] Therefore, the optimal feature enhancement encoder obtained by the above method can be used to enhance the features of the vehicle-induced vibration acceleration data, thereby effectively suppressing the interference of random loads in rail transit, enhancing the weak essential features related to defects in the vibration data, and providing a high-quality data foundation for subsequent accurate and reliable intelligent identification of rail transit structural conditions.

[0045] In summary, the structural vibration data feature enhancement method for rail transit under random load interference provided in the embodiments of this application constructs a three-level hierarchical labeling system of "structural state-geometric attributes-attribute values," providing fine-grained structured supervision signals for supervised comparative learning. This enables in-depth mining and utilization of prior knowledge of defects, laying the foundation for constructing a high-quality feature space. By adopting a supervised comparative learning framework and using "reducing the distance between positive sample pairs and increasing the distance between negative sample pairs" as the core optimization objective, the model is driven to focus on learning essential features sensitive to defect attributes, effectively suppressing noise interference caused by random loads and achieving significant enhancement of weak defect features at the feature level. By using intra-class similarity and silhouette coefficient as model evaluation indicators, quantitative evaluation of the feature enhancement effect and precise selection of the optimal encoder are achieved, ensuring the reliability and superiority of the obtained feature enhancement model. Ultimately, the synergistic effect of the above technical features jointly achieves stable enhancement and effective extraction of weak defect features under strong noise background, significantly improving the effective signal-to-noise ratio of vibration data and providing a reliable data foundation for accurate identification of rail transit structural states.

[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 method for enhancing the characteristics of structural vibration data under random load disturbances from rail transit, characterized in that, include: Obtain vehicle-induced vibration acceleration data of rail transit structures and the refined geometric attributes of corresponding defects at their locations; Based on the vehicle-induced vibration acceleration data and the refined geometric attributes of the disease, a training dataset is constructed, wherein each sample is labeled with the corresponding refined geometric attribute of the disease. A feature enhancement network is constructed and trained. The network includes a data augmentation module, a Transformer encoder, a projection head, and a contrastive learning module. The training of the network includes: data augmentation of sample data by the data augmentation module, feature extraction by the Transformer encoder, mapping to a low-dimensional space by the projection head, construction of positive and negative sample pairs by the contrastive learning module based on the refined geometric attribute labels of the disease, and optimization of the contrastive loss function to reduce the feature distance of positive sample pairs and increase the feature distance of negative sample pairs. Multiple trained Transformer encoders are evaluated using intra-class similarity and contour coefficients, and the optimal feature enhancement encoder is selected based on the evaluation results. The feature enhancement encoder is used to enhance the features of the vehicle-induced vibration acceleration data to improve the separability of the defect features.

2. The method according to claim 1, characterized in that, The acquisition of vehicle-induced vibration acceleration data for rail transit structures includes: The vehicle-induced vibration acceleration data are obtained by monitoring with accelerometers or fiber optic vibration sensors.

3. The method according to claim 2, characterized in that, After obtaining the vehicle-induced vibration acceleration data, the process further includes: Outlier removal is performed on the vibration acceleration data, and noise reduction preprocessing is performed using wavelet thresholding or empirical mode decomposition methods.

4. The method according to claim 1, characterized in that: The detailed geometric attributes of the disease include structural state and geometric attributes; The structural states include normal state, fastener failure, track slab cracking, track slab detachment, and tunnel wall detachment. The geometric attributes include the length, width, height, and area of ​​the lesion.

5. The method according to claim 4, characterized in that: The detailed geometric attributes of the disease are represented by a three-level hierarchical labeling system, which consists of: structural state, geometric attributes, and attribute values.

6. The method according to claim 1, characterized in that, The construction of the training dataset includes: The constructed labeling system includes m structural states, n geometric attributes, and p attribute values ​​to cover all fine-grained operating conditions that require diagnosis. For each fine-grained working condition uniquely determined by structural state, geometric properties, and attribute values, q vehicle-induced vibration acceleration data are collected to provide sufficient samples for the model to learn the stability characteristics of this type of working condition. The sampling frequency and duration of the vehicle-induced vibration acceleration data must meet the requirement of fully preserving the key characteristics of the vehicle-induced dynamic response of the rail transit structure.

7. The method according to claim 1, characterized in that, The data augmentation of the sample data includes: Add Gaussian noise to the vehicle-induced vibration acceleration data; and / or The amplitude of the vehicle-induced vibration acceleration data is scaled.

8. The method according to claim 1, characterized in that, The construction of positive and negative sample pairs based on the refined geometric attribute labels of diseases includes: Samples with the same refined geometric attribute labels for diseases are constructed into positive sample pairs; Sample features with different refined geometric attribute labels for diseases are constructed into negative sample pairs.

9. The method according to claim 1, characterized in that, The method of optimizing the contrastive loss function to reduce the feature distance of positive sample pairs and increase the feature distance of negative sample pairs includes: The parameters of the feature enhancement network are adjusted to reduce the distance between sample features with the same refined geometric attribute labels for diseases in the feature space, and to increase the distance between sample features with different refined geometric attribute labels for diseases in the feature space.

10. The method according to claim 1, characterized in that, The evaluation of the trained Transformer encoders using intra-class similarity and contour coefficients includes: The overall score (Score) for each candidate Transformer encoder is calculated using the following formula: in, Within-class similarity, representing the mean cosine similarity among feature vectors of all samples under the same disease attribute label. The silhouette coefficient is used to comprehensively evaluate the density of samples within a cluster and the separation of samples between clusters in the feature space.

Images