Sleeper internal damage degree identification method based on vibration signals

Abstract

A sleeper internal damage degree identification method based on a vibration signal relates to the technical field of sleeper damage identification, and comprises the following steps: arranging a vibration sensor array on a sleeper, a ballast bed and a related structure, and collecting the vibration signal; carrying out noise reduction processing on the collected vibration signals; extracting a time domain feature and a frequency domain feature from the denoised signal; inputting the extracted features into a trained classification identification model, identifying the degree of the internal crack of the sleeper, and outputting an identification result; according to the recognition result, sleeper health state evaluation and early warning are carried out; the method is used for solving the problems that traditional early crack recognition is difficult and low in recognition precision.

Description

Technical Field

[0001] This invention relates to the field of railway sleeper damage identification technology, specifically to a method for identifying the degree of internal damage to railway sleepers based on vibration signals. Background Technology

[0002] As the core load-bearing structure of urban rail transit systems, subway ballastless track's sleepers are crucial components for transmitting train loads and maintaining track geometry stability. Sleepers are typically made of prestressed concrete and partially embedded in the reinforced concrete track bed, forming a complex, multi-layered coupled system together with the rails, fasteners, and substructure. During long-term operation, sleepers are continuously subjected to cyclic impact loads from trains, braking stress, and the combined effects of environmental temperature and humidity changes, groundwater erosion, and other factors, making them highly susceptible to developing initial micro-cracks. These cracks often originate in stress concentration areas within the sleeper, such as around the reinforcing bars, at abrupt changes in cross-section, or at areas with casting defects. If not identified and addressed in time, these micro-cracks will gradually expand and connect under continuous dynamic loads, eventually forming through cracks. This leads to a reduction in the sleeper's effective load-bearing cross-section, stiffness degradation, and in severe cases, concrete spalling, accelerated steel corrosion, and even sleeper fracture, directly threatening train operation safety.

[0003] Currently, the detection of defects in subway sleepers mainly relies on periodic manual inspections and track inspection vehicles. Manual inspections, using experience-based methods such as visual inspection and tapping, can detect obvious surface defects, but are almost powerless to detect microscopic damage embedded in the track bed or in its early stages. Furthermore, manual inspections are inefficient and highly subjective. Track inspection vehicles collect track geometry data through onboard sensors, primarily for evaluating track smoothness. However, their indirect ability to reflect the internal structural condition of sleepers is limited, making it difficult to achieve precise location and quantitative assessment of damage.

[0004] In recent years, structural health monitoring technology based on vibration signal analysis has provided a new approach for identifying internal damage in railway sleepers. The basic principle of this technology is that structural damage alters its stiffness, mass, or damping characteristics, thereby affecting its dynamic response (vibration signal). Theoretically, by analyzing changes in vibration signal characteristics, the state of the structure can be inverted. However, applying this theory to subway sleeper damage identification, especially the identification of early micro-damage, faces significant challenges: First, signal acquisition and signal-to-noise ratio (SNR) are significant issues. The vibration response of a railway sleeper is a typical "source-path-sensor" coupled signal. Besides the weak "damage characteristic signal" caused by the damage itself, the vibration source primarily consists of strong background vibration generated by the interaction between the train wheels and rails, as well as environmental noise. In enclosed spaces such as tunnels, these vibration components superimpose and reflect each other, making the damage signal easily submerged, resulting in an extremely low SNR.

[0005] Secondly, damage feature extraction and quantification are difficult. The core technology lies in how to extract sensitive features related only to sleeper damage from the complex vibration response of the system. Single time-domain or frequency-domain indicators are often related to multiple factors and lack specificity. The signal changes caused by early micro-damage are extremely weak and may overlap with noise in the frequency band, making feature extraction and accurate quantification of damage severity (such as distinguishing between micro-cracks and through-cracks) a challenge.

[0006] Finally, the construction and generalization capabilities of intelligent recognition models need improvement. Even with the extraction of multi-dimensional features, establishing robust mapping relationships between these features and different damage types and severity requires efficient machine learning models. Some existing methods may rely on simple threshold judgments or shallow networks, which lack sufficient generalization ability and recognition accuracy when dealing with different track conditions, different train models, and different damage morphologies.

[0007] Therefore, we provide a method that can accurately distinguish different degrees of damage inside railway sleepers. Summary of the Invention

[0008] The purpose of this invention is to provide a method for identifying the degree of internal damage of railway sleepers based on vibration signals, which solves the problems of difficulty and low accuracy in identifying early cracks in traditional methods.

[0009] This invention is achieved through the following technical solution: A method for identifying the degree of internal damage to railway sleepers based on vibration signals, specifically including: Vibration sensor arrays are deployed on sleepers, track bed and related structures to collect vibration signals; The collected vibration signals are subjected to noise reduction processing; Extract time-domain and frequency-domain features from the denoised signal; The extracted features are input into the trained classification and recognition model to identify the degree of internal cracks in the sleepers and output the recognition results. Based on the identification results, the health status of railway sleepers is assessed and an early warning is issued.

[0010] Furthermore, the vibration sensor is a low-frequency piezoelectric vibration sensor with a frequency range of 0.2Hz to 1500Hz, and its placement locations include the rail web, rail bottom, sleeper surface, track bed surface, and tunnel wall.

[0011] Furthermore, the noise reduction process employs a fully adaptive noise ensemble empirical mode decomposition method, including the following steps: Inject adaptive Gaussian white noise into the original signal; Empirical mode decomposition is performed on the injected noise signal to extract multiple intrinsic mode function components; By performing multiple iterations and ensemble averaging, the influence of noise is eliminated and the signal is reconstructed.

[0012] Furthermore, the time-domain features include dimensional features and dimensionless features, wherein the dimensional features include mean, variance, root mean square value, and peak value; and the dimensionless features include peak factor, kurtosis coefficient, waveform factor, impulse factor, and margin factor.

[0013] Furthermore, the frequency domain features include centroid frequency, frequency variance, and mean square frequency.

[0014] Furthermore, the classification and recognition model is an extreme learning machine model based on the particle swarm optimization algorithm, and its training process includes: Randomly initialize the input weights and hidden layer biases; Optimize hidden layer node parameters using particle swarm optimization algorithm; An extreme learning machine model is constructed based on the optimized parameters, and the output weights are solved analytically using the generalized inverse matrix.

[0015] Furthermore, the particle swarm optimization algorithm uses the prediction error as the fitness function and iteratively updates the particle positions to find the global optimal solution.

[0016] Furthermore, the identification results are classified according to the severity of the crack, including no crack, microcrack, or through crack.

[0017] Furthermore, the assessment and early warning of sleeper health status based on the identification results specifically includes: Based on the identification results and historical feature data of the target sleeper in multiple consecutive monitoring cycles, the feature evolution gradient vector reflecting the change in its damage severity is calculated. Using a pre-trained damage dynamics model, based on the feature evolution gradient vector, the trend curve of the severity level of the damage changing with time or the number of load cycles under no intervention conditions is predicted. Based on the trend curve, calculate the inherent evolution probability of the damage's severity level jumping within the next preset long-term maintenance planning window; Calculate the probability of dynamic intervention risk using actual operation and maintenance constraint parameters; Set inherent risk thresholds and dynamic risk thresholds. When the probability of dynamic intervention risk is greater than or equal to the dynamic risk threshold, trigger an immediate action warning. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is greater than or equal to the inherent risk threshold, a planning warning is triggered. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is less than the inherent risk threshold, no warning is triggered. The final output is a structured report containing the warning level, damage parameters, recommended measures, and recommended time windows.

[0018] Furthermore, the specific steps for calculating the probability of dynamic intervention risk using actual operation and maintenance constraint parameters are as follows: Determine the nearest available repair times and repair durations for sunroofs; Assess the conditional probability that, given the current trend of damage evolution, the damage may accelerate and spread before the sunroof repair time, potentially affecting driving safety. By combining the estimated operation time required to complete the repair of the damage point with the matching degree of the maintenance window duration, the probability of dynamic intervention risk is calculated.

[0019] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention discloses a method for identifying the degree of internal damage of railway sleepers based on vibration signals. By deploying a sensor array to synchronously collect vibration signals and performing adaptive noise reduction, the signal-to-noise ratio of the damage signals is effectively improved. Then, by systematically extracting multi-dimensional time-domain and frequency-domain features and constructing a classification-based recognition model, the vibration signal features can be accurately mapped to different damage levels, significantly improving the ability to detect and identify early internal cracks.

[0020] In addition, by calculating the feature evolution gradient, predicting the inherent evolution probability, and combining the actual track maintenance time, operation duration and other constraints to calculate the dynamic intervention risk probability, a graded early warning and structured decision suggestion report is finally generated. This closed-loop design provides quantifiable and operable intelligent decision support for the health management of sleepers.

[0021] In addition, by assessing and issuing early warnings for health status, the abstract identification results are transformed into specific operational instructions such as "take immediate action" and "planning warnings," and associated with suggested time windows. This greatly enhances the practical value of early warning information, enables the optimized allocation and precise scheduling of limited maintenance resources, and technically ensures a balance between subway operation safety and maintenance efficiency. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a method for identifying the degree of internal damage of railway sleepers based on vibration signals according to the present invention. Figure 2 This is a schematic diagram of the network structure of the ELM of the present invention; Figure 3 This is a schematic diagram of the method for assessing and issuing early warnings for the health status of railway sleepers based on the identification results, according to the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Example 1 like Figure 1 The method for identifying the degree of internal damage of railway sleepers based on vibration signals, as shown, specifically includes: Vibration sensor arrays are deployed on sleepers, track bed and related structures to collect vibration signals; the setup of this vibration sensor array is a precisely designed three-dimensional, multi-source synchronous monitoring network based on the damage mechanism of subway sleepers and the vibration transmission path. In addition, the vibration sensor is a low-frequency piezoelectric vibration sensor with a frequency range of 0.2Hz to 1500Hz, and its placement locations include the rail web, rail bottom, sleeper surface, ballast surface, and tunnel wall. Among them, the acoustic emission or structural modal change signals generated by cracks inside the sleepers, especially early microcracks, under the excitation of train loads, are mainly concentrated in the low-frequency band. At the same time, the main interferences such as train wheel-rail impact also contain a large number of low-frequency components. Therefore, a low-frequency piezoelectric sensor with a frequency response range of 0.2Hz to 1500Hz is selected. The lower limit (0.2Hz) is intended to capture the extremely low-frequency modal drift or slow stiffness degradation signals caused by damage. The upper limit (1500Hz) is sufficient to cover most of the high-frequency resonance components related to damage, while avoiding the introduction of excessively high-frequency environmental electronic noise. This "narrow window width" selection is essentially a preset optimal physical filter for subsequent signal processing, realizing the maximum capture of the target signal frequency band and the initial suppression of non-target noise at the hardware level. Furthermore, since vibration follows a clear physical transmission path from the excitation source to the damage point and then to the sensor: rail (rail web / rail base) → sleeper → ballast bed → tunnel foundation, damage will change the vibration characteristics of each node along the path. Therefore, sensors deployed on the rail web, rail base, sleeper surface, ballast bed surface, and tunnel wall constitute an observation array covering the entire transmission chain. Sensors placed on the rail web or rail base can serve as excitation input reference points, monitoring the original load characteristics for subsequent signal comparison and eliminating the impact of changes in train operating status. Sensors placed on the sleeper surface can serve as near-field damage monitoring points, directly sensing changes in the vibration response of the sleeper itself. Sensors placed on the ballast bed surface can serve as energy transfer monitoring points, assessing the diffusion of vibration energy to the foundation, as damage may cause abnormal energy reflection at the sleeper-ballast interface. Sensors placed on the tunnel wall can serve as environmental background reference points, providing background vibration levels far from the vibration source, used to distinguish between overall system vibration and local sleeper damage response.

[0025] This topology enables the system to synchronously collect end-to-end information from "input-response-transmission-background," providing a data foundation for separating components related only to sleeper damage from complex coupled signals.

[0026] The collected vibration signals are subjected to noise reduction processing; Furthermore, the noise reduction process employs the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) method, which includes the following steps: Gaussian white noise is added to the signal to be decomposed. The new signal is obtained, and a total of N preprocessed sequences are constructed. , where n=1,2,3....N.

[0027]

[0028] in: These are the Gaussian white noise weighting coefficients; The Gaussian white noise sequence injected in the nth iteration.

[0029] For all preprocessed sequences Perform EMD (Empirical Mode Decomposition) to extract the first-order IMF set. By eliminating random noise interference through ensemble averaging, the first-order IMF (Intrinsic Mode Function) of CEEMDAN is generated. At the same time, the first residual sequence is obtained. ,as follows:

[0030]

[0031] In the In the first-order decomposition stage, the previous-order residuals are injected with Gaussian white noise to perform EMD decomposition, i.e.:

[0032]

[0033] in: The first term obtained from the CEEMDAN decomposition is... One IMF component; This indicates the extraction of the first... One IMF component; This indicates that the CEEMDAN decomposition is related to the first... The weights of the order residuals are added to Gaussian white noise; Indicates the first The residual of the order.

[0034] The decomposition process continues until the stopping condition of the EMD decomposition method is met, at which point the iteration terminates.

[0035] Furthermore, Empirical Mode Decomposition (EMD) is an adaptive signal processing method that decomposes complex signals into a series of IMF components with different frequency characteristics. These IMF components have clear physical meanings and can reflect the local characteristics of the signal at different time scales. The application of EMD in signal denoising is mainly based on the difference in frequency characteristics between noise and useful signals. By decomposing the signals into IMF components, the signal is reconstructed, thereby achieving the purpose of denoising.

[0036] There are several assumptions when performing EMD decomposition: The signal has at least two extreme points, one maximum and one minimum. The time-scale characteristic is determined by the time scale between two extreme points.

[0037] The basic principle of the EMD algorithm is: Find the original signal That is, preprocessing sequence The maximum and minimum points are identified, and then these extreme points are fitted using curve interpolation to obtain the upper envelope of the signal. and lower envelope .

[0038] Calculate the average value of the routes for the upper and lower packages:

[0039] Then subtract the mean from the original signal to obtain the remaining components after removing the low-frequency components. ,Right now:

[0040] The purpose of EMD decomposition is to decompose a signal f(t) into N IMF components and one residual component. Each IMF needs to satisfy the following two conditions: Throughout the entire data range, the number of local extrema and zero-crossing points must be equal, or the difference in number must be at most 1. At any given time, the average of the envelopes of the local maxima (upper envelope) and the envelopes of the local minima (lower envelope) must be zero.

[0041] Next, make a judgment. Does it meet the two conditions of the IMF? If so, then the remaining component... Save it as an IMF record. Otherwise, Treating this as the original signal, repeat steps 1-2 above to continue filtering. Let's assume the result after the k-th filtering is... Only if it meets the IMF criteria, then This is reserved for the first-tier IMF and is denoted as ,Right now:

[0042] minus We obtain the first residual, namely:

[0043] Again Treating it as a new set of signals, repeat the above three steps to obtain the IMF component that meets the conditions again. Repeat this process multiple times until... The decomposition terminates when the function is monotonic or less than the threshold. We can obtain:

[0044] in This is a residual quantity, reflecting the average trend of signal changes; Different IMF components are arranged in descending order of frequency.

[0045] The stopping of the screening is determined by the magnitude of the standard deviation (SD). That is:

[0046] Where: T represents the entire time span of the signal; SD ranges from 0.2 to 0.3; and k is the number of filtering iterations.

[0047] Extract time-domain and frequency-domain features from the denoised signal; The time-domain features include dimensional features and dimensionless features. Dimensional features include mean, variance, root mean square value, and peak value; dimensionless features include peak factor, kurtosis coefficient, waveform factor, impulse factor, and margin factor; and frequency-domain features include centroid frequency, frequency variance, and mean square frequency. The mean value represents the mathematical expectation of the static components of the signal. This parameter reflects the baseline value of the signal energy distribution, is less susceptible to transient shocks, and has significantly better stability than the peak value. Its expression is as follows:

[0048] Variance characterizes the fluctuation intensity of the dynamic components of a signal and quantifies the dispersion of data in the mean neighborhood. Its expression is as follows:

[0049] The root mean square (RMS) value is a time-domain statistical measure representing the energy of a signal. Its physical meaning is equivalent to the average power of the steady-state signal, and its expression is as follows:

[0050] The peak capture signal's maximum instantaneous amplitude, which exhibits high sensitivity to transient impact events, is expressed as follows:

[0051] The peak factor eliminates signal amplitude sensitivity through normalization. Its mathematical definition is the ratio of the signal peak value to the root mean square value, expressed as follows:

[0052] The kurtosis coefficient, by quantifying the tail characteristics of the data distribution, provides a key basis for identifying transient impact events in non-stationary signals. Its expression is as follows:

[0053] The waveform factor reflects the smoothness and shape characteristics of the signal waveform, and its expression is as follows:

[0054] The magnitude of the impulse factor reflects the significance of impulses in the signal, as expressed below:

[0055] The margin factor reflects the amplitude variation range and stability of the signal, and its expression is as follows:

[0056] The centroid frequency characterizes the location of the main energy distribution of a signal, and its expression is as follows:

[0057] Where: N represents the total number of frequency points involved in the calculation. This represents the frequency value corresponding to the i-th frequency point; This represents the energy value corresponding to the i-th frequency point.

[0058] Frequency variance is an important indicator in frequency domain analysis used to quantify the dispersion of the frequency components of a signal. It reflects the concentration or dispersion of the signal's dominant frequency components by statistically analyzing the distribution characteristics of the spectral energy. The expression is as follows:

[0059] in: f c This indicates the centroid frequency.

[0060] The mean square frequency is the square root of the second moment of the spectral energy, reflecting the degree of dispersion of the signal frequency distribution. It is used to distinguish between steady-state and transient vibrations, and its expression is as follows:

[0061] Frequency domain analysis, acting as a "frequency microscope" for vibration signal identification, overcomes the shortcomings of time domain analysis in its insensitivity to periodic features. By quantifying the frequency components and energy distribution of vibration signals, it provides a richer diagnostic basis for the state assessment of complex systems.

[0062] The extracted features are input into the trained classification and recognition model to identify the degree of internal cracks in the sleepers and output the recognition results. As required, the classification and recognition model is an extreme learning machine model based on the particle swarm optimization algorithm, and its training process includes: Randomly initialize the input weights and hidden layer biases; Optimize hidden layer node parameters using particle swarm optimization algorithm; An extreme learning machine model is constructed based on the optimized parameters, and the output weights are solved analytically using the generalized inverse matrix. The basic principle of the Extreme Learning Machine (ELM) is as follows: N samples are constructed based on the time-domain signal and the frequency-domain signal. ,in , For a single-hidden-layer neural network with L hidden layer nodes, i=1,2,…,N, the expression is:

[0063] in, These are weighting coefficients; It is an activation function; and These are parameters for hidden layer nodes.

[0064] Within the Extreme Learning Machine framework, the output weight matrix is... The optimization objective can be formalized as a loss function minimization problem with a regularization term, as follows:

[0065] in: It is the output matrix of the hidden layer; These are the output weights of the training data; It is the target output matrix for training.

[0066]

[0067]

[0068] Therefore, the optimal solution that can be derived from linear algebra is:

[0069] in: It is a generalized inverse matrix.

[0070] The ELM algorithm avoids the time cost of iterative parameter tuning in the traditional backpropagation algorithm by randomly initializing the weight matrix and bias vector from the input layer to the hidden layer and keeping them constant. It directly calculates the output layer weights based on the generalized inverse matrix of the hidden layer output, transforming the complex non-convex optimization problem into a system of linear equations to be solved, thus improving training efficiency by tens to hundreds of times.

[0071] At this point, constructing an Extreme Learning Machine (ELM) model using a simple method of randomly selecting initialization parameters suffers from drawbacks such as redundancy in hidden layer nodes, which hinders the accuracy of the ELM model to some extent. Furthermore, the ELM model also has the following problems: (1) Random initialization of hidden layer parameters may cause fluctuations in model performance, especially when the number of hidden layer nodes is insufficient or the data distribution is complex, it is easy to fall into suboptimal solutions; (2) Random parameters may not be able to fully capture the nonlinear relationships in the data, resulting in large prediction errors, especially in high-noise or high-dimensional scenarios where they are unstable. (3) Although the analytical solution of ELM can theoretically reach the global optimum, in practice the randomness of the hidden layer parameters may cause the solution of the output weight to deviate from the optimal direction.

[0072] Therefore, in the process of building a data prediction model, the optimal parameters of the model can be determined by the Particle Swarm Optimization (PSO) algorithm to ensure the accuracy of the ELM model.

[0073] In constructing the ELM prediction model, the parameter optimization capability of the PSO algorithm is utilized. A fitness function is established to calculate the fitness value, thus verifying the quality of particles in the PSO algorithm. Particles are continuously updated to ensure they constantly approach the global optimum. During prediction, the error between the predicted result and the true value is used as the fitness function. By finding the global optimum that minimizes the fitness function, a more accurate ELM prediction model is determined, providing a theoretical foundation for constructing a reasonable ELM prediction model.

[0074] In addition, the particle swarm optimization algorithm uses the prediction error as the fitness function and searches for the global optimum by iteratively updating the particle positions. Furthermore, the identification results are classified according to the severity of the crack, including no crack, micro-crack, or through crack.

[0075] Based on the identification results, a sleeper health status assessment and early warning will be conducted, specifically including: Based on the identification results and historical feature data of the target sleeper in multiple consecutive monitoring cycles, the feature evolution gradient vector reflecting the change in its damage severity is calculated. By calculating the feature evolution gradient vector, the rate of change of key damage indicators (such as kurtosis and center of gravity frequency) is mathematically described. Then, using a pre-trained damage dynamics model (which can be constructed based on a large amount of historical damage data or physical simulation), the current gradient vector is extrapolated to predict the development trajectory of the damage in the future. Using a pre-trained damage dynamics model, based on the feature evolution gradient vector, the trend curve of the severity level of the damage changing with time or the number of load cycles under no intervention conditions is predicted. Based on the trend curve, the inherent evolution probability of the damage undergoing a severity level jump within the next preset long-term maintenance planning window is calculated, reflecting the inherent possibility of the damage deteriorating under ideal conditions (without considering external constraints). Calculate the probability of dynamic intervention risk using actual operation and maintenance constraint parameters; Set inherent risk thresholds and dynamic risk thresholds. When the probability of dynamic intervention risk is greater than or equal to the dynamic risk threshold, trigger an immediate action warning. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is greater than or equal to the inherent risk threshold, a planning warning is triggered. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is less than the inherent risk threshold, no warning is triggered. The final output is a structured report containing the warning level, damage parameters, recommended measures, and recommended time windows.

[0076] Furthermore, the specific steps for calculating the probability of dynamic intervention risk using actual operation and maintenance constraint parameters are as follows: Determine the nearest available repair times and repair durations for sunroofs; Assess the conditional probability that, given the current trend of damage evolution, the damage may accelerate and spread before the sunroof repair time, potentially affecting driving safety. By combining the estimated operation time required to complete the repair of the damage point with the matching degree of the maintenance window duration, the probability of dynamic intervention risk is calculated; By incorporating the sunroof schedule and maintenance operation duration as core variables into the risk assessment formula, the early warning system gains resource awareness. Its output naturally matches the actual operation and maintenance capabilities, avoiding invalid alarms that cannot be responded to due to insufficient resources, and greatly improving the operability and authority of the early warning information.

[0077] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying the degree of internal damage to railway sleepers based on vibration signals, characterized in that, Specifically, it includes: Vibration sensor arrays are deployed on sleepers, track bed and related structures to collect vibration signals; The collected vibration signals are subjected to noise reduction processing; Extract time-domain and frequency-domain features from the denoised signal; The extracted features are input into the trained classification and recognition model to identify the degree of internal cracks in the sleepers and output the recognition results. Based on the identification results, the health status of railway sleepers is assessed and an early warning is issued.

2. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The vibration sensor is a low-frequency piezoelectric vibration sensor with a frequency range of 0.2Hz to 1500Hz, and its placement locations include the rail web, rail bottom, sleeper surface, ballast surface, and tunnel wall.

3. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The noise reduction process employs a fully adaptive noise ensemble empirical mode decomposition method, including the following steps: Inject adaptive Gaussian white noise into the original signal; Empirical mode decomposition is performed on the injected noise signal to extract multiple intrinsic mode function components; By performing multiple iterations and ensemble averaging, the influence of noise is eliminated and the signal is reconstructed.

4. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The time-domain features include dimensional features and dimensionless features, wherein the dimensional features include mean, variance, root mean square value, and peak value. Dimensionless features include peak factor, kurtosis coefficient, waveform factor, impulse factor, and margin factor.

5. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The frequency domain features include centroid frequency, frequency variance, and mean square frequency.

6. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The classification and recognition model is an extreme learning machine model based on the particle swarm optimization algorithm, and its training process includes: Randomly initialize the input weights and hidden layer biases; Optimize hidden layer node parameters using particle swarm optimization algorithm; An extreme learning machine model is constructed based on the optimized parameters, and the output weights are solved analytically using the generalized inverse matrix.

7. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 6, characterized in that: The particle swarm optimization algorithm uses prediction error as the fitness function and iteratively updates particle positions to find the global optimum.

8. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The identification results are classified according to the severity of the crack, including no crack, microcrack, or through crack.

9. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 1, characterized in that: The assessment and early warning of sleeper health status based on the identification results specifically includes: Based on the identification results and historical feature data of the target sleeper in multiple consecutive monitoring cycles, the feature evolution gradient vector reflecting the change in its damage severity is calculated. Using a pre-trained damage dynamics model, based on the feature evolution gradient vector, the trend curve of the severity level of the damage changing with time or the number of load cycles under no intervention conditions is predicted. Based on the trend curve, calculate the inherent evolution probability of the damage's severity level jumping within the next preset long-term maintenance planning window; Calculate the probability of dynamic intervention risk using actual operation and maintenance constraint parameters; Set inherent risk thresholds and dynamic risk thresholds. When the probability of dynamic intervention risk is greater than or equal to the dynamic risk threshold, trigger an immediate action warning. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is greater than or equal to the inherent risk threshold, a planning warning is triggered. When the probability of dynamic intervention risk is less than the dynamic risk threshold and the probability of inherent evolution is less than the inherent risk threshold, no warning is triggered. The final output is a structured report containing the warning level, damage parameters, recommended measures, and recommended time windows.

10. The method for identifying the degree of internal damage of railway sleepers based on vibration signals according to claim 9, characterized in that: The specific steps for calculating the probability of dynamic intervention risk using actual operation and maintenance constraint parameters are as follows: Determine the nearest available repair times and repair durations for sunroofs; Assess the conditional probability that, given the current trend of damage evolution, the damage may accelerate and spread before the sunroof repair time, potentially affecting driving safety. By combining the estimated operation time required to complete the repair of the damage point with the matching degree of the skylight duration, the probability of dynamic intervention risk is calculated.

Images