Railway line condition assessment method

Abstract

The application discloses a railway line state evaluation method, and belongs to the technical field of intelligent operation and maintenance of railway infrastructure. The method aims to solve the problem that the existing evaluation method cannot adapt to individual differences and long-term state dynamic changes of the line due to the use of fixed threshold or static model, and the "screening paradox" problem faced during online model updating. The method constructs a dynamic health baseline model based on historical data to output the normal value interval of the state parameter; the health index is calculated in real time and buffered; the buffered data is screened in two stages and concept drift is detected, wherein the preliminary screening uses an auxiliary reference benchmark independent of the baseline model; then, whether a significant concept drift occurs is determined based on whether the statistical distance exceeds the threshold value, and the model is updated online or the baseline is reset according to the decision. The method is mainly used for realizing continuous, self-adaptive and accurate evaluation and early warning of the state of the railway line, and providing decision support for predictive maintenance.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and operation and maintenance technology of railway infrastructure, and specifically relates to a method for assessing the condition of railway lines. Background Technology

[0002] In the field of railway engineering, accurate assessment of track conditions is crucial for ensuring transportation safety and guiding scientific maintenance. Currently, assessment methods based on fixed thresholds are the widely used standard practice. This method relies on uniform, pre-set static thresholds (such as limits on track geometry), directly comparing real-time monitoring data with these thresholds to determine whether the condition exceeds limits and trigger an early warning. Although this method is standardized and easy to operate, its inherent limitations have gradually become apparent in long-term practice.

[0003] First, the fixed threshold method fails to reflect the individual differences and dynamic evolution of railway line conditions. Different sections of a railway line differ due to variations in their basic conditions, operational load, and natural environment, resulting in inherently different baselines and rates of degradation in their health status. Applying a uniform static threshold is like using the same ruler to measure everyone, ignoring this individual specificity. This could lead to overprotection of some sections and underprotection of others, resulting in inaccurate assessments. More importantly, railway line conditions are a dynamic process that changes slowly over time with maintenance interventions. Fixed thresholds cannot track and adapt to this change. A threshold set at the beginning of operation may become inapplicable years later due to the slow decline in overall line performance, either generating numerous false alarms or underreporting real risks, demonstrating insufficient system adaptability.

[0004] Secondly, with advancements in data acquisition technology, methods for building state models using machine learning based on historical data have gained attention, aiming to achieve more intelligent evaluations. However, these data-driven methods face challenges in maintaining model validity during long-term online applications, primarily due to the "model drift" problem. Slow changes in line conditions, environmental factors, or the detection equipment itself can cause the distribution of newly acquired data to gradually deviate from the historical data distribution at the beginning of model training. If the model remains static, its prediction and evaluation accuracy will decay over time. To address this issue, the natural approach is to enable the model to learn from new data online and update itself. However, this leads to a more fundamental dilemma: if all new data is used to update the model indiscriminately, the random noise, short-term anomalies, and even erroneous data caused by equipment failures will contaminate the model, causing it to learn incorrect patterns. Conversely, if the goal is to select "normal" data for updates, a reliable selection criterion is lacking—because the benchmark used to determine whether data is "normal" is precisely the model itself, which may have become inaccurate due to drift. This creates a logical circular dependency, where a potentially faulty model is used to filter data to correct its own failure. This makes it difficult to reliably and automatically implement online adaptive updates of the model in practice.

[0005] Finally, existing condition assessment methods often focus on a single assessment dimension or data source. For example, they may only be based on track geometry data or only focus on the dynamic response of the vehicles. The overall health of a track is the result of the combined effects of multiple subsystems (such as track structure, vehicles, and wheel-rail relationships). Assessing from only a single perspective makes it difficult to comprehensively and holistically reflect the true condition of the system, and the assessment conclusions may be biased.

[0006] Therefore, in the field of railway line condition assessment, establishing a method that can adapt to the individual characteristics of different lines, track long-term dynamic changes in condition, reliably and stably achieve model self-updating, and integrate multi-dimensional information for comprehensive assessment is a pressing technical challenge. The difficulty lies in designing a mechanism that can overcome the rigidity of fixed thresholds and escape the logical dilemma of the "screening paradox" in data-driven models, thereby enabling the long-term, autonomous, and accurate operation of the assessment system. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0008] One objective of this invention is to address the problems of "model drift" and "data filtering paradox" encountered in the long-term online application of static data-driven models, which fail to reflect individual differences and long-term dynamic changes in railway line condition assessment due to the use of uniform static thresholds.

[0009] One objective of this invention is to address the problem of the lack of clear and objective construction standards for "auxiliary reference benchmarks".

[0010] One objective of this invention is to address the problem that the "secondary verification" step may be too mechanical and rigorous, thus incorrectly filtering out "marginal valid data" that slightly exceeds the "normal confidence interval" of the current model but actually reflects early and minor changes in the state.

[0011] One objective of this invention is to address the problem of misjudgment (frequent reset triggers) or missed judgment (insensitivity to slow drift) caused by data noise and short-term disturbances when determining "significance concept drift" based on a single instantaneous threshold.

[0012] One objective of this invention is to address the problem that when a model is continuously updated through incremental learning, it may gradually "forget" important knowledge (such as seasonal patterns and baseline performance) learned from historical data in the early stages that represents the long-term stability characteristics of the line, i.e., the "catastrophic forgetting" problem.

[0013] One objective of this invention is to address the problem that condition assessment may rely on a single type of data (such as track geometry only or vehicle response only), resulting in a single assessment dimension and an inability to comprehensively reflect the overall health status of the track-vehicle coupled system.

[0014] One objective of this invention is to address the problem that when railway lines undergo planned major overhauls, technical upgrades, or other known and controlled step changes in state, conventional adaptive evaluation processes may misjudge and filter out a large amount of "implicitly normal" data that characterizes the new state as abnormal, leading to the model's inability to learn new benchmarks and the long-term failure of the evaluation system.

[0015] One objective of this invention is to address the problem that if the "periodic historical review" step is simply retrained by merging data, it may lead to parameter conflicts or oscillations between "consolidating old knowledge" and "integrating new knowledge," affecting training efficiency and the final performance of the model.

[0016] One objective of this invention is to address the problem that using fixed weights in multi-source data fusion cannot adapt to changes in the relative importance of track conditions and vehicle dynamic response for safety assessment at different train operating speeds.

[0017] One objective of this invention is to address the problem that during the "reinforcement learning period," after the intelligent screening constraints are completely removed, obviously unreasonable or erroneous abnormal data may flow into and contaminate the model without hindrance due to factors such as occasional sensor malfunctions or extreme interference.

[0018] One object of the present invention is to provide a method for assessing the condition of a railway line, comprising the following steps: Obtain the time-series status monitoring data of the to-be-evaluated section of the target railway line within the historical monitoring period and perform preprocessing to form an initial training dataset; Based on the initial training dataset, use an unsupervised learning method to construct a dynamic health baseline model for the to-be-evaluated section. The dynamic health baseline model is used to output the normal value range of the status parameters at a given confidence level; Continuously obtain the current status monitoring data of the to-be-evaluated section, calculate the real-time health indicators according to the dynamic health baseline model, and store the current status monitoring data and the corresponding real-time health indicators in a first-in-first-out data buffer pool; at the same time, compare the real-time health indicators with the preset dynamic warning threshold or analyze their change trends to generate a health status evaluation result and a warning message; Perform online screening and concept drift detection on the data in the data buffer pool: First, based on an auxiliary reference benchmark, preliminarily screen the data in the data buffer pool to obtain the primary selected data; the auxiliary reference benchmark is independent of the dynamic health baseline model; subsequently, perform secondary verification on the primary selected data based on the dynamic health baseline model, and mark the data with the real-time health indicators within their normal confidence intervals as candidate update data; calculate the statistical distance between the data distribution characteristics of the candidate update data and the historical data distribution characteristics represented by the dynamic health baseline model, and when this statistical distance exceeds the preset first threshold, it is determined that a significant concept drift has occurred; if it is not determined that a significant concept drift has occurred, use the candidate update data to perform online update of the dynamic health baseline model through an incremental learning algorithm; if it is determined that a significant concept drift has occurred, trigger the baseline reset process, that is, re-train based on the new data after the drift occurs and replace the original dynamic health baseline model, and synchronously reset the dynamic warning threshold.

[0019] Preferably, in the railway line status evaluation method, the auxiliary reference benchmark is determined in the following manner: based on all the time-series status monitoring data collected within the preset historical period, calculate the statistical distribution of each status parameter, and define the preset high percentage range in the middle of the statistical distribution as the auxiliary reference benchmark.

[0020] Preferably, in the railway line status evaluation method, in the secondary verification step, if the real-time health indicator value x corresponding to the primary selected data is not within the normal confidence interval [L, U], and its absolute difference from the nearest boundary is less than a preset tolerance threshold δ, the primary selected data is still marked as candidate update data; where the absolute difference is defined as: L−x when x < L, and x−U when x > U.

[0021] Preferably, in the railway line status evaluation method, the determination of a significant concept drift is carried out in the following manner: Record the statistical distances calculated in each online screening and concept drift detection step to form a statistical distance sequence; Set a sliding decision window of fixed length and a frequency threshold; When tracing back a sliding decision window based on the current statistical distance, if the number of times the statistical distance exceeds the first threshold within the window reaches or exceeds the frequency threshold, a significant concept drift is determined to have occurred.

[0022] Preferably, in the railway line condition assessment method, the step of updating the dynamic health baseline model online using candidate update data through an incremental learning algorithm further includes a periodic historical review step: When the amount of candidate update data used in online incremental updates reaches a preset proportion of the initial training dataset, a portion of samples is extracted from the initial training dataset and merged with the recently accumulated candidate update data to retrain the dynamic health baseline model with all parameters.

[0023] Preferably, in the railway line condition assessment method, the condition monitoring data includes track geometric irregularity data and vehicle dynamic response data; the step of calculating real-time health indicators based on the dynamic health baseline model specifically includes: Sub-index calculation: Based on track geometric irregularity data and vehicle dynamic response data, the corresponding track state sub-health index and vehicle dynamic sub-health index are calculated respectively. Fusion assessment: Based on preset weighting coefficients, the track status sub-health index and the vehicle dynamic sub-health index are weighted and averaged to generate a comprehensive health index as a real-time health index.

[0024] Preferably, the railway line condition assessment method further includes a learning mode management step: When an external instruction is received indicating that the segment to be evaluated has entered the planned state change phase, the system enters a reinforcement learning period of a preset duration. During the reinforcement learning period, all screening and verification constraints on the data in the online screening and concept drift detection steps are temporarily lifted, so that the current state monitoring data received in the data buffer pool can be directly used as candidate update data for reinforcement incremental learning of the dynamic health baseline model. After the reinforcement learning period ends, the system automatically resumes the regular online screening and concept drift detection steps.

[0025] Preferably, in the railway line condition assessment method, the retraining in the periodic historical review step is implemented using an elastic weight consolidation strategy, specifically including: Parameter importance assessment: Calculate the Fisher information diagonal values ​​of each parameter in the dynamic health baseline model as the importance weights of each parameter; Constrained retraining: A regularization penalty term is added to the loss function during retraining. The value of this term is positively correlated with the importance weights of each parameter and their offsets relative to their values ​​before this retraining.

[0026] Preferably, in the railway line status assessment method, the preset weight coefficient is dynamically determined based on the current train operating speed, and the weight value of the vehicle dynamic sub-health index is configured to increase monotonically with the increase of operating speed.

[0027] Preferably, in the railway line condition assessment method, a data validity filtering step is included during the reinforcement learning period: The values ​​of each state parameter in the current state monitoring data received in the data buffer pool are compared with an absolute value safety threshold range obtained based on historical long-term monitoring data statistics. Only data where all parameter values ​​are within their respective absolute value safety threshold ranges are retained as candidate update data that can be directly used later.

[0028] The present invention has at least the following beneficial effects: This invention overcomes the rigidity of traditional fixed threshold methods by constructing a closed-loop system that includes independent auxiliary benchmark selection, secondary model verification, concept drift detection, and dual-path updates. This allows state assessment to adapt to the individual characteristics of different paths and track their long-term dynamic changes. Simultaneously, this method effectively resolves the "screening paradox" in online data-driven model updates, providing a feasible technical path for achieving reliable, automated, and adaptive modeling, thereby improving the continuous accuracy and adaptability of the assessment system during long-term operation.

[0029] This invention provides an objective and quantifiable standard for constructing an auxiliary reference benchmark by explicitly defining it as the high-percentage center range based on the statistical distribution of long-term historical data. This ensures that the benchmark possesses robust statistical properties and can remain an independent and reliable filtering basis even if the current dynamic health baseline model may drift. It provides stable support for the initial data screening stage of the entire adaptive update process, enhancing the system's robustness in the face of slow changes in state.

[0030] This invention introduces a preset tolerance threshold, allowing data slightly exceeding the current normal confidence interval to participate in model updates during secondary verification. This mechanism, while maintaining the overall stability of the model, increases the system's ability to perceive early, minor changes in the state. It helps avoid prematurely filtering out potentially effective learning data due to overly absolute evaluation criteria, enabling the model to more smoothly track the gradual degradation or improvement process of the line's state.

[0031] This invention employs a concept-based drift judgment method using a sliding time window and frequency threshold, elevating the judgment criterion from a single instance of statistical distance exceeding the limit to the frequency of exceeding the limit over a period of time. This design enhances the system's resistance to data noise and short-term disturbances, reducing the possibility of accidental model resets due to occasional anomalies. Simultaneously, it makes the system more sensitive to real, continuous systemic state changes, thereby improving the overall stability and reliability of model update decisions.

[0032] This invention incorporates a periodic historical review step. When online updates accumulate to a certain amount of data, a full-parameter retraining process that integrates historical and recent data is triggered. This approach helps the model review and reinforce long-term, stable patterns learned from historical data while continuously learning new knowledge. This mitigates the "catastrophic forgetting" problem that can occur with purely incremental learning and helps maintain the integrity of the baseline model's knowledge structure and the stability of its long-term predictions throughout its lifecycle.

[0033] This invention achieves a multi-dimensional comprehensive assessment of track health by extracting sub-health indicators from track geometry data and vehicle dynamic response data separately and then weighting and fusing them. This method fully utilizes the physical information (infrastructure deformation and system coupling dynamics) reflected by different data sources, enabling the generated comprehensive health indicators to more comprehensively depict the overall state of the "track-vehicle" system. The assessment results contain richer information, providing a better foundation for subsequent analysis.

[0034] This invention provides a controlled and rapid learning mode for the system to handle planned major changes in line status (such as after major overhauls) by introducing a reinforcement learning period triggered by external instructions. In this mode, temporarily removing conventional screening constraints allows "normal" data representing the new state to be efficiently injected into the model, thus avoiding the problem of conventional processes misclassifying this data as abnormal and causing long-term system failure. This ensures that the evaluation system can quickly rebuild an accurate baseline and maintain availability after the line undergoes controlled changes.

[0035] This invention employs a flexible weighting consolidation strategy during the retraining of historical reviews, imposing differentiated constraints on parameter updates based on their importance. This approach allows for more cautious modification of key model parameters during the learning process of consolidating historical knowledge, thereby reducing parameter conflicts and oscillations when integrating new and old knowledge. This contributes to a smoother and more efficient model review process, further enhancing the protection of key historical memories and the ability to maintain long-term performance.

[0036] This invention establishes a dynamic correlation between fusion weights and train operating speed, enabling the weights of vehicle dynamic sub-health indicators to monotonically increase with speed. This allows the generation process of the comprehensive health indicator to adapt to different operating conditions, placing greater emphasis on indicators reflecting the dynamic interaction stability of the system during high-speed operation. Therefore, the evaluation results better align with actual safety concerns at different speed levels, improving the scenario relevance and practicality of the condition assessment.

[0037] This invention provides a basic defense against extreme anomaly data contamination by adding a simple filtering step based on a long-term historical statistical absolute safety threshold during the reinforcement learning period, while maintaining a highly unobstructed data flow. This step intercepts erroneous data that clearly exceeds reasonable physical limits with very low computational cost, thus providing a minimum guarantee of data quality during the high-intensity learning period without affecting the main objective of rapidly learning new states, thereby enhancing the operational security of this mode.

[0038] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Detailed Implementation

[0039] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0040] One object of the present invention is to provide a method for assessing the condition of a railway line, comprising the following steps: Acquire and preprocess the time-series status monitoring data of the section of the target railway line to be evaluated within the historical monitoring period to form the initial training dataset; Based on the initial training dataset, an unsupervised learning method is used to construct a dynamic health baseline model for the segment to be evaluated. The dynamic health baseline model is used to output the normal value range of the state parameters at a given confidence level. The system continuously acquires current status monitoring data of the section to be evaluated, calculates real-time health indicators based on the dynamic health baseline model, and stores the current status monitoring data and corresponding real-time health indicators in a first-in-first-out data buffer pool. At the same time, it compares or analyzes the changing trends of the real-time health indicators with preset dynamic warning thresholds to generate health status assessment results and warning information. Perform online filtering and concept drift detection on the data buffer pool: First, the data in the data buffer pool is initially screened based on an auxiliary reference benchmark to obtain preliminary data; the auxiliary reference benchmark is independent of the dynamic health baseline model. Then, the preliminary data is re-verified based on the dynamic health baseline model, and data with real-time health indicators within their normal confidence intervals are marked as candidate update data. The statistical distance between the data distribution characteristics of the candidate update data and the historical data distribution characteristics represented by the dynamic health baseline model is calculated, and if the statistical distance exceeds a preset first threshold, a significant concept drift is determined to have occurred. If no significant concept drift is determined, the dynamic health baseline model is updated online using the candidate update data through an incremental learning algorithm. If a significant concept drift is determined, the baseline reset process is triggered, that is, the original dynamic health baseline model is retrained and replaced based on the new data after the drift occurs, and the dynamic warning threshold is reset simultaneously.

[0041] Currently, railway line condition assessment widely employs two methods: fixed threshold-based assessment and offline-trained static model-based methods. Fixed threshold methods judge whether a condition exceeds limits based on a uniform standard. Their drawback is that they cannot reflect the individual characteristics of different line sections, nor can they track the slow evolution of line condition over time, leading to potentially inaccurate assessment results. While offline static model methods can learn historical patterns to some extent, their models remain fixed once established. When line condition changes due to natural degradation, environmental changes, or maintenance work, model performance gradually declines, resulting in model drift. Attempting to make static models learn new data online to update themselves encounters a fundamental problem: the standard used to judge whether new data is "normal" and suitable for updating is precisely the potentially inaccurate model itself. This creates a logical circular dependency, making reliable adaptive updates difficult to achieve in practice.

[0042] To address the aforementioned problems, this invention provides a specific implementation method. First, the method collects time-series monitoring data on track geometry irregularities and vehicle dynamic responses within a target section over historical periods. After preprocessing such as cleaning and alignment, an initial training dataset is formed. Next, a Gaussian mixture model is used as an unsupervised learning method to train and generate a dynamic health baseline model for the section based on this initial dataset. In the specific construction, the optimal number of Gaussian mixture components is determined using the Bayesian information criterion, and the model parameters (including the mean vector, covariance matrix, and mixture weights of each Gaussian component) are optimized using the expectation-maximization algorithm. After training, for any given state parameter vector... x The model is able to calculate the probability density value of its belonging to the "normal state". p ( x By analyzing the model's performance on the initial training set, a probability density threshold can be determined. pth This ensures that approximately 95% of the samples in the initial data meet the requirements. p ( x )≥ p th This threshold p th The region enclosed by the corresponding probability density contour surface is defined as the normal range of the state parameter at that confidence level.

[0043] During operation, the method continuously acquires current monitoring data and uses the aforementioned dynamic health baseline model to calculate a real-time comprehensive health indicator, which reflects the degree of deviation of the current state from the individual's baseline. This current data and the corresponding health indicator are stored in a time-sequential, first-in, first-out data buffer. Simultaneously, the method compares this real-time health indicator with a preset, adjustable dynamic warning threshold, or analyzes its changing trends over a period of time, thereby generating assessment results and warning information.

[0044] The core of this method lies in a closed-loop process of online screening and model update decisions. This process first uses an auxiliary reference benchmark, independent of the aforementioned dynamic health baseline model, to initially screen new data in the data buffer pool. This auxiliary benchmark can be based on a relatively broad safety range derived from long-term historical data statistics across the entire network. Data that passes this screening is called preliminary selection data. Subsequently, the current dynamic health baseline model is used to perform a second verification on this batch of preliminary selection data, marking only those data whose health indicators fall within the normal confidence interval of the model's current output as candidate update data.

[0045] Next, the method calculates the sample mean vector of the candidate updated data. m new And the mean vector of the historical data distribution represented by the dynamic health baseline model. m hist (i.e., the weighted average of the overall mean of the model or the mean of each Gaussian component obtained during the initial training phase) is compared. The statistical distance used is Mahalanobis distance, calculated using the following formula: D =( m new - m hist ) T Σ ﹣1 ( m new - m hist ) Where Σ is the overall covariance matrix of the historical data distribution (which can be derived from the parameters of the Gaussian mixture model). The first threshold D thThe threshold can be pre-determined through Monte Carlo simulation based on the initial training dataset: multiple random sample sets of similar size to the candidate update data batch are drawn from the initial data, the Mahalanobis distance between each subset and the entire set is calculated, and the 95th percentile of these distance values ​​is taken as the first threshold. D th .

[0046] If this statistical distance D Not exceeding the threshold D th If no fundamental state concept drift has occurred, then this method utilizes this batch of candidate update data to fine-tune the dynamic health baseline model online through an incremental learning algorithm. When the model is a Gaussian mixture model, an online expectation-maximization algorithm can be used: the candidate update data is treated as a new batch of observations, and the current model parameters are used as priors. The data undergoes an E-step (calculating the posterior probability of each data point belonging to each Gaussian component) and an M-step (updating the mean, covariance, and mixture weights of each Gaussian component) to achieve incremental model updates.

[0047] If this statistical distance D Exceeded the threshold D th If a significant conceptual drift occurs, it indicates that the line status may have undergone a step change. In this case, the method will trigger a baseline reset process, which means stopping the use of the old model and instead retraining a completely new dynamic health baseline model based on the new data collected after the drift determination to replace the original model, while simultaneously updating the relevant warning thresholds.

[0048] Through the steps described above, this method can establish and maintain an individualized health baseline for each line segment. This baseline can adaptively evolve through two paths: for slow changes, incremental learning is used to track them; for drastic changes, the model is decisively reset to relearn. This mechanism effectively overcomes the rigidity problem of the fixed threshold method and avoids the logical paradox of static model updates, making the state assessment more realistic and possessing long-term applicability.

[0049] In a preferred embodiment, the auxiliary reference benchmark in the railway line condition assessment method is determined by: calculating the statistical distribution of each condition parameter based on all time-series condition monitoring data collected within a preset historical period, and defining the preset high percentage range in the center of the statistical distribution as the auxiliary reference benchmark.

[0050] Whether it is a fixed threshold method or other adaptive schemes, if a reference benchmark is introduced, its setting often depends on empirical values, theoretical calculation values ​​or simple statistics such as maximum and minimum values. These methods may be too sensitive or conservative and cannot stably reflect the natural fluctuation range of the line under long-term normal operation, resulting in insufficient reliability of the preliminary screening stage.

[0051] To address the aforementioned problems, this embodiment of the invention provides a benchmark determination method based on objective historical statistics. Specifically, this method first requires establishing a sufficiently long historical statistical period, such as a complete calendar year, to cover the environmental impacts of different seasons. Then, it collects all time-series status monitoring data for the target section or similar lines with similar operating conditions within this historical period; these data should represent the normal operating status of the line. Next, it performs statistical analysis on each status parameter in these historical data, such as track elevation deviation, track alignment deviation, and levelness, calculating their probability distribution.

[0052] The key step lies in the fact that this method does not simply take the average and subtract the standard deviation, nor does it directly use extreme values. Instead, it selects a preset high percentage range located in the center of the calculated statistical distribution as an auxiliary reference benchmark for the parameter. In a preferred embodiment, this range is defined as the 95% central interval of the parameter (i.e., the 2.5% quantile to the 97.5th quantile). This range can cover the vast majority of normal fluctuation data while effectively excluding extreme outliers. The percentage can be adjusted between 90% and 99% according to the stability of the line data and the requirements of risk assessment. Finally, combining these high percentage ranges calculated for each state parameter constitutes a multidimensional auxiliary reference benchmark.

[0053] The auxiliary reference benchmark determined through this implementation method is a conservative range with clear statistical significance based on a large amount of actual operating data. It is more representative of the inherent, long-term normal fluctuation characteristics of the line than a single threshold or a simple extreme value range. Therefore, in the initial screening, it can serve as a stable and reliable independent filter, effectively distinguishing normal fluctuation data from potentially abnormal data, and providing more controllable data input for subsequent processes.

[0054] In a preferred embodiment, in the railway line status assessment method, during the secondary verification step, if the real-time health indicator value corresponding to the initially selected data is... x Not within the normal confidence interval [ L , U Within [a certain range], and the absolute difference between it and the nearest boundary is less than a preset tolerance threshold. d If the initial selection data is still marked as candidate updated data, then the absolute difference is defined as: when x < L At that timeL - x ,when x > U At that time x - U .

[0055] The specific embodiments of this invention add a tolerance mechanism to the secondary verification step. In conventional model verification logic, the criteria for judging whether data is usable are usually binary and absolute, that is, data points must be completely within the normal range defined by the model to be considered valid. Although this strict rule is simple, it has significant shortcomings in practical applications, especially when dealing with slowly changing dynamic systems. It may incorrectly filter out data points that are slightly outside the current confidence interval due to slight lag in the model itself or small fluctuations in measurement, but which actually still reflect the normal or early changing state of the system. This may cause the model to lose the opportunity to learn these subtle but important changes, thereby affecting its ability to track the gradual degradation of the system or adapt to a new steady state.

[0056] To address the aforementioned issues, this embodiment of the invention introduces a preset tolerance threshold, providing flexibility to the boundary conditions of the secondary verification. Specifically, during the secondary verification, if the calculation finds that the real-time health indicator value corresponding to a certain initially selected data does not fall within the normal confidence interval currently output by the dynamic health baseline model, the method does not immediately discard it. Instead, it further calculates the absolute difference between the indicator value and the nearest boundary of the interval. For example, if the lower boundary of the interval is... L The upper boundary is U The indicator value is x and less than L The difference is L - x ;like x Greater than U The difference is x - U .

[0057] Next, the method compares this calculated difference with a pre-set tolerance threshold. d Compare this tolerance threshold. d This is a small positive number, the value of which can be set based on a balance between data noise levels and model update robustness, for example, five percent of the normal confidence interval width. If the absolute difference mentioned above is less than this tolerance threshold... d This means that the data point is only slightly out of bounds, in a "grey area" near the boundary. Therefore, the method will make an exception and still mark the initial data as candidate update data, allowing it to enter the subsequent model update process.

[0058] This implementation transforms the secondary verification process from a rigid "either / or" judgment to a flexible judgment with boundary tolerance. This allows the method to maintain the stability of the main model and resist obvious anomalies while incorporating marginal data containing potential state change information. This helps the model respond more smoothly and promptly to early or slow state transitions, avoiding model update stagnation or sluggishness that may result from overly stringent verification criteria, thus improving the sensitivity and continuity of the entire adaptive evaluation process.

[0059] In a preferred embodiment, the railway line status assessment method determines the occurrence of significant concept drift in the following manner: the statistical distances calculated in each online screening and concept drift detection step are recorded to form a statistical distance sequence; a sliding decision window of fixed length and a frequency threshold are set; when tracing back one sliding decision window based on the current statistical distance, if the number of times the statistical distance exceeds the first threshold within the window reaches or exceeds the frequency threshold, then a significant concept drift is determined to have occurred.

[0060] The specific embodiments of this invention optimize the judgment logic for saliency concept drift. Basic drift detection methods typically employ instantaneous decision rules, i.e., after calculating a statistical distance, it is immediately compared with a fixed threshold; if it exceeds the threshold, a drift is determined to have occurred. The drawback of this method is its weak anti-interference capability. Railway line monitoring data inevitably contains noise and abnormal fluctuations caused by accidental impacts, instantaneous sensor errors, or short-term environmental interference. These factors may cause a single calculated statistical distance to accidentally exceed the threshold. If this is used to determine drift and trigger a costly model reset, it will be a false alarm, disrupting the model's continuity and increasing unnecessary computational overhead. Conversely, for true, slow, but continuous drift, a single distance may not have exceeded the threshold, easily leading to missed detections.

[0061] To address the aforementioned issues, this embodiment of the invention employs a decision-making strategy based on frequency accumulation within a time window. Specifically, during the online screening and concept drift detection steps, the method continuously records the statistical distance between each calculated candidate update data and the historical data distribution, and saves these records in chronological order to form a statistical distance sequence.

[0062] Meanwhile, this method presets two key parameters: one is the length of the sliding decision window. W For example, one could set a detection cycle of 10 consecutive updates; another could be a frequency threshold. N For example, set it to 7 times. The length of the sliding judgment window. W (e.g., an update cycle of 10 times) should be sufficient to smooth out short-term noise; frequency threshold. NThe setting (e.g., 7 times) needs to strike a balance between avoiding false alarms and responding promptly to actual drift, and can be determined through historical data analysis or simulation. At each point where a decision needs to be made, the method uses the latest statistical distance as a reference point and traces back a complete sliding decision window length. W View the statistical distance values ​​recorded within this time window.

[0063] The judgment rule is no longer based on whether a single distance exceeds the limit, but rather on checking the number of times the distance value exceeds a first threshold within this window. If the cumulative number of times the distance exceeds the limit reaches or exceeds a preset frequency threshold, the judgment is made accordingly. N For example, if the standard is exceeded in 7 out of the last 10 times, then the method will finally determine that a significant concept drift has occurred.

[0064] This implementation method eliminates reliance on single, instantaneous data fluctuations, instead basing the judgment on the persistence of statistical trends over a period. Occasional noise spikes are unlikely to cause exceedances across multiple consecutive periods, thus reducing the likelihood of false judgments. True, continuous systemic state changes manifest as statistical distances consistently or frequently exceeding thresholds across multiple periods, making them reliably detectable. This significantly enhances the robustness of drift detection and the stability of decision-making, making model updates and resets more reliable.

[0065] In a preferred embodiment, the railway line condition assessment method further includes a periodic historical review step: when the amount of candidate update data used in the online incremental update reaches a preset proportion of the amount of data in the initial training dataset, a portion of the samples are extracted from the initial training dataset and merged with the recently accumulated candidate update data to retrain the dynamic health baseline model with all parameters.

[0066] The specific implementation of this invention adds a periodic historical review step. In purely online incremental learning, the model adapts to changes by continuously absorbing new data and adjusting its parameters. However, this approach has an inherent flaw known as catastrophic forgetting. When the model learns new knowledge or patterns, its parameters are updated to fit the new data. This process may overwrite or weaken the parameter patterns previously learned to fit historical data. For railway line assessment, this means that the model may gradually forget stable but important historical characteristics such as seasonal patterns and basic performance levels exhibited by the line during long-term operation, leading to a degradation of its long-term representation ability and potential bias in its judgment of certain historical normal states.

[0067] To address the aforementioned issues, this embodiment of the invention introduces a periodic, proactive review mechanism in addition to the conventional incremental update process. Specifically, during operation, the method continuously records and accumulates the amount of candidate update data for online incremental updates. Simultaneously, the method presets a ratio parameter. r For example, 50% or 100%.

[0068] When the cumulative amount of candidate update data used reaches this preset proportion of the initial training dataset amount... r (For example r When the percentage of historical data reaches 50% or 100%, the periodic historical review step is triggered. At this point, the method extracts a representative subset of samples from the initial training dataset, for example, through random sampling or stratified sampling. Subsequently, these historical samples are merged with a recently accumulated batch of candidate update data to form a hybrid dataset.

[0069] Next, the method uses this mixed dataset as training data to retrain the current dynamic health baseline model with all parameters. Unlike fine-tuning incremental learning, this training performs a complete optimization process using all parameters, aiming to allow the model to regain a good fit on the new mixed dataset.

[0070] Through this periodic review and retraining, the method forces the model to repeatedly encounter and learn core patterns from historical data while absorbing new information. This helps consolidate the model's memory of long-term, stable features, mitigating the forgetting of historical knowledge caused by continuous one-way learning of new data, thereby maintaining the integrity of the model's knowledge structure and the long-term consistency of its evaluation throughout its lifecycle. This is equivalent to establishing a mechanism for periodic review for the model to balance its adaptability and stability.

[0071] In a preferred embodiment, the railway line condition assessment method includes condition monitoring data such as track geometric irregularity data and vehicle dynamic response data. The step of calculating real-time health indicators based on the dynamic health baseline model specifically includes: sub-indicator calculation: calculating corresponding track condition sub-health indicators and vehicle dynamic sub-health indicators based on track geometric irregularity data and vehicle dynamic response data respectively; fusion assessment: calculating a weighted average of the track condition sub-health indicators and vehicle dynamic sub-health indicators according to preset weighting coefficients to generate a comprehensive health indicator as a real-time health indicator.

[0072] The specific embodiments of this invention refine the calculation method for real-time health indicators and adopt a multi-source data fusion evaluation strategy. Railway line condition assessment often relies on a single type of data source. Commonly, it primarily evaluates the smoothness and geometric condition of the line based on track geometry detection data (such as elevation, alignment, and level). Another approach is to primarily evaluate driving smoothness and wheel-rail interaction based on vehicle dynamic response data (such as car body acceleration and frame acceleration). Both of these single-dimensional assessment methods have limitations: relying solely on track geometry data makes it difficult to fully reflect the dynamic response quality and ride comfort when vehicles pass; relying solely on vehicle response data makes it difficult to directly locate specific defects in the line infrastructure. This results in incomplete assessment information, failing to comprehensively depict the overall health status of the rail-vehicle coupled system, and hindering accurate maintenance guidance.

[0073] To address the aforementioned issues, this embodiment of the invention explicitly requires the simultaneous collection and utilization of two types of core data: track geometric irregularity data and vehicle dynamic response data. When calculating real-time health indicators, this method does not simply merge the two types of data, but rather performs the calculation in two steps.

[0074] First, sub-indicators are calculated. This method calculates a sub-health index of orbital state based on orbital geometric irregularity data using a dynamic health baseline model (or its derived model). I t This indicator comprehensively reflects the health of the track's geometry. Furthermore, this method, based on vehicle dynamic response data, calculates a vehicle dynamic sub-health index through a corresponding analytical model. I v This indicator reflects the dynamic performance and stability of the vehicle during operation.

[0075] Next, a fusion evaluation is performed. This method is based on the current train speed. v The weighting coefficients are determined dynamically. Specifically, this is achieved through a pre-defined, monotonically non-decreasing weighting function. w v ( v This maps speed to weights for vehicle dynamic sub-health indicators. For example: The weights of the orbital state sub-health indicators are then respectively... w t ( v )=1﹣ w v ( v This design aligns with engineering principles: at low speeds, geometry dominates, while at high speeds, dynamic interactions become more critical. Then, a weighted average of these two sub-health indicators is calculated to generate a comprehensive health indicator. I综合 : I 综合 = w t ( v ) · I t + w v ( v ) · I v This comprehensive health index I 综合 This serves as a real-time health indicator.

[0076] This implementation method incorporates information from both infrastructure status and vehicle operational quality. The comprehensive health index includes direct information on track geometry as well as indirect feedback from vehicle dynamic response, thus providing a more comprehensive and integrated reflection of the overall service status of the track system. This offers richer decision-making support for maintenance and repair. For example, when the comprehensive index deteriorates, it's possible to further observe which sub-index contributes more significantly, thereby making a preliminary judgment on whether the problem is primarily due to track geometry degradation or abnormal vehicle dynamic response.

[0077] In a preferred embodiment, the railway line status assessment method further includes a learning mode management step: when an external instruction indicating that the section to be assessed has entered the planned status change stage is received, the system enters a reinforcement learning period of a preset duration; during the reinforcement learning period, all screening and verification constraints on the data in the online screening and concept drift detection steps are temporarily lifted, so that the current status monitoring data received in the data buffer pool can be directly used as candidate update data for enhanced incremental learning of the dynamic health baseline model; after the reinforcement learning period ends, the system automatically resumes the regular online screening and concept drift detection steps.

[0078] The specific implementation of this invention adds a learning mode management step. In a fully automated adaptive evaluation system, model updates are entirely data-driven, and the system operates autonomously according to preset algorithm rules. However, in actual operation, railway lines periodically undergo planned state change phases, such as large-scale comprehensive maintenance, line technical upgrades, and replacement of rails or turnouts. After these planned engineering operations, the physical state and performance benchmarks of the line will experience significant, controlled, and abrupt improvements or changes. At this time, the distribution characteristics of the monitoring data representing the new state will fundamentally differ from the distribution of historical data before the operation. If the system still strictly follows the conventional model trained on historical data of the old state for screening and verification, this large amount of healthy "new normal" data is very likely to be judged as "abnormal" by the current model and filtered out, causing the system to be unable to learn a new healthy baseline, thus remaining in a state of inaccurate evaluation for a long time after maintenance.

[0079] To address the aforementioned issues, this embodiment of the invention introduces a special learning mode triggered and managed by external commands. Specifically, when the railway maintenance management system plans to conduct major repairs or renovations on a certain section, operators or upstream systems can send a clear external command to this evaluation method, indicating that the section to be evaluated is about to enter the planned status change phase.

[0080] Upon receiving the instruction, the method automatically enters a reinforcement learning period of a preset duration. T 强化 This timeframe needs to be sufficient for the model to learn a new stable state, and is typically set based on engineering experience; for example, 20-30 days after a major track overhaul, and 7-14 days after rail grinding. Within this period, the method temporarily removes all data screening and verification constraints from the online screening and concept drift detection steps. This means that the initial screening of the auxiliary reference benchmark and the secondary verification of the dynamic health baseline model are both suspended. Current state monitoring data received in the data buffer pool, as long as it passes the most basic data integrity check, is directly used as candidate update data.

[0081] These data were used to perform augmented incremental learning on the dynamic health baseline model. Because the data flow was no longer strictly limited by the old cognitive model, a large amount of data reflecting the new state was rapidly and centrally injected into the model, driving significant adjustments to its parameters, thereby accelerating learning and establishing a new health baseline that matches the new state.

[0082] During the pre-set reinforcement learning period T 强化After completion, the method automatically exits this mode, and the system resumes executing the regular online screening and concept drift detection steps, transitioning to normal adaptive operation. Through this implementation, the method gains the flexibility to handle planned major state changes, enabling it to collaborate with operation and maintenance management processes and ensuring that the evaluation system can quickly and safely rebuild accurate evaluation capabilities after changes to the physical baseline of the line.

[0083] In a preferred embodiment, the retraining in the periodic historical review step of the railway line status assessment method is performed using an elastic weight consolidation strategy, specifically including: parameter importance assessment: calculating the Fisher information diagonal values ​​of each parameter of the dynamic health baseline model as the importance weights of each parameter; constrained retraining: adding a regularization penalty term to the loss function of the retraining, the value of which is positively correlated with the importance weights of each parameter and their offsets relative to the values ​​before this retraining.

[0084] The specific implementation of this invention optimizes the retraining method in the periodic historical review step by employing an elastic weight consolidation strategy. Retraining typically involves merging historical and new data and then performing routine, undifferentiated full-parameter training on the model. While this method serves a review function, it carries the risk of parameter conflicts. All model parameters are treated equally during training and adjusted simultaneously to fit both old and new data. This can lead to the model over-adjusting parameters that are equally crucial for representing long-term historical patterns in order to better fit recent data patterns. Consequently, while reviewing old knowledge, it may unintentionally weaken or destroy valuable new knowledge learned recently. This conflict can sometimes reduce training efficiency and affect the model's final performance.

[0085] To address the aforementioned issues, this embodiment of the invention introduces a mechanism for parameter importance evaluation and differential constraints during retraining. This process specifically consists of two steps.

[0086] The first step is parameter importance assessment. Before initiating retraining, this method first needs to evaluate the parameters in the current dynamic health baseline model. i i The importance of this. One specific implementation involves calculating the Fisher information diagonal value for each parameter. The Fisher information diagonal value... F i It can be efficiently computed using the following approximation method: on the initial training dataset (or a retained subset of historical data) D old Above, for each parameter i i Calculate the log-likelihood function of the model L The expected value of the square of the gradient of this parameter is: , Where E is the mathematical symbol for expectation, representing the "average". The subscript "x~D" is used. old "This means that the average value is taken with respect to the variable x, and x follows a distribution D." old Here, D old This refers to preserved historical datasets (such as the initial training set or a representative subset thereof). In actual computation, this can be achieved through... D old A small batch of samples is extracted, the square of the gradient is calculated, and then the average is taken to approximate the expected value. This is used as the importance weight for each parameter.

[0087] In actual calculations, it can be done through D old A small batch of samples is extracted, the square of the gradient is calculated, and then the average is taken to approximate the expected value. This is used as the importance weight for each parameter.

[0088] The second step is constrained retraining. When retraining using a mixed dataset, this method does not simply minimize the prediction error. It modifies the training loss function. L total Add a dedicated regularization penalty item. L reg : in, L data This is the standard loss for the model on mixed datasets (such as negative log-likelihood). l The regularization coefficient is . i i old The old values ​​of the parameters before retraining. F i For parameters i i Importance weights. The core design principle of this item is: for parameters with high importance ( F i (Large), this item will change its amount in this training. For parameters of lower importance, a larger penalty is imposed; for parameters of lower importance, a smaller penalty is imposed. This means that the training process is encouraged to primarily adjust parameters that are not important to historical knowledge to fit new data, while maintaining relative stability for important historical parameters.

[0089] This implementation transforms the retraining process from a "peer review" to an "intelligent consolidation." While enabling the model to learn new data and adapt to changes, it consciously protects the model parameters that carry key historical memories, thus more finely balancing the contradiction between integrating new knowledge and retaining old knowledge, making the periodic historical review more stable and efficient.

[0090] In a preferred embodiment, in the railway line status assessment method, the preset weight coefficient is dynamically determined based on the current train operating speed, and the weight value of the vehicle dynamic sub-health index is configured to increase monotonically with the increase of operating speed.

[0091] The specific implementation of this invention optimizes the weighting coefficient setting rules for multi-source data fusion. In the basic fusion method, the weighting coefficients between the track state sub-health index and the vehicle dynamics sub-health index are usually preset fixed values. For example, the track state weight might be set to 0.7 and the vehicle dynamics weight to 0.3 based on experience, and these values ​​remain unchanged in all evaluation scenarios. The drawback of this fixed weighting is that it cannot adapt to varying operating conditions. When trains pass at low speeds, the geometric state of the track is often the main constraint on safety and comfort; however, when trains are running at high speeds, the dynamic interaction between the vehicle and the track becomes intense and critical, and the impact of the vehicle's dynamic response (such as vibration acceleration) on safety and stability increases significantly. Fixed weighting cannot reflect this evaluation focus that changes with speed, resulting in the need to improve the relevance and accuracy of the comprehensive evaluation results under different speed conditions.

[0092] To address the aforementioned issues, this embodiment of the invention replaces the fixed weighting coefficients with variables dynamically determined based on the real-time train speed. Specifically, during each fusion evaluation step, the method needs to synchronously acquire or receive the current train speed value. Based on this speed value, a predefined rule is used to determine the weighting value of the vehicle dynamic sub-health index.

[0093] The core feature of this rule is that the weights of the vehicle dynamic sub-health indicators are configured to increase monotonically with increasing train speed. This means that the higher the speed, the greater the weight of vehicle dynamic response data in the final comprehensive health indicator. The rationale for this weighting configuration is that at low speeds, track geometry has a significant impact on driving safety; at high speeds, wheel-rail dynamic interaction intensifies, making vehicle dynamic response a key factor in the assessment. Through dynamic adjustment of the weights, the comprehensive health indicator can better align with the safety assessment priorities at different speed levels. Simultaneously, the weights of the track condition sub-health indicators are correspondingly reduced to ensure that the total weight sums to 1.

[0094] This implementation method enables the generation of comprehensive health indicators to be scenario-adaptive. At low speeds, the assessment focuses more on track geometry indicators reflecting infrastructure status; at high speeds, the assessment automatically shifts towards vehicle response indicators reflecting the quality of dynamic system interaction. This allows the final assessment results to better align with actual safety concerns and performance shortcomings at different operating speeds, improving the rationality and practicality of the condition assessment and providing a more reliable basis for differentiated and precise operation and maintenance decisions.

[0095] In a preferred embodiment, the railway line status assessment method includes a data validity filtering step during the reinforcement learning period: comparing the status parameter values ​​in the current status monitoring data received in the data buffer pool with an absolute value safety threshold range obtained based on historical long-term monitoring data statistics; retaining only the data where all parameter values ​​are within their respective absolute value safety threshold ranges as candidate update data that can be directly used subsequently.

[0096] The specific implementation of this invention adds a data validity filtering step during the reinforcement learning period. In the basic reinforcement learning mode, to maximize learning efficiency, all model-based intelligent filtering constraints are temporarily removed, allowing data to flow in almost unimpeded. While this ensures that "normal" data reflecting the new state is learned efficiently, it also introduces a significant risk: after removing all constraints, extreme abnormal data that is clearly beyond physical possibilities, generated by sudden sensor failures, momentary communication errors, or extreme, accidental interference (such as strong impacts from non-train loads), will also enter the model update process without hindrance. This data is invalid or even harmful "noise," and if learned by the model, it will severely pollute the newly established healthy baseline, causing deviations in the learning direction and affecting the accuracy of the evaluation system after the reinforcement learning period ends.

[0097] To address the aforementioned issues, this embodiment of the invention establishes a simple yet necessary safety checkpoint at the front end of the data stream during the reinforcement learning period. This method does not involve complex model validation; instead, it presets a set of absolute value safety threshold ranges based on domain-specific physical common sense and long-term observational experience. The method for determining the absolute value safety threshold range is as follows: for each state parameter... s Its mean is calculated based on its long-term historical monitoring data (e.g., complete annual data covering different seasons and operating conditions). m s and standard deviation s s The upper limit of the aforementioned security threshold range. U abs,s and lower limit L abs,s It can be set as follows: Uabs,s = m s +k · s s , L abs,s = m s - k · s s in, k For safety, the value is typically between 4 and 6. This range is designed to intercept outliers that deviate significantly from physical probability, such as those caused by equipment malfunctions.

[0098] During the reinforcement learning period, whenever the data buffer receives new current state monitoring data, the method performs a data validity filtering step: each state parameter value in the data is compared with its corresponding preset absolute value safety threshold range. Only when all parameter values ​​in a data point are within their respective safety threshold ranges is the data allowed as candidate update data for subsequent direct use. If any parameter value exceeds its safety range, the entire data entry is filtered out.

[0099] This implementation method, while maintaining smooth data flow and promoting rapid learning, introduces a bottom-line protection based on physical extreme values ​​during the reinforcement learning period. It can effectively intercept obviously unreasonable and highly likely error-prone extreme data with extremely low computational cost, providing a basic data quality guarantee for the high-intensity, high-degree-of-freedom learning process and enhancing the robustness and security of this special operating mode. Example

[0100] To illustrate the invention more clearly, the following description uses a section of a high-speed railway with an operating speed of 200 km / h as an example.

[0101] 1. Data and Model Initialization: Track geometry (elevation, alignment, and level) and vehicle dynamic response (vertical / lateral acceleration) data for this section over the past year were collected to form the initial training set. A Gaussian mixture model was used to construct a dynamic health baseline model: the optimal components were determined to be 5 using the Bayesian information criterion, and the model was trained using the expectation-maximization algorithm. A probability density threshold p was set. th This ensures that 95% of the initial training data is identified as normal, and the corresponding probability density contour is the 95% confidence interval for each parameter.

[0102] 2. Auxiliary reference benchmark: Based on 3 years of historical data for the entire line, the 99th percentile (0.5% quantile to 99.5% quantile) of each parameter is calculated as an auxiliary reference benchmark.

[0103] 3. Real-time evaluation and updates: Real-time calculation of orbital state sub-indices I t and vehicle dynamics sub-indicator I v The current train speed is v = 180 km / h. The weighting function is used to calculate w. v If (v) = 0.38, then w t (v)=0.62, calculate the comprehensive health index I 综合 =0.62×I t +0.38×I v .

[0104] After initial screening by the auxiliary benchmark, if the comprehensive index value of the new data is within the current 95% confidence interval of the model, or exceeds the interval but is within the tolerance threshold δ (set as 5% of the interval width), it is marked as candidate update data.

[0105] Calculate the Mahalanobis distance D between the candidate data and the historical data. The first threshold D... th =3.0 was pre-determined through Monte Carlo simulation: 1000 subsets of the same size as the candidate batch were randomly selected from the initial training set, the Mahalanobis distance was calculated, and the 95th largest value (i.e., the 95th percentile) was taken. The sliding window length W=10 and the frequency threshold N=7 were set.

[0106] If the Mahalanobis distance D exceeds 3.0 7 times in the most recent 10 tests, a significant concept drift is determined to have occurred, triggering a baseline reset; otherwise, the Gaussian mixture model is incrementally learned using the online expectation-maximization algorithm on this batch of data.

[0107] 4. Periodic Historical Review: When the cumulative updated data reaches half of the initial training set (i.e., proportion ρ=50%), a review is triggered. An elastic weight consolidation strategy is used: Based on a subset of the initial dataset, the diagonal value of the Fisher information for each parameter is approximated as an importance weight using mini-batch gradient squared average. A regularization penalty term is added to the retraining loss function that merges old and new data to protect important historical parameters.

[0108] 5. Planned Change Handling: Rail replacement will be carried out in this section. The system receives the instruction and enters a period of T... 强化 =15-day reinforcement learning period. During this period, only absolute value safety thresholds are used (based on historical data, with a safety factor k=5, i.e., upper and lower limits are μ). s ±5σ s Filter out obviously erroneous data, and use the remaining data directly for rapid updates using the online expectation-maximization algorithm of the model. Resume the normal process after 15 days.

[0109] Through the above process, this method achieves continuous, adaptive and accurate evaluation of the state of the segment.

[0110] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily implemented by those skilled in the art. Therefore, the present invention is not limited to the specific details without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method for assessing the condition of a railway line, characterized in that, Includes the following steps: Acquire and preprocess the time-series status monitoring data of the section of the target railway line to be evaluated within the historical monitoring period to form the initial training dataset; Based on the initial training dataset, an unsupervised learning method is used to construct a dynamic health baseline model for the segment to be evaluated. The dynamic health baseline model is used to output the normal value range of the state parameters at a given confidence level. Continuously acquire current status monitoring data of the section to be evaluated, calculate real-time health indicators based on the dynamic health baseline model, and store the current status monitoring data and the corresponding real-time health indicators into a first-in-first-out data buffer pool. At the same time, real-time health indicators are compared with preset dynamic warning thresholds or their changing trends are analyzed to generate health status assessment results and warning information. Perform online filtering and concept drift detection on the data buffer pool: First, the data in the data buffer pool is initially screened based on an auxiliary reference benchmark to obtain preliminary selected data; The auxiliary reference benchmark is independent of the dynamic health baseline model; Subsequently, the initial data were validated a second time based on the dynamic health baseline model, and data whose real-time health indicators were within their normal confidence intervals were marked as candidate update data. The statistical distance between the data distribution characteristics of the candidate update data and the historical data distribution characteristics represented by the dynamic health baseline model is calculated. If the statistical distance exceeds a preset first threshold, a significant concept drift is determined to have occurred. If no significant concept drift is determined, the dynamic health baseline model is updated online using the candidate update data through an incremental learning algorithm. If a significant concept drift is determined, the baseline reset process is triggered, that is, the original dynamic health baseline model is retrained and replaced based on the new data after the drift occurs, and the dynamic warning threshold is reset simultaneously.

2. The railway line condition assessment method according to claim 1, characterized in that, The auxiliary reference benchmark is determined by calculating the statistical distribution of each state parameter based on all time-series state monitoring data collected within a preset historical period, and defining the preset high percentage range in the center of the statistical distribution as the auxiliary reference benchmark.

3. The railway line condition assessment method according to claim 2, characterized in that, In the secondary verification step, if the real-time health indicator value x corresponding to the initial data is not within the normal confidence interval [L, U], and the absolute difference between it and the nearest boundary is less than a preset tolerance threshold δ, then the initial data is still marked as candidate update data; where the absolute difference is defined as: Lx when x < L, and xU when x > U.

4. The railway line condition assessment method according to claim 1, characterized in that, The determination of significant conceptual drift is performed as follows: Record the statistical distances calculated in each online screening and concept drift detection step to form a statistical distance sequence; Set a sliding decision window of fixed length and a frequency threshold; When tracing back a sliding decision window based on the current statistical distance, if the number of times the statistical distance exceeds the first threshold within the window reaches or exceeds the frequency threshold, a significant concept drift is determined to have occurred.

5. The railway line condition assessment method according to claim 1, characterized in that, The steps of updating the dynamic health baseline model online using candidate update data and an incremental learning algorithm also include a periodic historical review step: When the amount of candidate update data used in online incremental updates reaches a preset proportion of the initial training dataset, a portion of samples is extracted from the initial training dataset and merged with the recently accumulated candidate update data to retrain the dynamic health baseline model with all parameters.

6. The railway line condition assessment method according to claim 1, characterized in that, Condition monitoring data includes track geometry irregularities and vehicle dynamic response data; The specific steps for calculating real-time health indicators based on the dynamic health baseline model include: Sub-index calculation: Based on track geometric irregularity data and vehicle dynamic response data, the corresponding track state sub-health index and vehicle dynamic sub-health index are calculated respectively. Fusion assessment: Based on preset weighting coefficients, the track status sub-health index and the vehicle dynamic sub-health index are weighted and averaged to generate a comprehensive health index as a real-time health index.

7. The railway line condition assessment method according to claim 1, characterized in that, It also includes learning mode management steps: When an external instruction is received indicating that the section to be evaluated has entered the planned state change phase, the system enters a reinforcement learning period of a preset duration. During the reinforcement learning period, all data screening and verification constraints in the online screening and concept drift detection steps are temporarily lifted, so that the current state monitoring data received in the data buffer pool can be directly used as candidate update data for reinforcement incremental learning of the dynamic health baseline model. After the reinforcement learning period ends, the system automatically resumes the regular online screening and concept drift detection steps.

8. The railway line condition assessment method according to claim 5, characterized in that, The retraining process in the periodic history review step is implemented using an elastic weight consolidation strategy, specifically including: Parameter importance assessment: Calculate the Fisher information diagonal values ​​of each parameter in the dynamic health baseline model as the importance weights of each parameter; Constrained retraining: A regularization penalty term is added to the loss function during retraining. The value of this term is positively correlated with the importance weights of each parameter and their offsets relative to their values ​​before this retraining.

9. The railway line condition assessment method according to claim 6, characterized in that, The preset weighting coefficients are dynamically determined based on the current train speed, and the weighting values ​​of the vehicle dynamic sub-health indicators are configured to increase monotonically with the increase of the operating speed.

10. The railway line condition assessment method according to claim 7, characterized in that, During the reinforcement learning period, a data validation filtering step is implemented: The values ​​of each state parameter in the current state monitoring data received in the data buffer pool are compared with an absolute value safety threshold range obtained based on historical long-term monitoring data statistics. Only data where all parameter values ​​are within their respective absolute value safety threshold ranges are retained as candidate update data that can be directly used later.

Images