Rail profile wear prediction method and system based on multi-source data fusion

Abstract

The present application belongs to the field of railway rail detection and intelligent operation and maintenance technology, and discloses a rail profile wear prediction method and system based on multi-source data fusion. The method collects multi-source data including rail profile data, train operation data, environment data and maintenance record data; the multi-source data is weighted and fused according to the data type to obtain a fusion feature data set; the fusion feature data set is input into a locally deployed time series neural network model to output the wear trend prediction result of the rail profile in the future period of time; the differential early warning threshold is set according to the rail type and operation life, and the early warning is triggered when the prediction result exceeds the threshold; the fusion feature data set, model weight and prediction result are synchronously stored in the local storage and cloud storage, and the prediction is independently completed based on the local storage when the network is interrupted. The present application is improved in the three dimensions of prediction accuracy, field availability and early warning pertinence.

Description

Technical Field

[0001] This invention relates to the field of railway rail inspection and prediction technology, specifically to a method and system for predicting rail profile wear based on multi-source data fusion. Background Technology

[0002] During long-term service, rails gradually experience railhead wear and profile deformation due to the combined effects of train loads, track geometry, environmental conditions, and maintenance procedures. Failure to predict wear trends in a timely manner will directly impact train operation safety and the scientific validity of track maintenance plans.

[0003] Existing methods and equipment for detecting rail wear have several shortcomings. First, existing rail profile measurement equipment mostly focuses on the static identification of the current rail profile, only outputting the current wear status. It typically cannot provide quantitative predictions about how wear will develop in the future or when it might reach maintenance limits. Second, most existing wear prediction methods are based on single types of data, such as relying solely on historical profile data or extrapolating from train passage counts. They fail to consider the coupled effects of train load, environmental temperature, humidity, wind speed, and maintenance operations, limiting prediction accuracy. Furthermore, the integration of rail profile data with other relevant data is low, resulting in insufficient data utilization and difficulty in supporting accurate predictions. Third, existing prediction devices rely heavily on cloud-based data processing platforms for centralized computation, lacking independent reasoning capabilities on-site. When the network is interrupted or the signal is poor, on-site inspectors cannot obtain real-time wear predictions and maintenance guidance. Additionally, some existing prediction models are complex in structure, consume significant computational resources, and are not easily deployed on embedded terminals in the field. Fourth, existing early warning systems mostly use uniform thresholds or coarse-grained limit comparisons, without classifying thresholds according to different factors such as rail type, service life, and track curve conditions, resulting in insufficient targeting and operability of early warnings. At the same time, existing early warning systems mostly operate in a single-level mode of "alarm when the limit is reached," lacking the linkage between graded prompts and targeted maintenance suggestions, making it difficult to support the track maintenance department in making graded responses and maintenance plans according to the wear evolution stages.

[0004] To address the aforementioned issues, there is an urgent need for a method and system that can predict rail profile wear trends in real time and stably on-site, and issue early warnings based on differences in rail type and service life. This would enable accurate prediction of wear trends, independent output of prediction results on-site even when network quality is poor or interrupted, and targeted maintenance decision support based on differentiated thresholds. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a rail profile wear prediction method and system based on multi-source data fusion. It comprehensively utilizes multi-source information such as rail profile data, train operation data, environmental data, and maintenance record data to achieve accurate prediction and differentiated early warning of rail profile wear trends on site. It can also independently complete the prediction based on local storage even when the network is interrupted, thus solving the problems of insufficient prediction accuracy, low data utilization, lack of real-time performance and availability during network outages, and insufficient targeted early warning in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution provided by this invention includes: A rail profile wear prediction method based on multi-source data fusion includes: Collect multi-source data, including rail profile data, train operation data, environmental data, and maintenance record data; The multi-source data are weighted according to data type and then fused to obtain a fused feature dataset. The fused feature dataset is input into a locally deployed temporal neural network model, which outputs a prediction result of the wear trend of the rail profile over a future period. Differentiated warning thresholds are set according to rail type and service life, and a warning is triggered when the wear trend prediction result exceeds the warning threshold; The fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results are synchronously stored in local storage and cloud storage; in the event of a network interruption, the wear trend prediction results are output independently based on the weights stored in the local storage.

[0007] Preferably, in the weighted fusion, the weight of the rail profile data is greater than the weight of the train operation data, the weight of the train operation data is greater than the weight of the environmental data and the weight of the maintenance record data, and the weight of the environmental data is equal to the weight of the maintenance record data.

[0008] Preferably, the method further includes: preprocessing the multi-source data before the weighted fusion, wherein the weighted fusion is performed on the preprocessed multi-source data; the preprocessing includes: removing data noise by means filtering; completing missing data by linear interpolation; mapping the data to a preset interval by normalization; and removing abnormal data according to the 3σ criterion.

[0009] Preferably, the temporal neural network model is an improved LSTM neural network model, which includes: a time attention mechanism applied to the hidden layer output of the improved LSTM neural network model; and a multi-feature split-encoding structure for split-encoding the rail profile data, the train operation data, the environmental data, and the maintenance record data before fusion.

[0010] Preferably, the improved LSTM neural network model further includes inter-layer residual connections; the improved LSTM neural network model is a lightweight structure that has undergone pruning and quantization; the training of the improved LSTM neural network model adopts an adaptive learning rate and an early stopping strategy.

[0011] Preferably, the setting of the warning threshold includes: establishing a benchmark limit based on the railway line wear and repair industry standard; establishing a graded threshold library according to rail type; correcting the graded threshold library according to the line curve radius, operating years and transportation intensity; statistically calibrating the corrected threshold based on historical wear and failure data; and dividing the calibrated threshold into three levels of warning: yellow warning, orange warning and red warning, according to the proportion of the benchmark limit.

[0012] Preferably, when the warning is triggered, it also includes pushing targeted maintenance suggestions, which include at least one of rail grinding, weld repair and rail replacement.

[0013] Preferably, when the network is restored, the local storage and the cloud storage perform bidirectional synchronization on the weights of the temporal neural network model, the fused feature dataset, and the wear trend prediction results.

[0014] This invention also discloses a rail profile wear prediction system based on multi-source data fusion, comprising: The data acquisition module is used to collect rail profile data, train operation data, environmental data, and maintenance record data as multi-source data; The data fusion module is used to preprocess the multi-source data and perform weighted fusion according to data type to obtain a fused feature dataset; The wear prediction module is used to input the fused feature dataset into a locally deployed temporal neural network model and output the wear trend prediction result of the rail profile over a future period of time. The early warning module is used to set differentiated early warning thresholds according to rail type and service life, and to trigger an early warning when the wear trend prediction result exceeds the early warning threshold; The storage module includes a local storage unit and a cloud storage unit, used to store the fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results, so that the wear prediction module can independently output the wear trend prediction results based on the weights in the local storage unit when the network is interrupted. The interactive module is used to display the wear trend prediction results and early warning information, and to receive maintenance data and early warning threshold adjustment instructions entered by the operator.

[0015] Preferably, the data acquisition module includes a profile acquisition unit, which uses a combination of laser acquisition and visual acquisition methods to acquire the rail profile data.

[0016] Beneficial effects 1. This invention uses four types of multi-source data—rail profile data, train operation data, environmental data, and maintenance record data—as input, and performs weighted fusion according to data type. The weighting method of having the highest weight for profile data, followed by train operation data, and then having environmental data and maintenance record data in parallel, is consistent with the rail wear mechanism, enabling the fused feature dataset to comprehensively cover the key influencing factors of wear. On this basis, a locally deployed temporal neural network is used to model the fused feature dataset. Compared with existing methods that extrapolate from single data or use only static recognition, this method can significantly improve the prediction accuracy of wear trends in the future.

[0017] 2. This invention deploys the temporal neural network model on the local side of the device and combines it with the lightweight structure of the improved LSTM neural network model, enabling the field device to complete prediction and inference without relying on the remote platform for real-time scheduling. The single prediction response is rapid and adaptable to the computing power conditions of the field embedded terminal.

[0018] 3. This invention integrates the feature dataset, the weights of the temporal neural network model, and the wear trend prediction results, and stores them simultaneously in local storage and cloud storage. When the network is interrupted, the device can independently output the wear trend prediction results based on the weights in local storage, and can continue to work without relying on the cloud. After the network is restored, bidirectional synchronization can be performed between local storage and cloud storage to ensure that the data accumulated by the device during continuous operation is consistent with other sites. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of a rail profile wear prediction method based on multi-source data fusion according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of a rail profile wear prediction system based on multi-source data fusion according to an embodiment of the present invention. Figure 3This is a schematic diagram of the internal architecture of an improved LSTM neural network model according to an embodiment of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. Adjustments and substitutions made by those skilled in the art to the following embodiments according to actual needs without departing from the principles and spirit of the present invention shall all fall within the scope of protection of the present invention.

[0021] Example 1 like Figure 1 As shown, this embodiment uses a vehicle-mounted device as a typical example to specifically illustrate the rail profile wear prediction method based on multi-source data fusion of the present invention. It should be understood that the present invention is also applicable to other device forms such as handheld portable and trackside fixed devices. The method of this embodiment includes steps S1 to S8, which will be elaborated one by one below.

[0022] In one specific embodiment, this method is applied to an operating section of a main line of high-speed trains. The train type on the line is high-speed trains, with an average operating speed of about 250 km / h, an average load weight of about 800 t, an average number of passes per day of about 30, an on-site temperature of about 25℃, a humidity of about 60%RH, and a wind speed of about 3 m / s. The method is illustrated using a section of the line with 60 kg / m rails and an operating life of about 3 years as an example.

[0023] Step S1: Collect multi-source data, including rail profile data, train operation data, environmental data, and maintenance record data.

[0024] Rail profile data includes the geometric contours of the rail cross-sections at discrete sampling points along the longitudinal direction of the track and their derived values, specifically including features such as rail head width, rail top radius, vertical wear, and lateral wear. Existing historical profile data can be retrieved simultaneously during data collection to form a historical sequence for subsequent time-series modeling. Train operation data includes parameters affecting wear excitation, such as train type, speed, load weight, and number of passes, which can be obtained from the railway operation system via a wireless communication interface. Environmental data includes parameters affecting rail material properties and the wear process, such as on-site temperature, humidity, and wind speed. Maintenance record data includes historical maintenance operation records such as rail grinding, weld repair, and rail replacement, which can be entered by operators on-site through an interactive interface or imported from existing maintenance information systems.

[0025] In one specific embodiment, the collected current rail profile data includes typical values ​​such as vertical wear of 0.5 mm and lateral wear of 0.6 mm, and simultaneously calls historical profile data from the past 12 months for time-series modeling; train operation data retrieves section operation records from the past 3 months; environmental data is obtained from real-time on-site measurements; maintenance record data includes maintenance operation records from the past 6 months, such as "two months ago, a rail grinding operation was performed on this section, with a grinding accuracy of approximately 0.05 mm".

[0026] To address the shortcomings of existing technologies that only collect single types of data and fail to cover all factors influencing wear, thus hindering accurate prediction, this step simultaneously collects four types of data: rail profile, train operation, environment, and maintenance records. This provides a complete data foundation covering wear mechanisms for subsequent weighted fusion and time-series modeling, improving prediction accuracy from the data source.

[0027] It should be understood that the specific methods for collecting the aforementioned data are not unique. Rail profile data can be obtained using laser scanning, machine vision, or a combination of both; train operation data can be obtained through railway operation system interfaces, onboard information systems, or trackside detection devices; environmental data can be measured by temperature, humidity, and wind sensors integrated into the device itself, or obtained from railway meteorological information systems; maintenance record data can be entered on-site or synchronized online through maintenance information systems. The specific data collection methods and communication interfaces under different implementation forms can be flexibly selected by those skilled in the art according to the device form.

[0028] Step S2: Preprocess the multi-source data.

[0029] The collected multi-source data is preprocessed to improve data quality and lay the foundation for weighted fusion and subsequent time series modeling.

[0030] In some preferred embodiments, the preprocessing includes: removing data noise using mean filtering; completing missing data using linear interpolation; mapping the data to a preset interval using normalization; and removing outlier data according to the 3σ criterion.

[0031] To address the shortcomings of existing technologies where multi-source data is affected by sensor noise, sampling discontinuities, dimensional differences, and individual extreme values, resulting in compromised fusion and modeling effects, this step applies four processing steps—mean filtering, linear interpolation, normalization, and the 3σ criterion—to the multi-source data in sequence. This ensures that the data fed into the fusion process are consistent in scale, completeness, and reliability. Compared to simply cleaning the data, this step significantly improves the usability and stability of the fused feature dataset.

[0032] Specifically, mean filtering is used to suppress high-frequency noise in profile and environmental data. In one specific embodiment, the sliding window length of the mean filter is 5, meaning the arithmetic mean of 5 consecutive samples replaces the original value of the intermediate point. Linear interpolation is used to complete missing samples caused by communication interruptions or packet loss. In one specific embodiment, if a train operation record is missing, it is completed using the linear interpolation results of two adjacent records. Normalization is used to map multi-source data with different dimensions to a unified interval to eliminate the impact of dimensional differences on weighted fusion. In one specific embodiment, the data is mapped to the interval 0 to 1, calculated using the following formula: ; Where x is the original data sample value before normalization. It is the minimum value in the sample set of the corresponding data type. Let x' be the maximum value in the sample set corresponding to the data type, and x' be the normalized data sample value mapped to the interval between 0 and 1. The 3σ criterion is used to remove individual outliers that deviate from the overall trend, and the criterion is defined as follows: like If so, sample x is determined to be an outlier and is removed. Where x is the sample value to be judged, μ is the mean of the sample set of the corresponding data type, and σ is the standard deviation of the sample set of the corresponding data type. In a specific embodiment, if a profile data point is identified by the 3σ criterion as significantly deviating from the overall trend, that sample is removed before participating in subsequent fusion.

[0033] It should be understood that the window length of the mean filter, the normalization mapping interval, and the statistical criteria for anomaly removal can be adjusted accordingly in other embodiments. For example, normalization can use other mapping intervals or standardization processing, and anomaly removal can be combined with other statistical criteria such as box plots, all without departing from the technical concept of the present invention.

[0034] Step S3: Assign weights to the preprocessed multi-source data according to the data type and perform weighted fusion.

[0035] The preprocessed multi-source data are weighted according to data type and then fused to obtain a fused feature dataset.

[0036] In some preferred embodiments, in the weighted fusion, the weight of the rail profile data is greater than the weight of the train operation data, the weight of the train operation data is greater than the weight of the environmental data and the weight of the maintenance record data, and the weight of the environmental data is equal to the weight of the maintenance record data.

[0037] To address the shortcomings of existing technologies that often involve concatenating or averaging multi-source data by simply listing table fields without distinguishing the different contributions of various data types to wear prediction, resulting in insufficient data utilization, this step applies differentiated weights to four types of data based on the wear mechanism: Profile data directly reflects the wear state and historical development trend of rails, serving as the core basis for wear prediction, hence its highest weight; train operating load and passing frequency are the main excitation sources of wear, playing a dominant role in the wear rate, and thus have the second highest weight; environmental conditions affect rail material properties and the wear process, while maintenance operations alter the wear baseline and subsequent development trend; both are important influencing factors, therefore they are assigned equal auxiliary weights. Compared to equal-weighted or indiscriminate fusion methods, the weighted fusion method in this step ensures that the contributions of each component of the fused feature dataset match the wear mechanism, thereby improving the accuracy of subsequent predictions.

[0038] In one specific embodiment, the weight of rail profile data is 0.4, the weight of train operation data is 0.3, and the weights of environmental data and maintenance record data are 0.15 each. The fused feature dataset is calculated using the following formula: ; Where w1=0.4, w2=0.3, w3=0.15, w4=0.15. Where F is the fused feature dataset. This is the preprocessed rail profile data. For preprocessed train operation data, For preprocessed environmental data, The data represents the pre-processed maintenance records; w1, w2, w3, and w4 are the weights corresponding to the four types of data mentioned above. These weight values ​​were determined through wear mechanism analysis, field data statistics, and verification by multiple sets of comparative tests, and are used to reflect the primary and secondary relationship between profile as the core basis for wear prediction, train load as the main excitation source, and environment and maintenance as auxiliary influencing factors.

[0039] It should be understood that, while maintaining the above relative relationship, other weight combinations can also be used in other embodiments; for example, they can be allocated in combinations such as 0.45 / 0.25 / 0.15 / 0.15 or 0.5 / 0.3 / 0.1 / 0.1, or new weights can be calibrated according to the wear pattern of a specific line, all without departing from the technical concept of the present invention.

[0040] Step S4: Input the fused feature dataset into the locally deployed temporal neural network model and output the wear trend prediction result.

[0041] The fused feature dataset is input into a locally deployed temporal neural network model, which outputs a prediction result of the wear trend of the rail profile over a future period.

[0042] In some preferred embodiments, the temporal neural network model is an improved LSTM neural network model, which includes: a time attention mechanism applied to the hidden layer output of the improved LSTM neural network model; and a multi-feature split-encoding structure for split-encoding the rail profile data, the train operation data, the environmental data, and the maintenance record data before fusion.

[0043] To address the shortcomings of existing technologies that primarily employ static identification or traditional machine learning methods for wear prediction, which fail to fully utilize the time dependence of wear evolution and differentiate between various types of data with different physical meanings, this step utilizes an improved LSTM neural network model. This model strengthens the time steps highly relevant to wear prediction through a time attention mechanism and avoids mutual interference between data with different physical meanings in early layers through a multi-feature split-path encoding structure. This enables the model to simultaneously possess the ability to distinguish between the time dimension and the feature dimension, significantly improving prediction accuracy and long-term trend stability compared to traditional methods.

[0044] like Figure 3 As shown, the overall architecture of the improved LSTM neural network model in a specific embodiment includes: a multi-feature split-path encoding structure at the input end encodes four types of data—rail profile, train operation, environment, and maintenance records—into independent embedding sub-networks; the outputs of each split-path encoding are concatenated and fused in a shared layer; the concatenated sequence is fed into a shared LSTM backbone; a time attention mechanism is superimposed on the LSTM backbone to perform weighted summation of the hidden layer outputs at each time step; finally, the wear trend prediction result for a future period is obtained through the output layer.

[0045] In one specific embodiment, the temporal attention mechanism calculates the attention weights for each time step and performs a weighted summation using the following formula: ; ; ; in, This represents the hidden layer output of the improved LSTM neural network model at time step t. and Learnable parameters for scoring attention. Let be the score at time step t. Let be the attention weight at the t-th time step after normalization, T be the total number of time steps, and c be the output vector after weighted summation of time attention. Through this weighted summation method, the model can adaptively focus on time steps highly correlated with wear prediction, such as recent profile change trends.

[0046] In some preferred embodiments, the improved LSTM neural network model further includes inter-layer residual connections; the improved LSTM neural network model is a lightweight structure that has undergone pruning and quantization; and the training of the improved LSTM neural network model employs an adaptive learning rate and an early stopping strategy.

[0047] To address the shortcomings of existing temporal neural networks in rail wear prediction scenarios—namely, vanishing gradients in deep networks, large model size making deployment difficult on embedded terminals, and susceptibility to overfitting—this step further enhances the aforementioned improved LSTM model with three underlying algorithmic improvements: Interlayer residual connections allow information to flow directly between layers of the deep LSTM via additive pathways, mitigating vanishing gradients and improving model fitting ability; pruning and quantization significantly compress the model size and reduce computational requirements by removing connections with near-zero absolute weights and performing low-bit quantization on parameters; and adaptive learning rate and early stopping strategies dynamically adjust the learning rate based on the validation set loss during training and terminate training early if the validation set loss does not decrease for several consecutive rounds, preventing overfitting and improving generalization. These three improvements, combined with the top-level temporal attention mechanism and multi-feature split-path encoding structure, simultaneously improve prediction accuracy, lightweight deployment, and training generalization, enabling a single model to meet the requirements of real-time performance, accuracy, and stability in the field.

[0048] In one specific embodiment, the improved LSTM neural network model has 12 neurons in the input layer, 3 hidden layers with 64 neurons per layer, and 1 neuron in the output layer. During training, the Adam optimizer is used with an initial learning rate of 0.001 and a maximum of 1000 iterations. The historical fusion feature dataset from the past 12 months is divided into training and testing sets in a 7:3 ratio. This division ensures sufficient training samples to fit long-range dependencies while retaining ample test samples for generalization validation. An adaptive learning rate and early stopping strategy are employed during training. After training, structural pruning and parameter quantization are performed to obtain a lightweight deployment model. Under the above parameter configuration, the improved LSTM neural network model achieves a test set validation accuracy of 96.2% after training to convergence. When deployed on an embedded terminal, the model performs a wear trend prediction for the next 1 to 6 months based on the current fusion feature dataset, with a single prediction taking less than 5 seconds. Typical prediction outputs are: vertical wear of approximately 0.6 mm and lateral wear of approximately 0.7 mm after 1 month; vertical wear of approximately 0.75 mm and lateral wear of approximately 0.85 mm after 3 months; and vertical wear of approximately 0.95 mm and lateral wear of approximately 1.05 mm after 6 months. Based on these, a wear trend curve is generated to determine the wear rate and the expected time to reach the warning threshold.

[0049] It should be understood that the above-mentioned parameters such as input dimension, number of hidden layers and neurons, optimizer type, learning rate and number of iterations are only one preferred example; in other embodiments, the network size and training hyperparameters can be adjusted according to the on-site computing power and training data volume, and other temporal modeling structures such as improved LSTM and temporal convolutional networks can also be used in combination, all without departing from the technical concept of the present invention.

[0050] Step S5: Set differentiated early warning thresholds based on rail type and service life, and trigger early warnings.

[0051] Different warning thresholds are set according to rail type and service life, and a warning is triggered when the wear trend prediction result exceeds the warning threshold.

[0052] In some preferred embodiments, the setting of the warning threshold includes: establishing a benchmark limit based on the railway line wear and repair industry standard; establishing a graded threshold library according to rail type; correcting the graded threshold library according to the line curve radius, operating years and transportation intensity; statistically calibrating the corrected threshold based on historical wear and failure data; and dividing the calibrated threshold into three levels of warning: yellow warning, orange warning and red warning, according to the proportion of the benchmark limit.

[0053] To address the shortcomings of existing technologies that use a uniform threshold or coarse-grained comparison without considering differences in rail type, service life, track curves, and transport intensity, resulting in insufficient targeted early warning, this paper proposes a five-step threshold determination method. This method consists of five sequential steps: establishing a benchmark limit, constructing a tiered threshold library, correcting for on-site conditions, calibrating historical fault statistics, and classifying the warning into three levels. Compared to existing methods that only use a uniform threshold, this method can provide differentiated and tiered warning signals for different rail types, service lives, and track conditions, significantly improving the targetedness and operability of the early warning.

[0054] Specifically, the benchmark limit establishment stage, based on industry standards for railway line wear repair, determines the permissible limits for vertical and lateral rail wear under different scenarios, serving as the basis for subsequent threshold settings. The tiered threshold library construction stage establishes a tiered benchmark threshold library according to rail type to accommodate differentiated thresholds for different rail types such as 43 kg / m, 50 kg / m, 60 kg / m, and 75 kg / m. The field condition correction stage adjusts the tiered threshold library using curve radius (small-radius curves have relatively higher wear rates), service life (older lines have relatively stronger cumulative wear effects), and transport intensity (heavy-load or high-density lines accelerate wear) as correction factors. The historical fault statistics calibration stage statistically calibrates the corrected thresholds based on fault and accident data caused by historical wear on this line or adjacent lines, making the threshold settings more reasonable.

[0055] The three-level early warning classification process divides the calibrated thresholds into three levels of early warning based on the proportion of the baseline limit. In a specific embodiment, the hierarchical triggering threshold relationship is as follows: ; ; ; The triggering rule is: when ≤ < When a yellow alert is triggered, it prompts you to plan maintenance in advance; when ≤ < When an orange alert is triggered, an immediate response is required; when ≥ At that time, a red alert was triggered, and emergency response was initiated. Among these, These are the baseline limits obtained after correction for on-site conditions and statistical calibration based on historical fault data. , , The trigger thresholds for yellow, orange, and red alerts are respectively. This is the wear trend prediction value output by the improved LSTM neural network model.

[0056] In a specific embodiment, taking a line section with 60 kg / m rails and an operating life of 3 years as an example, the aforementioned 5-step determination method yields a vertical wear baseline limit of 1.0 mm and a lateral wear baseline limit of 1.2 mm. The vertical wear warning threshold generally falls within the range of 0.8 to 1.2 mm, and the lateral wear warning threshold generally falls within the range of 1.0 to 1.5 mm. Comparing the predicted vertical wear value of 0.95 mm for the next 6 months in step S4 with the aforementioned baseline limit, the predicted value reaches approximately 95% of the baseline limit, falling into the yellow warning trigger range, and the audible and visual alarm unit issues a yellow warning.

[0057] In some preferred embodiments, when the warning is triggered, it also includes pushing targeted maintenance recommendations, which include at least one of rail grinding, weld repair and rail replacement.

[0058] To address the shortcomings of existing early warning systems that are disconnected from maintenance decisions and only issue alarm signals without providing specific maintenance recommendations, this step provides targeted maintenance recommendations to the site simultaneously with the triggering of the early warning. The types of maintenance recommendations are stratified according to the nature and magnitude of the predicted wear: when the predicted wear is mainly characterized by changes in the rail head profile and is within the yellow warning zone, it is recommended to restore the profile through rail grinding; when there is localized accelerated wear at the weld joints or a trend towards weld-related faults in the predicted section, it is recommended to perform weld repair; when the predicted wear is close to or exceeds the red warning trigger threshold and conventional maintenance is insufficient to restore the condition, it is recommended to replace the rail. Taking the aforementioned 60 kg / m rail as an example, when the predicted vertical wear value reaches approximately 95% of the baseline limit after 6 months, the maintenance recommendation sent along with the yellow warning is "It is recommended to perform rail grinding on this section within the next 4 to 5 months."

[0059] It should be understood that the specific types, triggering conditions, and timing of maintenance recommendations can be adjusted accordingly based on the maintenance processes and maintenance plan management methods commonly used in the field; in other embodiments, multiple maintenance recommendations can also be combined and pushed, all without departing from the technical concept of the present invention.

[0060] Step S6: Synchronously store the fused feature dataset, model weights, and prediction results to local storage and cloud storage; output prediction results independently based on local storage when the network is interrupted.

[0061] The fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results are synchronously stored in local storage and cloud storage; in the event of a network interruption, the wear trend prediction results are output independently based on the weights stored in the local storage.

[0062] To address the shortcomings of existing prediction devices that rely heavily on cloud-based data processing platforms, lack independent inference capabilities on-site, and are unable to provide wear prediction results in the event of a network outage, this step constructs a local / cloud dual-storage bidirectional synchronization architecture. When the network is available, local and cloud storage work together to synchronously persist the fused feature dataset, model weights, and wear trend prediction results on both the device's local side and the cloud side. During a network outage, the device continues to perform inference on the currently collected and subsequently acquired fused feature datasets based on the model weight snapshot in local storage, independently outputting wear trend prediction results. After the network is restored, consistency is restored between the two storage systems according to the bidirectional synchronization process. Compared to existing methods that solely rely on the cloud, this step significantly improves the device's availability in unstable network environments along railway lines.

[0063] In some preferred embodiments, upon network recovery, the local storage and the cloud storage perform bidirectional synchronization of the weights of the temporal neural network model, the fused feature dataset, and the wear trend prediction results.

[0064] To address the data inconsistency risks inherent in unidirectional cloud downloads or local uploads, this step performs bidirectional synchronization upon network recovery: The fused feature dataset accumulated during the network outage and the newly generated wear trend prediction results are uploaded to cloud storage. Simultaneously, the latest model weight snapshot is retrieved from cloud storage, and the local model weights are version-verified to update the local model. Compared to unidirectional synchronization, this step simultaneously facilitates the upward aggregation of newly added local data during the network outage and the downward deployment of the latest cloud model to the local system, forming a consistent, traceable, and rollback-capable bidirectional data flow.

[0065] In one specific embodiment, when a network interruption occurs in the segment where the device is located, the device continues to perform multiple wear trend predictions based on the model weights in the local storage unit, and each prediction can be completed within 5 seconds. After the network is restored, the local storage unit and the cloud storage unit are synchronized according to the aforementioned bidirectional synchronization process: the newly generated fusion feature dataset and prediction results during the network outage are uploaded to the cloud storage unit as backup and cross-site aggregation, and the latest improved LSTM model weight snapshot in the cloud storage unit is sent to the local storage unit, and the local side completes version verification and model update accordingly.

[0066] It should be understood that the specific triggering time for bidirectional synchronization can be the instant of network recovery or a timed synchronization after network recovery; version verification can be implemented in various ways such as weighted hashing, version number or timestamp; in other embodiments, conflict resolution strategies can also be added to handle the situation where the same object has new versions on both sides, all without departing from the technical concept of the present invention.

[0067] Step S7: Display the prediction results and early warning information, and receive maintenance data and threshold adjustment instructions.

[0068] The wear trend prediction results, early warning information, and the aforementioned maintenance suggestions are displayed to the operator through the interactive interface; at the same time, the maintenance record data and early warning threshold adjustment instructions entered by the operator are received through the interactive interface and written back to the corresponding data link - the maintenance record data is written back to the maintenance record data collection entry in step S1 to participate in the weighted fusion of subsequent cycles; the early warning threshold adjustment instructions are written back to the threshold setting entry in step S5 for manual correction of the graded threshold library.

[0069] To address the shortcomings of existing devices, such as limited interactive information and fragmented maintenance data acquisition and early warning threshold adjustment links, this step centrally displays wear trend curves, three-level early warning information, and maintenance suggestions on the same interactive interface. It also integrates maintenance data entry and threshold adjustment into the interactive interface, forming a closed loop of "display-entry-writeback". Compared to methods that only display without interaction or only interact without writeback, this step enables frontline maintenance personnel to complete information acquisition and feedback with the shortest operation path, improving on-site operation efficiency.

[0070] It should be understood that the specific form of the interactive interface in other embodiments may be a touch screen, physical buttons, or a combination of both, or it may be presented through remote access via a mobile terminal, without departing from the technical concept of the present invention.

[0071] In some preferred embodiments, step S8 is also included: periodic retraining.

[0072] Repeat steps S1 to S7 at preset intervals to update multi-source data, retrain the improved LSTM neural network model, and optimize prediction accuracy.

[0073] To address the shortcomings of existing models whose parameters remain unchanged after training and are unable to adapt to the dynamic evolution of rail wear with seasonal changes, traffic volume, and maintenance, this step involves retraining the model at fixed intervals. This allows the model weights to evolve with the latest fused feature dataset, ensuring that the wear prediction results consistently match the actual trends on the ground. Compared to static model methods, this step can maintain prediction accuracy over long-term operation.

[0074] In one specific embodiment, the interval between periodic retraining is 1 month; every 1 month, the device automatically re-executes steps S1 to S7, adopts the aforementioned adaptive learning rate and early stopping strategy during the retraining process, uses the old model weights as the initial values ​​for fine-tuning training, and performs pruning and quantization processing again after training to obtain the deployment model.

[0075] It should be understood that the retraining cycle in other embodiments can be adjusted accordingly based on the rate of change of line traffic volume and wear. For example, it can be shortened to half a month for heavy-load or high-density trunk lines, and extended to 2 to 3 months for lines with lower traffic volume, without departing from the technical concept of the present invention.

[0076] Example 2 like Figure 2 As shown, this embodiment provides a rail profile wear prediction system based on multi-source data fusion. As the hardware configuration for executing the method described in Method Embodiment 1, it includes a data acquisition module, a data fusion module, a wear prediction module, an early warning module, a storage module, and an interaction module. These six modules are cascaded sequentially according to the data flow. The core algorithm principle, parameter value basis, and alternative solutions of the system have been detailed in Method Embodiment 1 and will not be elaborated here. Only the hardware configuration and structural relationship of each module will be continuously described.

[0077] The data acquisition module is used to collect rail profile data, train operation data, environmental data, and maintenance record data as multi-source data. It consists of a profile acquisition unit, a train operation data acquisition unit, an environmental data acquisition unit, and a maintenance data acquisition unit. The profile acquisition unit uses a combination of laser and visual acquisition methods to obtain rail profile data. In one specific embodiment, it is equipped with a 660 nm wavelength line laser emitter and a 12-megapixel industrial camera, achieving an acquisition accuracy of ±0.05 mm, and can simultaneously retrieve historical profile data from the past 12 months. The train operation data acquisition unit connects to the railway operation system via a 4G or 5G wireless communication interface to acquire data such as train type, operating speed, load weight, and number of passes. The environmental data acquisition unit is equipped with a temperature sensor (range -40 to 85℃, accuracy ±0.5℃), a humidity sensor (range 0 to 100%RH, accuracy ±2%RH), and a wind speed sensor (range 0 to 30 m / s, accuracy ±0.1). The data acquisition unit (m / s) is used to collect on-site temperature, humidity, and wind speed data. The maintenance data acquisition unit receives maintenance records such as rail grinding, weld repair, and rail replacement data entered by operators through an interactive interface. It should be understood that the profile acquisition unit combines laser and visual acquisition methods. This leverages the complementary characteristics of laser measurement (high distance and profile accuracy) and visual acquisition (rich surface information), allowing for stable acquisition of rail profile data even under conditions of ambient light and dirt. In other embodiments, the profile acquisition unit may use only laser or only visual methods, without departing from the technical concept of this invention.

[0078] The data fusion module is used to preprocess multi-source data and perform weighted fusion according to data type to obtain a fused feature dataset. It consists of a data preprocessing unit and a multi-source data fusion unit. The data preprocessing unit is responsible for denoising, completing, normalizing, and removing anomalies from the collected multi-source data. The specific principles and parameter selection basis of the mean filtering, linear interpolation, normalization, and 3σ criterion used have been detailed in step S2 of the method embodiment and will not be elaborated here. The multi-source data fusion unit performs weighted fusion according to the aforementioned weight relationship. The corresponding weighted fusion principle, weight value basis, and formula and formula symbol explanation have been detailed in step S3 of the method embodiment and will not be elaborated here.

[0079] The wear prediction module is used to input the fused feature dataset into a locally deployed temporal neural network model and output the wear trend prediction result of the rail profile over a future period. It consists of a model training unit and a prediction unit. The model training unit trains and retrains the improved LSTM neural network model based on historical fused feature datasets, while the prediction unit performs inference on the current fused feature dataset based on the trained model. The overall architecture of the improved LSTM neural network model, its temporal attention mechanism, multi-feature split-path encoding structure, inter-layer residual connections, pruning and quantization, adaptive learning rate, and early stopping strategy, along with their technical principles and stacking order, have been detailed in step S4 of the method embodiment and will not be elaborated here. The wear prediction module uses an embedded computing unit to carry out the inference process, enabling the device to independently complete wear trend prediction on-site.

[0080] The early warning module is used to set differentiated early warning thresholds according to rail type and service life. An early warning is triggered when the wear trend prediction exceeds the early warning threshold. It consists of a threshold setting unit and an audible and visual alarm unit. The threshold setting unit implements the aforementioned five-step threshold determination method and three-level early warning classification. Its specific determination process, the relationship between the classification thresholds, and typical values ​​have been detailed in step S5 of the method embodiment and will not be elaborated here. The audible and visual alarm unit simultaneously emits sound and light to transmit early warning signals to on-site personnel. In one specific embodiment, the alarm light brightness is not less than 200 cd, and the volume is not less than 80 dB. Different frequencies, rhythms, and light color combinations are used to distinguish the three levels of warning (yellow, orange, and red) to differentiate the warning levels. Simultaneously, it collaborates with the interaction module to push targeted maintenance suggestions to the site.

[0081] The storage module includes a local storage unit and a cloud storage unit, used to store the fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results. This allows the wear prediction module to independently output wear trend prediction results based on the weights in the local storage unit when the network is interrupted. In one specific embodiment, the local storage unit uses a 128GB solid-state drive, which can store nearly 36 months of multi-source data, prediction results, and model weights. The cloud storage unit uses a cloud server cluster to achieve data backup, multi-device data synchronization, and cross-site data aggregation, and supports remote querying. The bidirectional synchronization process between the local storage unit and the cloud storage unit, the independent local operation during network outages, and the version verification after network recovery have been detailed in step S6 of the first embodiment of the method, and will not be elaborated here.

[0082] The interaction module, used to display wear trend prediction results and early warning information, and to receive maintenance data and early warning threshold adjustment instructions entered by the operator, consists of a touch screen and operation buttons. In one specific embodiment, the touch screen is 10.1 inches in size with a resolution of 1920×1080, and there are 5 operation buttons, corresponding to typical operations such as data entry, threshold adjustment, historical record viewing, early warning confirmation, and menu switching. The information write-back path and closed-loop logic between the interaction module, the data acquisition module, and the early warning module have been detailed in step S7 of the method embodiment, and will not be elaborated here.

[0083] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting rail profile wear based on multi-source data fusion, characterized in that, include: Collect multi-source data, including rail profile data, train operation data, environmental data, and maintenance record data; The multi-source data are weighted according to data type and then fused to obtain a fused feature dataset. The fused feature dataset is input into a locally deployed temporal neural network model, which outputs a prediction result of the wear trend of the rail profile over a future period. Differentiated warning thresholds are set according to rail type and service life, and a warning is triggered when the wear trend prediction result exceeds the warning threshold; The fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results are synchronously stored in local storage and cloud storage. In the event of a network outage, the wear trend prediction result is output independently based on the weights stored locally.

2. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, In the weighted fusion, the weight of the rail profile data is greater than the weight of the train operation data, the weight of the train operation data is greater than the weight of the environmental data and the weight of the maintenance record data, and the weight of the environmental data is equal to the weight of the maintenance record data.

3. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, Also includes: Before the weighted fusion, the multi-source data is preprocessed, and the weighted fusion is performed on the preprocessed multi-source data. The preprocessing includes: removing data noise by using mean filtering; completing missing data by using linear interpolation; mapping the data to a preset interval by normalization; and removing outlier data according to the 3σ criterion.

4. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, The temporal neural network model is an improved LSTM neural network model, which includes: A temporal attention mechanism is applied to the hidden layer output of the improved LSTM neural network model; The multi-feature split coding structure splits and encodes the rail profile data, the train operation data, the environmental data, and the maintenance record data before fusing them.

5. The rail profile wear prediction method based on multi-source data fusion according to claim 4, characterized in that: The improved LSTM neural network model also includes inter-layer residual connections; The improved LSTM neural network model is a lightweight structure that has undergone pruning and quantization. The improved LSTM neural network model is trained using an adaptive learning rate and an early stopping strategy.

6. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, The setting of the early warning threshold includes: Establish benchmark limits based on industry standards for railway line wear and repair; Establish a graded threshold library based on rail type; The graded threshold library is modified according to the curve radius of the line, the years of operation, and the transportation intensity; The corrected threshold was statistically calibrated based on historical wear and failure data. The calibrated thresholds are divided into three levels of warning: yellow, orange, and red, based on the proportion of the aforementioned baseline limit.

7. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, When the warning is triggered, it also includes pushing targeted maintenance recommendations, which include at least one of rail grinding, weld repair and rail replacement.

8. The rail profile wear prediction method based on multi-source data fusion according to claim 1, characterized in that, When the network is restored, the local storage and the cloud storage perform bidirectional synchronization on the weights of the temporal neural network model, the fused feature dataset, and the wear trend prediction results.

9. A rail profile wear prediction system based on multi-source data fusion, characterized in that, include: The data acquisition module is used to collect rail profile data, train operation data, environmental data, and maintenance record data as multi-source data; The data fusion module is used to preprocess the multi-source data and perform weighted fusion according to data type to obtain a fused feature dataset; The wear prediction module is used to input the fused feature dataset into a locally deployed temporal neural network model and output the wear trend prediction result of the rail profile over a future period of time. The early warning module is used to set differentiated early warning thresholds according to rail type and service life, and to trigger an early warning when the wear trend prediction result exceeds the early warning threshold; The storage module includes a local storage unit and a cloud storage unit, used to store the fused feature dataset, the weights of the temporal neural network model, and the wear trend prediction results, so that the wear prediction module can independently output the wear trend prediction results based on the weights in the local storage unit when the network is interrupted. The interactive module is used to display the wear trend prediction results and early warning information, and to receive maintenance data and early warning threshold adjustment instructions entered by the operator.

10. The system according to claim 9, characterized in that, The data acquisition module includes a profile acquisition unit, which uses a combination of laser acquisition and visual acquisition methods to acquire the rail profile data.

Images