Data missing value interpolation method based on diffusion model

Abstract

The invention provides a data missing value interpolation method based on a diffusion model. The method comprises the following steps: collecting a track geometric dynamic detection data set, and carrying out data preprocessing on the track geometric dynamic detection data set; constructing a data missing value interpolation model, and training the data missing value interpolation model by using the preprocessed track geometry dynamic detection data set to obtain a trained data missing value interpolation model; and inputting the original track geometric dynamic detection data with the missing value into the trained data missing value interpolation model, recovering the missing value of the original track geometric dynamic detection data by the trained data missing value interpolation model, and obtaining the complete track geometric dynamic detection data after interpolation. According to the method, a missing value interpolation algorithm based on a diffusion and structured state space sequence model is provided, and missing values existing in animation detection data are complemented and filled.

Description

Technical Field

[0001] The present invention relates to the technical field of track geometry dynamic detection, and in particular to a data missing value interpolation method based on a diffusion model. Background Art

[0002] Since the early 21st century, China's high-speed rail industry has experienced rapid growth and development, firmly establishing itself as the world's leading rail transport industry. By the end of 2023, China's total operating mileage of high-speed rail will have reached 4.5 billion kilometers, accounting for 28.3% of the nation's total operating mileage. Between 2018 and 2023, China's high-speed rail operating mileage exhibited a 32.4% compound annual growth rate, a remarkable achievement that demonstrates the rapid pace of China's high-speed rail construction. By 2023, China's railway passenger volume is expected to reach approximately 3.85 billion, of which approximately 2.94 billion will be carried by high-speed rail, an increase of approximately 24% compared to 2019. This indicates that the proportion of high-speed rail in total passenger volume continues to rise. High-speed rail, in particular, demonstrates significant time efficiency and convenience on medium- and long-distance routes of 400 to 800 kilometers. As high-speed rail becomes an increasingly popular mode of transportation, safety issues during travel have become increasingly important.

[0003] Mining high-speed rail track dynamic inspection data primarily involves interpolating missing data, mileage registration based on track record data, detecting and classifying anomalies in invalid data, and more. During the data collection process, missing values ​​often exist in dynamic inspection data due to sensor failures or human error. Incomplete observations significantly impact the quality of downstream mining tasks related to dynamic inspection data and can even affect the safe operation of high-speed rail. Interpolating missing data in dynamic inspection data can optimize data quality, improve the efficiency and accuracy of downstream task algorithms, and better mine the potential information within the data.

[0004] Currently, there is no method in the prior art to effectively interpolate missing data in animal examination data. Summary of the Invention

[0005] An embodiment of the present invention provides a method for interpolating missing data values ​​based on a diffusion model, so as to effectively interpolate missing data in track geometry dynamic detection data.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solutions.

[0007] A data missing value interpolation method based on a diffusion model, comprising:

[0008] collecting a track geometry dynamic detection data set, and performing data preprocessing on the track geometry dynamic detection data set;

[0009] Constructing a data missing value interpolation model, and training the data missing value interpolation model using the preprocessed track geometry dynamic detection data set to obtain a trained data missing value interpolation model;

[0010] The original track geometry dynamic detection data with missing values ​​is input into a trained data missing value interpolation model, and the trained data missing value interpolation model restores the missing values ​​of the original track geometry dynamic detection data to obtain the complete track geometry dynamic detection data after interpolation.

[0011] Preferably, the collecting of the track geometry dynamic detection dataset and the data preprocessing of the track geometry dynamic detection dataset include:

[0012] A track geometry dynamic detection dataset is collected. The track geometry dynamic detection data includes data on track gauge, height, and track direction detection items, as well as auxiliary detection data related to track geometry. The track geometry dynamic detection dataset is preprocessed through data scanning and analysis algorithms to identify the location and number of missing values ​​in the track geometry dynamic detection dataset. According to the distribution characteristics of the data, the distribution pattern of missing values, and actual application requirements, an appropriate masking strategy is selected.

[0013] Preferably, the construction of the data missing value interpolation model includes:

[0014] Construct and train a data missing value interpolation model, which includes a mask processing module, a conditional feature extraction module, and a noise estimation module; set the hyperparameters of the data missing value interpolation model according to the characteristics of the track geometry dynamic detection data, the data volume, and the expected interpolation accuracy requirements. The hyperparameters include parameters related to the diffusion model in the conditional feature extraction module and parameters related to the structured state space sequence model S4 in the noise estimation module, where the parameters related to the diffusion model include: the maximum noise level β T and noise level series

[0015] Preferably, the method of training the missing data interpolation model using the preprocessed track geometry dynamic detection data set to obtain a trained missing data interpolation model includes:

[0016] The pre-processed orbit geometry dynamic observation data X is input into the data missing value interpolation model. The mask processing module uses the mask strategy to mask the input observation data X to obtain the interpolation target data. Then interpolate the target data Simple linear interpolation is performed to obtain enhanced interpolation condition information χ.

[0017] The interpolation target The interpolation condition information χ is input into the conditional feature extraction module. In each training iteration, the conditional feature extraction module performs the interpolation of the interpolation target in the forward diffusion process of the diffusion model. Sampling is performed based on the maximum noise level β T and the noise level series β T Parameters and interpolation condition information χ select the diffusion step value of the current iteration, generate Gaussian noise that meets the requirements, and add the Gaussian noise to the original interpolation target In the process, the noise samples are obtained, and the conditional feature extraction module also outputs the interpolation conditional features used as prior knowledge for noise estimation;

[0018] The noisy samples and interpolation condition features are input into a noise estimation module. The noise estimation module uses the diffusion step value to denoise the noisy samples and updates the gradient. The combined structured state space sequence model S4 in the noise estimation module captures the long-term temporal dependency of the track geometry dynamic detection data and captures the spatiotemporal dependency of the track geometry dynamic detection data through an attention mechanism. The long-term temporal dependency and the spatiotemporal dependency jointly construct data structure features. The noise estimation module uses the interpolation condition features to calculate the denoised and restored data features under the current iteration, and generates interpolated data based on the denoised and restored data features and the data structure information.

[0019] After each iterative training is completed, whether the data missing value interpolation model has converged is judged according to the preset convergence standard. If the data missing value interpolation model has not converged, the next iteration is continued; if the data missing value interpolation model has converged, the training is stopped.

[0020] Preferably, the convergence criteria of the data missing value interpolation model include that the loss function value is less than a preset threshold, the change of the model parameters is less than a preset threshold, or the model performance no longer improves within a certain number of iterations.

[0021] Preferably, the step of inputting the original track geometry dynamic detection data with missing values ​​into a trained data missing value interpolation model, wherein the trained data missing value interpolation model restores the missing values ​​of the original track geometry dynamic detection data to obtain the complete track geometry dynamic detection data after interpolation, comprises:

[0022] Inputting the track geometry dynamic detection data with missing values ​​into a trained data missing value interpolation model, the data missing value interpolation model performs linear interpolation processing on the track geometry dynamic detection data to form continuous time series data, and adding noise to the missing values ​​in the continuous time series data to obey Gaussian distribution to obtain interpolation target data;

[0023] According to the inverse denoising process, the interpolation target data is denoised using the formula and parameter calculations learned during the training process to remove the noise in the interpolation target data, restore the missing values ​​of the original interpolation target, and obtain the complete track geometry dynamic detection data after interpolation.

[0024] It can be seen from the technical solutions provided by the above embodiments of the present invention that the present invention proposes a missing value interpolation algorithm based on diffusion and structured state space sequence models to complete and fill missing values ​​in animal inspection data.

[0025] Additional aspects and advantages of the present invention will be set forth in part in the following description, will be obvious from the following description, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 A schematic diagram illustrating the implementation principle of a data missing value interpolation method based on a diffusion model provided in an embodiment of the present invention;

[0028] Figure 2 A processing flow chart of a data missing value interpolation method based on a diffusion model provided by an embodiment of the present invention;

[0029] Figure 3 and Figure 4 A schematic diagram of the training process of a data missing value interpolation model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The 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 throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0031] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the description of the present invention refers to the presence of the 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, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" used herein includes any unit and all combinations of one or more associated listed items.

[0032] It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention pertains. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art and, unless defined as such herein, will not be interpreted in an idealized or overly formal sense.

[0033] To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings. However, each embodiment does not constitute a limitation on the embodiments of the present invention.

[0034] The implementation principle of the data missing value interpolation method based on the diffusion model provided by the embodiment of the present invention is as follows: Figure 1 As shown in the figure, the overall architecture of the missing value interpolation model for track geometry dynamic detection data based on the diffusion model is presented, which mainly includes the conditional feature extraction module and the noise estimation module.

[0035] The processing flow of a data missing value interpolation method based on a diffusion model provided by an embodiment of the present invention is as follows: Figure 2 As shown, the following steps are included:

[0036] Step S1: Complete the preparation of track geometry dynamic detection data.

[0037] First, a track geometry dynamic detection data set is collected. The track geometry dynamic detection data includes data on track gauge, height, and track direction detection items, as well as other auxiliary detection data related to track geometry. These track geometry dynamic detection data must accurately reflect the actual status and changes of the track geometry.

[0038] Data preprocessing is performed on the track geometry dynamic detection dataset. This stage involves identifying the location and number of missing values ​​in the dataset through data scanning and analysis algorithms. The data for each detection item is then individually examined to ensure accurate identification of all possible missing values. An appropriate masking strategy is selected based on the data distribution characteristics, the distribution pattern of missing values, and the actual application requirements. If the data missingness is relatively random and lacks obvious patterns, a random masking strategy is preferred. If the data exhibits a certain structure or pattern and can be referenced by other relevant data, a hybrid strategy can be used.

[0039] Step S2: Construct and train a data missing value interpolation model, which includes a mask processing module, a conditional feature extraction module and a noise estimation module.

[0040] The hyperparameters of the missing data interpolation model are set based on the characteristics of the track geometry dynamic detection data, the data volume, and the expected interpolation accuracy requirements. The hyperparameters include the parameters related to the diffusion model in the conditional feature extraction module and the parameters related to the structured state space sequence model S4 in the noise estimation module. The parameters related to the diffusion model are: the maximum noise level β T , used to control the noise intensity that is ultimately added to the data during the forward diffusion process of the diffusion model. During the diffusion process, the noise level sequence determines the degree of noise added at each diffusion step, and its value affects the speed and degree of the transition of the data from the original state to the noisy state.

[0041] Figure 3 、 Figure 4 Schematic diagram of the training process of the data missing value interpolation model according to an embodiment of the present invention. It includes masking, diffusion, denoising, and gradient updating. The training process of the data missing value interpolation model includes:

[0042] Step 21: Input the pre-processed orbit geometry dynamic observation data X into the above data missing value interpolation model. First, the mask processing module uses the mask strategy to mask the input observation data X to obtain the interpolation target data. Then interpolate the target data Simple linear interpolation is performed to obtain enhanced interpolation condition information χ.

[0043] When using a random masking strategy, observations in the track geometry dynamic detection dataset are randomly masked according to preset probabilities, generating a mask matrix that identifies which data are masked as missing. When using a hybrid strategy, masking operations are performed based on a combination of data characteristics and other factors. For example, for data with periodic characteristics, partial masking can be performed based on the periodic pattern; for data associated with other data, masking can be performed based on the missing patterns of the associated data. Through these operations, the target data for interpolation is obtained.

[0044] Step 22: The above interpolation target and interpolation condition information χ are input into the conditional feature extraction module. In each training iteration, the conditional feature extraction module performs the interpolation on the above interpolation target in the forward diffusion process of the diffusion model. Sampling is performed based on the maximum noise level β T and the noise level series β T Parameters and interpolation condition information χ stepwise to the original interpolation target Gaussian noise ε is added to obtain noisy samples. The conditional feature extraction module also outputs interpolated conditional features used as prior knowledge for noise estimation.

[0045] In the forward diffusion process of the diffusion model, the original interpolation target is given according to the set rules and parameters. Adding Gaussian noise gradually transforms the original data distribution into a Gaussian noise distribution. This helps to discover latent features, construct multi-noise layer representations, and improve the model's ability to learn complex data distributions.

[0046] In the forward diffusion process of the diffusion model, the diffusion step value of the current iteration is selected from the preset diffusion step range according to a certain probability distribution. The diffusion step value clarifies the stage framework of denoising. When the original data is restored by reverse denoising, the model determines the number of denoising iterations and rhythm based on this diffusion step value. At the same time, in the denoising process, the diffusion step value assists the model in reconstructing data features and predicting missing values. Sample the Gaussian noise, generate the required Gaussian noise according to the preset noise distribution model (such as normal distribution) and the noise level parameter of the current iteration, and add the Gaussian noise to the original interpolation target. middle.

[0047] In the face of noise data generated by sensor failure, the model can enhance the understanding of vehicle vibration characteristics under track unevenness, improve feature extraction and expression capabilities, lay a solid foundation for accurate interpolation of missing values ​​and anomaly detection, and enhance the effect of deep data mining.

[0048] Step 23: Input the above noise samples and interpolation condition features into the noise estimation module.

[0049] After the above-mentioned noisy samples and interpolation condition features are input into the noise estimation module, the noise estimation module uses the diffusion step value to denoise the above-mentioned noisy samples and updates the gradient. The noise of the denoising process includes the noisy original data and the noisy data added during the diffusion process.

[0050] The combined structured state space sequence model S4 in the noise estimation module captures the long-term temporal dependencies of the track geometry dynamic detection data, and the spatiotemporal dependencies of the track geometry dynamic detection data are captured through the attention mechanism. The above long-term temporal dependencies and spatiotemporal dependencies jointly construct the data structure features.

[0051] The noise estimation module calculates the denoised and restored data features in the current iteration using the interpolation condition features used as prior knowledge for noise estimation, and then generates interpolated data based on the denoised and restored data features and the data structure information.

[0052] In dynamic inspection data processing based on diffusion models, the data features recovered after denoising and the aforementioned data structure information become key guides for interpolating missing values. When interpolating missing values ​​for orbital geometry parameters, the accurately restored track elevation and track direction data trends and fluctuations after denoising help the model accurately estimate missing values ​​based on the data's inherent logic and relevance, improving data integrity and continuity. A neural network is used to fit the denoising process, using interpolation condition information to aid denoising and predict noise. Training is performed using the maximum likelihood method to minimize the loss function for predicting noise.

[0053] Dynamic inspection data is multi-dimensional and features complex relationships between sensor data. The diffusion step breaks down the inherent structure of the data, enabling the model to capture hidden temporal and dimensional dependencies, laying the foundation for accurate interpolation of missing values. Properly setting the diffusion step value is crucial. If the value is too small, the data is insufficiently noisy, making it difficult to expose underlying features. For example, the complex relationships between track geometry and vehicle response parameters are difficult for the model to learn, limiting interpolation accuracy. If the value is too large, the data is excessively noisy, distorting the original features, making reverse denoising difficult, and interpolated values ​​tend to deviate from the true values, compromising data integrity and model reliability. Furthermore, the diffusion step value affects computational cost; increasing it significantly increases both computational effort and time. Based on extensive experiments, the paper determined a denoising step value, T, of 50 steps for dynamic inspection data. This balances sufficient learning of data features with reasonable computational cost, ensuring efficient application of the model in real-world engineering scenarios and enhancing its practicality and stability. The noise level sequence, including parameters such as the starting and ending noise levels and the pattern of noise level variation, should be set based on the noise content of the data and the desired denoising effect.

[0054] Step 24: After each training iteration, determine whether the missing data interpolation model has converged based on a preset convergence criterion. This convergence criterion may include a loss function value less than a preset threshold, a change in model parameters less than a preset threshold, or no significant improvement in model performance within a certain number of iterations. If the missing data interpolation model has not converged, proceed to the next iteration. If the missing data interpolation model has converged, training is terminated.

[0055] In the dynamic detection data processing algorithm based on the diffusion model, the loss function calculation is the core link in optimizing the model parameters. MSE (mean-square error) or MAE (mean absolute error) is used to measure the difference between the predicted noise and the actual noise. θ Taking the real added Gaussian noise ε as an example, the MSE is calculated as MAE is N is the number of samples, is the diffusion step data, and χ is the interpolation condition information. During model training, minimizing the loss function makes the predicted noise approach the actual noise, improving denoising capabilities and missing value interpolation accuracy. This ensures that the model accurately learns the noise patterns and feature dependencies of the animal inspection data, enhancing data processing reliability.

[0056] Step S3: interpolation operation stage.

[0057] First, track geometry dynamic detection data with missing values ​​is input into a trained data missing value interpolation model. Prior to input, the data undergoes necessary formatting and preprocessing to ensure it meets the model's input requirements. Simple linear interpolation is then performed on the missing data to form continuous time series data as the underlying information. The linear interpolation method calculates an estimate of the missing value based on a linear relationship between the observed values ​​before and after the missing value, thereby constructing a continuous time series. This constructed continuous time series data is then further processed (such as standardization and normalization) to improve the model's data processing effectiveness and accuracy.

[0058] The interpolation target (i.e., the missing values ​​in the continuous time series data after the interpolation condition information is constructed) is noised to conform to a Gaussian distribution. The noise addition process is similar to that used in the model training phase. The track geometry dynamic detection data with missing values ​​is manipulated according to a preset noise level and diffusion step to achieve an appropriate noise state. The interpolation target data is then obtained, providing a foundation for subsequent denoising and interpolation operations.

[0059] Starting with an interpolation target that follows a Gaussian distribution, the interpolation target data is denoised using the formulas and parameters learned during training, following an inverse denoising process. This inverse denoising process, based on the denoising principles of the diffusion model, gradually removes noise from the interpolation target data based on the conditional information, noise level, and model structure parameters determined during training, restoring the missing values ​​of the original interpolation target. During the denoising process, the denoised values ​​are adjusted and verified as necessary based on the actual data and the model output. For example, if the denoised values ​​fall outside a reasonable range, they need to be corrected based on the physical meaning of the data and the actual situation. If there is a significant inconsistency between the denoised values ​​and the surrounding data, further analysis and adjustments are required. Through these operations, the missing values ​​of the original interpolation target are gradually restored, ultimately obtaining complete interpolated data.

[0060] In summary, the missing value interpolation algorithm based on the diffusion and structured state space sequence model in the embodiment of the present invention can fill in the missing values ​​in the animal inspection data.

[0061] In data processing, faced with the complex temporal and dimensional dependencies and missing data challenges of multi-sensor data, DTSIS4 utilizes the S4 module to capture long-term dependencies and the attention mechanism to extract dimensional associations, accurately interpolating missing values. Comparative experiments on real-world dynamic inspection datasets show that its mean absolute error (MAE) and mean square error (MSE) outperform traditional methods such as mean, median, and BRITS at most missing rates. For example, at a missing rate of 20%, the MAE is approximately 22% lower than BRITS, and the interpolated data is highly consistent with the true value. This provides a solid data foundation for accurate assessment of railway track status and precise investigation of safety hazards, improves the scientific nature of operation and maintenance decisions, and reduces safety risks caused by missing data.

[0062] Those skilled in the art will appreciate that the accompanying drawings are merely schematic diagrams of an embodiment, and the modules or processes in the accompanying drawings are not necessarily required to implement the present invention.

[0063] From the above description of the embodiments, it can be seen that those skilled in the art can clearly understand that the present invention can be implemented by means of software plus the necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0064] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device or system embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, refer to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without making any creative efforts.

[0065] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A method for interpolating missing data values ​​based on a diffusion model, characterized in that: include: Collecting a track geometry dynamic detection data set, and performing data preprocessing on the track geometry dynamic detection data set; Constructing a data missing value interpolation model, and training the data missing value interpolation model using the preprocessed track geometry dynamic detection data set to obtain a trained data missing value interpolation model; The original track geometry dynamic detection data with missing values ​​is input into a trained data missing value interpolation model, and the trained data missing value interpolation model restores the missing values ​​of the original track geometry dynamic detection data to obtain the complete track geometry dynamic detection data after interpolation.

2. The method according to claim 1, characterized in that The collecting of the track geometry dynamic detection data set and performing data preprocessing on the track geometry dynamic detection data set include: The track geometry dynamic detection data set is collected, which includes data on track gauge, height, and track direction detection items, as well as auxiliary detection data related to track geometry. The track geometry dynamic detection data set is preprocessed through data scanning and analysis algorithms to identify the location and number of missing values ​​in the track geometry dynamic detection data set. According to the distribution characteristics of the data, the distribution law of missing values, and actual application requirements, a suitable mask strategy is selected.

3. The method according to claim 2, characterized in that The method of constructing a missing value interpolation model includes: A data missing value interpolation model is constructed and trained, and the data missing value interpolation model includes a mask processing module, a conditional feature extraction module, and a noise estimation module; hyperparameters of the data missing value interpolation model are set according to the characteristics of the track geometry dynamic detection data, the data volume, and the expected interpolation accuracy requirements, and the hyperparameters include parameters related to the diffusion model in the conditional feature extraction module and parameters related to the structured state space sequence model S4 in the noise estimation module, wherein the parameters related to the diffusion model include: maximum noise level β T and the noise level series 4. The method according to claim 3, characterized in that: The method of training the missing data interpolation model using the preprocessed track geometry dynamic detection data set to obtain a trained missing data interpolation model includes: The preprocessed orbital geometry dynamic observation data X is input into the data missing value interpolation model. The mask processing module uses the mask strategy to mask the input observation data X to obtain the interpolation target data Then interpolate the target data Simple linear interpolation is performed to obtain enhanced interpolation condition information χ. The interpolation target and interpolation condition information χ are input into the conditional feature extraction module. In each training iteration, the conditional feature extraction module extracts the interpolation target in the forward diffusion process of the diffusion model. Sampling is performed according to the maximum noise level β T and the noise level series β T Parameters and interpolation condition information χ select the diffusion step value of the current iteration, generate Gaussian noise that meets the requirements, and add the Gaussian noise to the original interpolation target In the above example, a noisy sample is obtained, and the conditional feature extraction module also outputs interpolation conditional features used as prior knowledge for noise estimation; The noisy samples and interpolation condition features are input into the noise estimation module. The noise estimation module uses the diffusion step value to perform denoising on the noisy samples and updates the gradient. The combined structured state space sequence model S4 in the noise estimation module captures the long-term time dependency of the track geometry dynamic detection data, and captures the spatiotemporal dependency of the track geometry dynamic detection data through the attention mechanism. The long-term time dependency and the spatiotemporal dependency jointly construct the data structure feature. The noise estimation module uses the interpolation condition features to calculate the denoised and restored data features under the current iteration, and generates interpolated data according to the denoised and restored data features and the data structure information. After each iteration of training is completed, whether the data missing value interpolation model has converged is determined according to a preset convergence standard. If the data missing value interpolation model has not converged, the next iteration is continued; if the data missing value interpolation model has converged, the training is stopped.

5. The method according to claim 4, characterized in that The convergence criteria of the data missing value interpolation model include that the loss function value is less than a preset threshold, the change of the model parameter is less than a preset threshold, or the model performance no longer improves within a certain number of iterations.

6. The method according to claim 4 or 5, characterized in that: The method of inputting the original track geometry dynamic detection data with missing values ​​into the trained data missing value interpolation model, wherein the trained data missing value interpolation model restores the missing values ​​of the original track geometry dynamic detection data to obtain the complete track geometry dynamic detection data after interpolation, comprises: Inputting the track geometry dynamic detection data with missing values ​​into a trained data missing value interpolation model, the data missing value interpolation model performs linear interpolation processing on the track geometry dynamic detection data to form continuous time series data, and adding noise to the missing values ​​in the continuous time series data to obey Gaussian distribution to obtain interpolation target data; According to the inverse denoising process, the interpolation target data is denoised using the formula and parameter calculations learned during the training process to remove the noise in the interpolation target data, restore the missing values ​​of the original interpolation target, and obtain the complete track geometry dynamic detection data after interpolation.

Images