Railway rail wear prediction method and system in non-periodic track maintenance scenario

Abstract

The application provides a rail wear prediction method and system in a non-periodic line maintenance scene, belongs to the technical field of rail wear prediction based on image processing, and obtains point cloud data of a rail profile; a pre-trained rail wear prediction model is used to process the obtained point cloud data, and a rail wear prediction result is obtained; wherein the rail wear prediction model comprises a point cloud coding layer and a distribution accumulation prediction unit; the point cloud data is processed through the point cloud coding layer, and a feature vector is extracted; the feature vector is input into the distribution accumulation prediction unit, based on a given point cloud feature vector and a target time interval T, and a wear prediction result that monotonically increases with time is output. The application introduces a distribution accumulation prediction with a unique structure, internalizes the physical law of wear development as a structural constraint of the model, ensures that the prediction result strictly follows the monotonic increasing characteristic, and significantly improves the physical rationality of the output and the credibility of long-term prediction.

Description

Technical Field

[0001] This invention relates to the field of rail wear prediction technology based on image processing, and specifically to a method and system for predicting rail wear in non-periodic track maintenance scenarios. Background Technology

[0002] Rail wear prediction is a core component of railway infrastructure condition monitoring and maintenance decision-making. Currently, data-driven time-series prediction methods are the mainstream approach in this field. These methods typically use historically collected rail wear measurements to construct mathematical models to predict future wear trends.

[0003] Existing schemes based on traditional statistical models, such as the Autoregressive Integrated Moving Average (ARIMA) model and the Grey Model (GM(1,1)), capture the statistical regularity of historical wear value sequences over time by modeling them. However, they inherently require the input data to be time series with uniform time intervals, making it difficult to directly handle non-uniform time-interval data collected by track flaw detection vehicles in actual engineering. More importantly, the outputs of these models lack physical constraints and cannot guarantee that the prediction results conform to the physical law of monotonically increasing wear.

[0004] Deep learning-based temporal modeling approaches, including Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM) networks, and the latest Transformer architecture variants (such as Autoformer), learn temporal dependencies in wear data through more complex network structures. However, they also rely on the fundamental assumption that the input should be a long sequence with uniform temporal intervals, which fundamentally contradicts the actual characteristics of railway inspection data. Although preprocessing techniques such as linear interpolation can force alignment of temporal intervals, this method introduces errors and fails to fully utilize the rich geometric information contained in the original point cloud data. Furthermore, the outputs of these deep learning models also lack constraints based on physical laws, potentially leading to unrealistic fluctuations in predicted trends.

[0005] It is worth noting that while some existing studies have attempted to incorporate physical constraints into deep learning models, no effective solution has yet been found in the specific field of rail wear prediction that can simultaneously address the two key issues of "adapting to non-uniform temporal data" and "ensuring compliance with physical constraints." In particular, in processing raw track profile point cloud data, most existing methods remain at the level of using pre-calculated wear values ​​as input, failing to fully extract deeper geometric features from the point cloud data.

[0006] In summary, existing rail wear prediction technologies have the following three main drawbacks:

[0007] First, there is poor data adaptability. Mainstream time series prediction methods (including traditional statistical models and deep learning models) are all based on the fundamental assumption that the input data is a long sequence with uniform time intervals. However, the track profile point cloud data collected by track flaw detection vehicles in actual engineering naturally has the characteristics of short sequences and non-uniform time intervals. This data characteristic is fundamentally contradictory to the model requirements. Although interpolation preprocessing can force a transformation, it will introduce errors and lose the true geometric features of the original data. Second, there is a lack of physical constraints. Rail wear is essentially a monotonically increasing cumulative process. Unless manual rail replacement is carried out, the amount of wear will not decrease. However, the output of existing prediction models generally lacks constraints based on this physical law, which may lead to fluctuations or declines in the prediction results that violate the actual trend, seriously affecting the reliability of long-term predictions and their engineering application value. Third, the features are not fully utilized. Most existing methods are based on pre-calculated wear values ​​for prediction, failing to fully explore and utilize the rich geometric morphological information contained in the original track profile point cloud data. This neglect of the underlying data features limits the model's ability to perceive subtle changes in the rail condition, thus affecting the prediction accuracy. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for predicting rail wear in non-periodic track maintenance scenarios, so as to solve at least one of the technical problems existing in the background art.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, the present invention provides a method for predicting rail wear in non-periodic track maintenance scenarios, comprising:

[0011] Obtain point cloud data of the rail profile;

[0012] The acquired point cloud data is processed using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud data is processed by the point cloud encoding layer to extract feature vectors. The feature vectors are input into the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0013] As a further limitation of the first aspect of the present invention, the point cloud encoding layer includes: converting the two-dimensional profile data of the rail cross section into an image matrix representation; performing multi-level feature encoding on the image matrix using a ResNet architecture; and performing average pooling operation on the feature vectors of multiple sampled cross sections within the same rail segment to generate a feature representation with global representation capabilities.

[0014] As a further limitation of the first aspect of the present invention, the distributed cumulative prediction unit realizes wear prediction by means of the distributed cumulative prediction mechanism. By decomposing the total prediction problem into the sum of multiple sub-prediction problems and ensuring that each sub-prediction result is non-negative, the final prediction result is guaranteed to strictly follow the physical law of monotonically increasing with time from the model structure level.

[0015] As a further limitation of the first aspect of the present invention, the distributed cumulative prediction unit takes the comprehensive feature vector output by the point cloud coding layer as input and is responsible for outputting the total wear prediction value within any target time interval T in the future, including: dividing the entire prediction period T into sub-intervals in months, and calculating the total wear by weighted aggregation of the output results of each monthly prediction sub-network.

[0016] As a further limitation of the first aspect of the present invention, the distributed cumulative prediction unit consists of 12 monthly prediction networks with independent parameters; each monthly prediction network is a three-layer perceptron network, with the same comprehensive feature vector output by the preceding point cloud feature encoding layer as its input; each monthly prediction network works independently, and its function is to output a daily average growth rate of wear for the corresponding month, and to use the ELU activation function in the output layer to ensure that the growth rate is non-negative.

[0017] As a further limitation of the first aspect of the invention, based on the output and input time interval T of these monthly prediction networks, the total wear is calculated using the following formula:

[0018] .

[0019] Secondly, the present invention provides a rail wear prediction system for non-periodic track maintenance scenarios, comprising:

[0020] The acquisition module is used to acquire point cloud data of the rail profile;

[0021] The processing module is used to process the acquired point cloud data using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud encoding layer processes the point cloud data to extract feature vectors. The feature vectors are input to the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0022] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the rail wear prediction method under non-periodic track maintenance scenarios as described in the first aspect.

[0023] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the rail wear prediction method under the non-periodic track maintenance scenario as described in the first aspect.

[0024] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the rail wear prediction method under the non-periodic track maintenance scenario as described in the first aspect.

[0025] Terminology Explanation: Rail Profile Point Cloud Data: This refers to the set of discrete three-dimensional coordinate points that characterize the geometric contour of the rail cross-section, collected by a detection system mounted on a rail flaw detection vehicle using technologies such as structured light 3D imaging. This data is the foundation for quantitative analysis and prediction of rail wear. Distributed Cumulative Prediction Module: The core prediction module proposed in this invention, its function is to embed physical constraints into the model structure. This module consists of multiple sub-networks of monthly prediction modules with independent parameters. By performing time-based weighted accumulation of the non-negative daily growth rate outputs of each sub-network, it ensures that the final wear prediction result strictly follows the monotonically increasing physical law. Monthly Prediction Module: A sub-network in the distributed cumulative prediction module, typically a three-layer MLP structure. Each monthly prediction module is responsible for predicting the average daily growth rate of rail wear within a specific month. Its output is constrained to a non-negative value through the ELU activation function, providing a fundamental guarantee for the monotonic output of the entire prediction model.

[0026] The beneficial effects of this invention are as follows: By designing a dedicated point cloud feature encoding layer, the core problem of mismatch between existing technologies and actual data characteristics is successfully solved, and efficient feature extraction of raw sampled point cloud data is achieved; by introducing a distribution cumulative prediction head with a unique structure, the physical laws of wear and tear development are internalized into the structural constraints of the model, and the prediction results are guaranteed from a mathematical mechanism to strictly follow the monotonically increasing characteristics, which significantly improves the physical rationality of the output and the credibility of long-term prediction.

[0027] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a diagram illustrating the overall architecture of the rail wear prediction model as described in an embodiment of the present invention.

[0030] Figure 2 This is a structural diagram of the distributed cumulative prediction module according to an embodiment of the present invention. Detailed Implementation

[0031] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0032] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0034] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0035] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0036] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0038] Currently, mainstream time-series prediction methods in the field of rail wear prediction (such as ARIMA, LSTM, and Transformer) typically require long sequences of historical data with uniform time intervals as input. However, in actual railway operation and maintenance, rail profile inspection relies on work windows and monthly maintenance plans, resulting in irregular data collection time points and an inherent characteristic of uneven time intervals in the obtained data sequences. This data characteristic, constrained by actual engineering conditions, is fundamentally mismatched with the standard uniform time-series input required by existing prediction models. Forcibly converting such unevenly spaced time-series data into a standard uniform sequence through preprocessing methods such as interpolation will not only distort the true dynamics of rail wear development and introduce artificial fitting errors, but also obscure the key local geometric deformation information contained in the original point cloud data. Furthermore, rail wear essentially follows a monotonically increasing physical cumulative law, while the output of existing prediction models generally lacks a constraint mechanism on physical laws, which may lead to non-physical fluctuations in the prediction results that contradict the actual trend, seriously affecting their credibility in maintenance decisions. Meanwhile, most existing methods are based on abstracted and calculated wear values ​​for modeling, failing to fully explore and utilize the high-dimensional geometric morphology and spatial structure information contained in the original track profile point cloud. Therefore, this invention provides a rail wear prediction method that can directly adapt to track profile point cloud data with non-uniform temporal intervals. It aims to overcome the dependence of existing prediction technologies on data regularity, construct a novel prediction model that strictly follows physical laws and deeply integrates geometric features, thereby improving the accuracy, reliability, and engineering applicability of the prediction results. The wear prediction method based on rail profile point cloud data of this invention achieves the following objectives by designing a dedicated point cloud feature encoding layer and introducing a physically constrained distributed accumulation prediction head: 1) It can directly process original point cloud data with non-uniform temporal intervals, fully exploring its geometric features; 2) It ensures that the prediction results strictly follow the physical law of monotonically increasing wear from the model structure level; 3) It improves the accuracy and engineering applicability of the prediction while reducing dependence on the length of historical data.

[0039] Example 1

[0040] In this embodiment 1, a rail wear prediction system for non-periodic track maintenance scenarios is first provided, including: an acquisition module for acquiring point cloud data of rail profile; and a processing module for processing the acquired point cloud data using a pre-trained rail wear prediction model to obtain rail wear prediction results.

[0041] In this embodiment, the above-described system is used to implement a rail wear prediction method for non-periodic track maintenance scenarios, including: acquiring point cloud data of the rail profile using an acquisition module; processing the acquired point cloud data using a pre-trained rail wear prediction model using a processing module to obtain a rail wear prediction result; wherein, the rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit; the point cloud data is processed by the point cloud encoding layer to extract feature vectors; the feature vectors are input to the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0042] The point cloud encoding layer includes: converting the two-dimensional profile data of the rail cross-section into an image matrix representation; performing multi-level feature encoding on the image matrix using a ResNet architecture; and performing average pooling on the feature vectors of multiple sampled cross-sections within the same rail segment to generate a feature representation with global representation capabilities. The distributed cumulative prediction unit achieves wear prediction through a distributed cumulative prediction mechanism. By decomposing the total prediction problem into the sum of multiple sub-prediction problems and ensuring that each sub-prediction result is non-negative, the model structure guarantees that the final prediction result strictly follows the physical law of monotonically increasing over time. The distributed cumulative prediction unit takes the comprehensive feature vector output by the point cloud encoding layer as input and is responsible for outputting the total wear prediction value within any future target time interval T. This includes: dividing the entire prediction period T into monthly sub-intervals and calculating the total wear by weighted aggregation of the output results of each monthly prediction sub-network.

[0043] The distributed cumulative prediction unit consists of 12 independent monthly prediction networks. Each monthly prediction network is a three-layer perceptron network, which takes the same comprehensive feature vector output from the previous point cloud feature encoding layer as its input. Each monthly prediction network works independently and its function is to output the average daily wear rate for the corresponding month and use the ELU activation function in the output layer to ensure that the growth rate is non-negative.

[0044] Based on the outputs of these monthly prediction networks and the input time interval T, the total wear is calculated using the following formula:

[0045] .

[0046] Example 2

[0047] In this embodiment 2, a wear prediction method based on rail profile point cloud data is provided. The workflow is as follows: First, the original point cloud data of the rail profile is collected by a rail flaw detection vehicle; then, the point cloud data is processed by a point cloud feature encoding layer to extract feature vectors representing the geometric state of the rail; next, the feature vectors are input to the distributed cumulative prediction module; finally, the distributed cumulative prediction module outputs a wear prediction result that monotonically increases with time based on a given target time interval T.

[0048] The point cloud encoding layer includes: converting the two-dimensional profile data of the rail cross section into an image matrix representation; using the ResNet architecture to perform multi-level feature encoding on the image matrix; and performing average pooling on the feature vectors of multiple sampled cross sections within the same rail segment to generate a feature representation with global representation capabilities.

[0049] The distributed cumulative prediction module achieves wear prediction through a distributed cumulative prediction mechanism. Its overall process is as follows: This module acts as the prediction head of the model, taking the comprehensive feature vector output from the point cloud feature encoding layer as input, and is responsible for outputting the total wear prediction value within any future target time interval T (in days). Its core working principle is to divide the entire prediction period T into monthly sub-intervals, and calculate the total wear by weighted aggregation of the outputs of each monthly prediction sub-network. This design, by decomposing the total prediction problem into the sum of multiple sub-prediction problems and ensuring that each sub-prediction result is non-negative, guarantees from the model structure level that the final prediction result strictly follows the physical law of monotonically increasing over time.

[0050] The distributed cumulative prediction module is specifically implemented based on multiple parameter-independent monthly prediction networks: the distributed cumulative prediction module consists of 12 parameter-independent monthly prediction networks. Each monthly prediction network is a three-layer perceptron network, using the same comprehensive feature vector output from the preceding point cloud feature encoding layer as its input. Each monthly prediction network operates independently, its function being to output the average daily wear rate for the corresponding month, and to use the ELU activation function in the output layer to ensure that the growth rate is non-negative. Based on the outputs of these monthly prediction networks and the time interval T between the inputs, the total wear is calculated using the following formula:

[0051]

[0052] The parameter independence of each monthly prediction network allows it to focus on learning the possible differentiated wear patterns in different months, avoiding mutual interference. This calculation paradigm based on the monthly accumulation of non-negative growth rates mathematically guarantees that for any T1 < T2, the predicted wear amount w(T1) < w(T2) must hold true.

[0053] The overall architecture of the rail wear prediction model proposed in this embodiment is as follows: Figure 1 As shown, the architecture consists of two parts: a point cloud feature encoding layer and a distributed cumulative wear prediction head. The point cloud feature encoding layer is responsible for converting the original two-dimensional track profile point cloud into an image matrix and extracting the comprehensive feature vector of the entire rail segment using a ResNet architecture combined with average pooling. The distributed cumulative prediction module, through its unique 12 independent monthly prediction module sub-networks and weighted accumulation mechanism, internalizes physical laws into the structural constraints of the model, ensuring the monotonically increasing characteristics of the prediction results. Its detailed structure is shown in [link to detailed structure]. Figure 2 The purpose of the step-by-step weighted accumulation mechanism is to obtain the final predicted wear from the output of the entire monthly forecasting module. Since wear growth is an irreversible process at the physical level, each monthly forecasting module predicts the wear increment for that month, and uses the ELU activation function at the end of the output layer to ensure the final output is non-negative. Finally, the distributed weighted accumulation mechanism performs a comprehensive weighted accumulation based on the time and wear increment of each month to obtain the total wear amount, as shown in the following formula:

[0054]

[0055] in the formula This represents the final predicted total wear. This represents the wear increment output by the prediction module for the i-th month. The expected time interval for input.

[0056] To verify the effectiveness of the model proposed in this embodiment, it was systematically compared with seven mainstream time series prediction models, including NBEATS, MLP, Autoformer, LSTM, RNN, GM(1,1), and ARIMA. The experiment was conducted on actual detection data provided by a railway bureau. This dataset contains track profile point clouds of approximately 150 kilometers of track, from which 3290 rail segment samples in 10-meter units were extracted and divided into training, validation, and test sets in an 8:1:1 ratio.

[0057] The model's hyperparameters are set as follows: 300 training epochs, using the Adam optimizer, with an initial learning rate of 10. -4 Learning rate decay was applied during training. Mean squared error (MSE) was used as the loss function. All experiments were conducted in the same hardware and software environment to ensure fairness in the comparisons.

[0058] The quantitative evaluation metrics used are the mean absolute error (MAE) and the Spearman rank correlation coefficient (SRC). MAE measures the absolute accuracy of the prediction, and a smaller value is better. SRC is used to evaluate the consistency between the predicted sequence and the actual wear growth trend. In this task, the closer the SRC is to 1, the better the model conforms to the physical law of monotonically increasing wear.

[0059] As shown in Table 3, the experimental results demonstrate that the model proposed in this embodiment exhibits optimal or near-optimal performance in all four prediction tasks (left-side wear, right-side wear, left-side vertical wear, and right-side vertical wear). Specifically, regarding prediction accuracy (MAE), the model in this embodiment has the lowest MAE across all four tasks. For example, in left-side vertical wear prediction, the MAE is 0.343, which is approximately 8.5% higher than the near-optimal MLP (0.375). Regarding trend consistency (SRC), the model maintains a constant SRC of 1.0 across all tasks, strictly guaranteeing the monotonically increasing nature of the prediction results, while the best SRC value of the comparison model fluctuates between 0.740 and 1.0.

[0060] Table 3

[0061]

[0062] Furthermore, the effectiveness and synergistic effect of point cloud data and the distributed cumulative prediction module were further verified through systematic ablation experiments (as shown in Table 4). Experimental results show that the complete model (Model D), which integrates point cloud data and the distributed cumulative prediction module, achieved the lowest MAE in all prediction tasks, especially in the right rail vertical wear prediction, where the error was reduced by 33.2% compared to the baseline model (Model A).

[0063] Table 4

[0064]

[0065] In summary, the method proposed in this embodiment not only outperforms existing mainstream time-series models in prediction accuracy, but more importantly, its prediction results strictly conform to the physical laws of rail wear, thus possessing higher application value in actual railway maintenance work. The key point of this embodiment lies in constructing a prediction scheme capable of directly processing rail point cloud data with non-uniform time intervals, and ensuring through model structure design that the prediction results strictly follow the monotonically increasing physical law of wear.

[0066] In this embodiment, the point cloud encoding layer first converts the point cloud data into a two-dimensional image. The specific formula for converting point cloud coordinates into image coordinates is as follows:

[0067]

[0068]

[0069] in The x-coordinate of the image after transforming the i-th point of a certain frame's point cloud profile is given. The ordinate of the image represents the transformed image of the i-th point in a given frame of the point cloud profile. This represents the x-coordinate of the i-th point in a given frame of point cloud before transformation. This represents the ordinate of the i-th point in a given frame of point cloud before transformation. Represents the x-coordinate of the minimum point cloud. Represents the maximum point cloud ordinate. and This refers to the horizontal and vertical scaling ratio. and For fixed horizontal and vertical offsets, and This is a hyperparameter.

[0070] After obtaining the 2D image, the ResNet architecture is used to encode the image, ultimately obtaining the encoded feature vector of the outline image. The encoding process is as follows:

[0071]

[0072]

[0073]

[0074]

[0075]

[0076] in This represents the l-th residual block, containing multiple convolutional layers, batch normalization, and ReLU activation. Represents the total number of residual blocks. This represents a global average pooling operation, which ultimately yields a d-dimensional feature vector. .

[0077] The training set used by the distributed cumulative prediction module includes point clouds, wear values, and mileage information for each profile in the road segment. The entire dataset is divided into training, validation, and test sets in an 8:1:1 ratio, and the best result from the validation set during training is selected as the convergent model. The distributed cumulative prediction design primarily addresses the problem of sparse supervision signals in multi-step prediction, where only a few months have ground truth values ​​to supervise the model, but the model needs to be able to output predictions of arbitrary step lengths. A cumulative supervision strategy is employed, where the model first predicts future... Monthly growth over the past month Then calculate the cumulative growth over the first k months:

[0078]

[0079] For each month with actual observations By comparison Compared with real cumulative growth To construct the loss function:

[0080]

[0081] This design allows each observable point to provide a gradient signal for the forecasts of all previous months, thus maintaining the consistency of the forecast sequence under sparse supervision.

[0082] Example 3

[0083] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the rail wear prediction method under the non-periodic track maintenance scenario described above. The method includes:

[0084] Obtain point cloud data of the rail profile;

[0085] The acquired point cloud data is processed using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud data is processed by the point cloud encoding layer to extract feature vectors. The feature vectors are input into the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0086] Example 4

[0087] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, and the memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the rail wear prediction method under the non-periodic track maintenance scenario described above. The method includes:

[0088] Obtain point cloud data of the rail profile;

[0089] The acquired point cloud data is processed using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud data is processed by the point cloud encoding layer to extract feature vectors. The feature vectors are input into the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0090] Example 5

[0091] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions to implement the rail wear prediction method under the non-periodic track maintenance scenario described above. The method includes:

[0092] Obtain point cloud data of the rail profile;

[0093] The acquired point cloud data is processed using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud data is processed by the point cloud encoding layer to extract feature vectors. The feature vectors are input into the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

[0094] In summary, this invention constructs a prediction scheme capable of directly processing non-uniform time-series rail point cloud data and ensuring that the prediction results strictly adhere to the monotonically increasing physical law of wear through model structure design. Compared with existing prediction techniques based on traditional statistical models or deep learning time-series models, this invention achieves significant progress in data adaptability, adherence to physical laws, and engineering practicality. Existing technologies, due to their inherent dependence on long-series, uniform time-series data, struggle to effectively process short-series, non-uniform time-series point cloud data collected by track flaw detection vehicles. Furthermore, their prediction results lack physical constraints, potentially leading to outputs that violate the monotonically increasing law of wear, thus limiting their reliable application in engineering practice. This invention, by designing a dedicated point cloud feature encoding layer, successfully solves the core problem of mismatch between existing technologies and actual data characteristics, achieving efficient feature extraction from the original sampled point cloud data. More importantly, by introducing a distributed cumulative prediction head with a unique structure, the physical law of wear development is internalized as a structural constraint of the model, mathematically ensuring that the prediction results strictly adhere to the monotonically increasing characteristic, significantly improving the physical rationality of the output and the credibility of long-term predictions. Experimental verification fully demonstrates that this invention not only outperforms existing mainstream methods in prediction accuracy, significantly reducing the mean absolute error (MAE), but also exhibits excellent trend consistency, with a Spearman rank correlation coefficient (SRC) consistently at 1.0. Simultaneously, the model reduces the requirement for the length of historical data, enhancing its adaptability to data across different time intervals. This results in stronger generalization ability and application value in complex real-world engineering environments, providing more reliable technical support for rail maintenance decision-making.

[0095] In specific implementations, those skilled in the art can make appropriate adjustments without departing from the core idea of ​​this invention: the point cloud feature encoding layer can adopt other network architectures besides ResNet that can effectively process two-dimensional geometric data converted from point clouds or directly process three-dimensional point cloud data (such as PointNet++, CNN and GNN hybrid architecture, etc.); the number of monthly prediction sub-networks in the distributed cumulative prediction module can be adjusted according to the prediction granularity requirements, and its specific structure can also be changed, as long as the parameters remain independent and the output is a non-negative growth rate; in addition, the core idea of ​​this invention is also applicable to railway infrastructure condition prediction tasks such as rail corrugation prediction and track geometry prediction.

[0096] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A method for predicting rail wear in a non-periodic track maintenance scenario, characterized in that, include: Obtain point cloud data of the rail profile; The acquired point cloud data is processed using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud data is processed by the point cloud encoding layer to extract feature vectors. The feature vectors are input into the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

2. The method for predicting rail wear in non-periodic track maintenance scenarios according to claim 1, characterized in that, The point cloud encoding layer includes: converting the two-dimensional profile data of the rail cross section into an image matrix representation; using the ResNet architecture to perform multi-level feature encoding on the image matrix; and performing average pooling on the feature vectors of multiple sampled cross sections within the same rail segment to generate a feature representation with global representation capabilities.

3. The method for predicting rail wear in non-periodic track maintenance scenarios according to claim 1, characterized in that, The distributed cumulative prediction unit achieves wear prediction through a distributed cumulative prediction mechanism. By decomposing the total prediction problem into the sum of multiple sub-prediction problems and ensuring that each sub-prediction result is non-negative, the final prediction result strictly follows the physical law of monotonically increasing with time from the model structure level.

4. The method for predicting rail wear in non-periodic track maintenance scenarios according to claim 3, characterized in that, The distributed cumulative prediction unit takes the comprehensive feature vector output by the point cloud coding layer as input and is responsible for outputting the total wear prediction value within any target time interval T in the future. This includes: dividing the entire prediction period T into sub-intervals in months, and calculating the total wear by weighted aggregation of the output results of each monthly prediction sub-network.

5. The rail wear prediction method under non-periodic track maintenance scenarios according to claim 4, characterized in that, The distributed cumulative prediction unit consists of 12 independent monthly prediction networks. Each monthly prediction network is a three-layer perceptron network, which takes the same comprehensive feature vector output from the previous point cloud feature encoding layer as its input. Each monthly prediction network works independently and its function is to output the average daily wear rate for the corresponding month and use the ELU activation function in the output layer to ensure that the growth rate is non-negative.

6. The method for predicting rail wear in non-periodic track maintenance scenarios according to claim 5, characterized in that, Based on the outputs of these monthly prediction networks and the input time interval T, the total wear is calculated using the following formula: 。 7. A rail wear prediction system for non-periodic track maintenance scenarios, characterized in that, include: The acquisition module is used to acquire point cloud data of the rail profile; The processing module is used to process the acquired point cloud data using a pre-trained rail wear prediction model to obtain rail wear prediction results. The rail wear prediction model includes a point cloud encoding layer and a distributed cumulative prediction unit. The point cloud encoding layer processes the point cloud data to extract feature vectors. The feature vectors are input to the distributed cumulative prediction unit, and based on the given point cloud feature vectors and the target time interval T, a wear prediction result that monotonically increases with time is output.

8. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the rail wear prediction method under non-periodic track maintenance scenarios as described in any one of claims 1-6.

9. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the rail wear prediction method under non-periodic track maintenance scenarios as described in any one of claims 1-6.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the rail wear prediction method under the non-periodic track maintenance scenario as described in any one of claims 1-6.

Images