Subgrade freeze injury diagnosis method based on satellite remote sensing data enhancement and SCNN-BN

Abstract

The invention discloses a roadbed freezing damage diagnosis method based on satellite remote sensing data enhancement and SCNN-BN. The method comprises the following steps: firstly, acquiring PS-InSAR settlement time sequence data of a railway roadbed, and preprocessing the data to eliminate noise and improve data quality; then, performing time sequence alignment and enhancement on the original data by adopting an FDTW algorithm, and generating a training data set with higher representativeness and balance; then, a convolutional neural network model based on a sequence structure is designed, and the convergence speed, stability and diagnosis precision of the model are remarkably improved in combination with a batch normalization technology; and finally, through experimental verification, the performance of the proposed method is comprehensively evaluated. Experimental results show that the method has a high accuracy rate of 96.76% in railway roadbed freeze injury diagnosis, has excellent performance in a noise environment, and is significantly superior to that of a traditional model. The method provides an efficient and reliable solution for intelligent monitoring of the railway infrastructure.

Description

Technical Field

[0001] This invention relates to the monitoring and diagnosis of railway infrastructure, and in particular to a method for diagnosing roadbed frost damage based on satellite remote sensing data augmentation and a deep learning model (SCNN-BN). Background Technology

[0002] Railway subgrade is a crucial component of railway infrastructure, and its stability and safety directly impact the efficiency and reliability of railway transportation. With the intensification of global climate change and the increasing frequency of extreme weather events, the problem of frost damage to railway subgrades is becoming increasingly severe. Frost damage causes changes in the physical and mechanical properties of subgrade materials, leading to uneven ground settlement and track deformation. In severe cases, it can even cause track breakage, significantly increasing the risks to train operation.

[0003] Traditional methods for diagnosing frost damage to railway subgrades, such as manual inspection, ground-penetrating radar (GPR) surveys, and Global Navigation Satellite System (GNSS) measurements, while capable of identifying some anomalies, suffer from drawbacks such as high cost, low efficiency, and limited coverage, making them unsuitable for large-scale, real-time monitoring. In recent years, PS-InSAR technology has been widely applied to frost damage monitoring of railway subgrades due to its high spatial resolution and long-term monitoring capabilities. However, PS-InSAR technology still faces many challenges in practical applications, such as weak radar image signals caused by the regional and random characteristics of frost damage, and data noise interference.

[0004] Traditional machine learning methods (such as support vector machines, random forests, and K-nearest neighbors) face challenges when processing frost damage data, including over-reliance on manual feature engineering, model performance limitations imposed by expert knowledge, and data imbalance. While advancements in deep learning technologies, such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), have been applied in frost damage diagnosis, problems remain, including the need for high-quality labeled data and overfitting.

[0005] Therefore, this invention proposes a method for diagnosing frost damage to railway subgrade based on satellite remote sensing data augmentation and the SCNN-BN model, aiming to improve the accuracy, efficiency, and robustness of frost damage diagnosis. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN. This method, by combining PS-InSAR technology, FDTW data augmentation, and the SCNN-BN deep learning model, effectively improves the accuracy and efficiency of frost damage diagnosis, solving problems such as difficulty in extracting frost damage features, data imbalance, and insufficient noise resistance in existing technologies. This method provides railway transportation departments with a more efficient, real-time, and reliable solution for monitoring roadbed frost damage, and has significant application value.

[0007] The above objectives are achieved through the following technical solutions:

[0008] The present invention provides a method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN, comprising the following steps:

[0009] S1. Data Acquisition and Preprocessing: The railway subgrade is monitored using PS-InSAR technology. Time series data of railway subgrade frost damage settlement are collected as the raw dataset. The collected raw dataset is then preprocessed. The data preprocessing steps include removing missing values, removing outliers, and normalizing or standardizing the data.

[0010] S2. Classification of Frost Damage Types and Intervention Measures: The railway subgrade frost damage settlement time series data after preprocessing in step S1 is labeled to form a labeled dataset. The labeled dataset is divided into four types of frost damage: seasonal deformation, non-seasonal deformation, continuous uplift deformation, and continuous settlement deformation.

[0011] S3, FDTW satellite remote sensing time series data enhancement: The Dynamic Time Warping (DTW) algorithm (FDTW for short) is improved by using a limited search window and a multi-resolution hierarchical mechanism to enhance the labeled dataset obtained in step S2, generating an enhanced dataset. The enhancement specifically includes aligning the time series data of adjacent regions to generate synthetic samples. FDTW generates enhanced samples by calculating the optimal alignment path between two time series data and interpolating the time series points.

[0012] S4. SCNN-BN Model Design and Training: Based on the augmented dataset obtained in step S3, an SCNN-BN model based on Convolutional Neural Network (CNN) and Batch Normalization (BN) technology is designed and trained. The SCNN-BN model automatically extracts temporal features from the settlement data through CNN and accelerates network training through BN. The training of the SCNN-BN model improves the accuracy of frost damage diagnosis by optimizing network parameters, including convolution kernel, bias, and learning rate.

[0013] S5 and SCNN-BN Model Evaluation and Validation: The performance of the SCNN-BN model was validated using accuracy, precision, recall, and F1 score evaluation metrics. In addition, the confusion matrix was used to analyze the predictive ability of the SCNN-BN model for different types of frost damage. Cross-validation was used to further verify the generalization ability of the SCNN-BN model and ensure its stability under different environmental conditions.

[0014] S6. Anti-interference capability verification: In order to verify the anti-interference capability of the SCNN-BN model, noise was added to the test set for interference testing. The performance of the SCNN-BN model under different noise interference conditions was evaluated by calculating the signal-to-noise ratio (SNR).

[0015] S7. The SCNN-BN model, which meets the evaluation criteria and passes the anti-interference capability verification after training, is applied to the diagnosis of frost damage in railway subgrade.

[0016] Furthermore, the labeled dataset mentioned in step S2 is divided into four types of frost damage: seasonal deformation, non-seasonal deformation, continuous uplift deformation, and continuous subsidence deformation. The characteristics of these four types of frost damage are specifically described by mathematical formulas, as shown below:

[0017] Seasonal deformation is related to seasonal freeze-thaw cycles, manifesting as periodic uplift and subsidence. The types of seasonal deformation can be represented as periodic functions.

[0018] D seasonal (t)=A·sin(ωt+φ)+B (1)

[0019] Where: D seasonal (t) represents the seasonal deformation function; A is the settlement amplitude, representing the maximum change in settlement or uplift during the freeze-thaw process; ω is the periodic angular frequency, controlling the periodicity of seasonal changes; t is the time variable; φ is the phase angle, used to adjust the starting point of the cycle; B is the baseline settlement value, representing the settlement value when no frost damage has occurred.

[0020] Non-seasonal deformation is caused by sudden or extreme changes, and the magnitude of these changes is greater than that of seasonal deformation. Non-seasonal deformation is represented by a step function and is characterized by its suddenness.

[0021]

[0022] Where: D non-seasonal (t) represents the non-seasonal deformation function; t0 is the time point when the deformation occurs, indicating the moment when the abrupt change occurs; C is the settlement value after the abrupt change;

[0023] Continuous uplift deformation refers to the continuous upward bulge of the ground, caused by hydrological changes, soil problems, or temperature fluctuations. Continuous uplift deformation is represented by a linear growth function, indicating the continuous upward trend of the ground.

[0024] D uplift (t)=α·t+D0 (3)

[0025] Where: D uplift (t) represents the deformation function of continuous uplift; α is the rate of settlement change, with a positive value indicating the speed of uplift; D0 is the initial settlement value, representing the ground position before frost damage; t is the time variable;

[0026] Continuous settlement deformation refers to the continuous subsidence of the ground, usually caused by foundation problems or water loss. This type of deformation can be represented by an exponential decay model, indicating that the rate of settlement gradually slows down over time.

[0027] D subsidence (t)=D0·e -βt (4)

[0028] Where: D0 is the initial settlement value, which is usually a large positive value, representing the height at which settlement begins; β is the decay rate, representing the acceleration of settlement, which is usually a positive value; t is a time variable, representing the duration of settlement.

[0029] Furthermore, the FDTW satellite remote sensing time-series data enhancement in step S3 involves the following steps:

[0030] S3.1 Initial timing sequence definition: Let S1 = {s 11 s 12 ...s 1m} represents the deformation observation sequence of a certain region, S2={s 21 s 22 ...s 2n} represents the deformation sequence of the region adjacent to this region; the lengths of the two sequences are m and n, respectively, representing the lengths of sequences S1 and S2, and the time steps corresponding to sequences S1 and S2 are T1={t 11 , t 12 ...t 1m} and T2={t 21 , t 22 ...t 2n};

[0031] S3.2 The alignment path between S1 and S2 is calculated using the improved DTW algorithm, FDTW: The optimal alignment path between the two time series data points is constructed by calculating the minimum distance between each pair of time series points. The FDTW algorithm updates the cost matrix recursively, with the cost of each pair of time series points being the Euclidean distance between them. By limiting the search window and using a multi-resolution hierarchical mechanism, the FDTW algorithm reduces the computational complexity from the traditional O(n log n) to O(n log n). 2 The computation time is reduced to O(n), and the formula for calculating the alignment path is:

[0032]

[0033] Where D(i,j) is the calculated cumulative cost, representing the cost from time point [i,j]. arrive The minimum cost between; Let be the Euclidean distance between time steps i and j; D(i-1,j-1), D(i,j-1), and D(i-1,j) are the costs of the candidate paths. The path with the minimum cost is selected to find the optimal alignment path. An alignment path is generated based on the FDTW algorithm.

[0034] path = {(w i1 ,w j1 ),(w i2 ,w j2 ),…,(w iL ,w jL (6)

[0035] Where w iL and w jL These represent the synchronization points S1 and S2 in the sequence alignment path, respectively;

[0036] S3.3 Synthetic Sample Generation and Interpolation: New synthetic samples are generated using the optimal alignment path calculated by FDTW. Specifically, the FDTW algorithm calculates the difference between each pair of time points through interpolation, thereby generating new synthetic samples. The generated synthetic samples are represented by the following formula.

[0037]

[0038] Where λ is a random interpolation factor satisfying 0 ≤ λ ≤ 1, s pl and t pl These are the synthetic sample values ​​and their corresponding time steps; This represents the deformation value of the i-th time series in the time series data at step k. This represents the deformation value of the j-th time series sequence at step k in the time series data. This represents the time value of the i-th time series sequence in the time series data at step k. This represents the time value of the j-th time sequence in the time series at step k.

[0039] S3.4 Time Consistency Adjustment: Perform time consistency adjustment on the synthesized samples generated in step S3.3 to ensure that the time step of the synthesized samples is consistent with the original data. The adjustment rules are as follows:

[0040] (a)t j ∈T p , where t j T represents the time value corresponding to time step j after adjustment. p If the original set of time steps is represented, then the numerical value is taken directly:

[0041] s j =s pj ,t j =t pj (9)

[0042] (b) If Then use linear interpolation:

[0043]

[0044] Among them, sj s represents the deformation value of the j-th sample in the adjusted time series data. pj t represents the deformation value of the j-th sample in the synthetic sample. pj s represents the time value of the j-th sample in the synthetic sample; j This represents the settlement value of the j-th sample in the adjusted time series data. Indicates the closest sample to t in the synthetic sample. j The previous time step, Indicates the closest sample to t in the synthetic sample. j The next time step, Indicates the time step in the synthetic sample The corresponding deformation value, Indicates the time step in the synthetic sample The corresponding deformation value;

[0045] S3.5 Generates different samples by randomly selecting a random interpolation factor λ and repeating steps S3.1-S3.4 to generate a diverse set of enhanced samples {S}. p1 ,S p2 ,…,S pN These synthetic samples were subsequently used for foundation deformation analysis and diagnosis.

[0046] Furthermore, the SCNN-BN model described in step S4 employs a module combining multiple CNN networks and BN networks. The modules are trained sequentially from front to back; that is, the first CNN-BN module is trained first, and its parameters are fixed after reaching the optimum. Then, the combined network of the first and second modules is trained to determine the parameters of the second network module, and so on, until the global optimum is reached. The trained network is used for frost damage diagnosis of railway subgrade. Specific information about the SCNN-BN model is as follows:

[0047] S4.1 Constructing the CNN-BN Module: In the railway frost damage diagnosis task, the Convolutional Neural Network (CNN) extracts local temporal features from the time-series data, including periodic changes and abrupt changes that occur when frost damage occurs. Assuming the input data to the CNN is... Where T is the number of time steps, F is the feature dimension, and the output after convolution is given by the following formula:

[0048] Y = Conv(X; W, b) (11)

[0049] Where Y represents the convolution output, W represents the convolution kernel, b is the bias term, and Conv represents the convolution operation;

[0050] The mathematical expression for batch normalized BN is:

[0051]

[0052] Where μ and σ 2 These represent the mean and variance of the current mini-batch, respectively, while ∈ is a small constant used to prevent division by zero;

[0053] The mathematical expression for the sparse classification cross-entropy (SCCE) loss function is:

[0054]

[0055] Where y is the true category label (represented by an integer), This is the predicted class probability distribution, where N is the total number of samples.

[0056] S4.2 Constructing an SCNN-BN model for railway subgrade frost damage diagnosis, as detailed below:

[0057] Step 1: In the first stage, the labeled dataset in S2 is standardized and divided into training and testing sets. The goal of the first stage of training is to optimize the initial model parameters θ0.

[0058]

[0059] Where L1(θ0) is the loss function for the first stage. These are the optimization parameters obtained from the first stage;

[0060] Step 2: Add more complex network layers and adjust the learning rate. The goal of the second stage is to optimize the parameters θ1 of the new model.

[0061]

[0062] Where L2(θ1) is the loss function for the second stage. These are the optimization parameters obtained from the second stage;

[0063] Step 3: The third stage involves comprehensive training and fine-tuning. The entire network, including all convolutional layers, pooling layers, and fully connected layers, will be jointly trained. The learning rate is gradually increased to accelerate convergence to the optimal solution. The regularization term and learning rate are dynamically adjusted to ensure the model maintains stability and generalization ability in later training stages. The goal of this stage is to optimize all network parameters θ*.

[0064]

[0065] Where L(θ) is the loss function in the final stage, θ * These are the optimal parameters obtained after the final stage of training.

[0066] Furthermore, the specific formula for the evaluation index in step S5, specifically for model evaluation and validation, is as follows:

[0067]

[0068] Wherein, Accuracy represents accuracy, Precision represents precision, Recall represents recall, F1 represents F1-Score, TP represents the number of actual positive samples that were predicted as positive samples, TN represents the number of actual negative samples that were predicted as negative samples, FP represents the number of actual negative samples that were predicted as positive samples, and FN represents the number of actual positive and negative samples that were predicted as negative samples.

[0069] Furthermore, the anti-interference capability verification in step S6 first verifies its robustness by testing the model performance under different noise levels. The signal-to-noise ratio (SNR) is calculated using the following formula:

[0070]

[0071] Where P s P represents signal power. n Indicates noise power;

[0072] When the SNR is less than 0, the energy of the noise exceeds the energy of the original signal, which makes it particularly difficult to extract meaningful features from the noisy composite signal.

[0073] The advantages of this invention compared to the prior art are:

[0074] This invention presents a railway subgrade frost damage diagnosis method based on satellite remote sensing data augmentation and the SCNN-BN model. Addressing the problems of low accuracy, data imbalance, difficulty in feature extraction, and insufficient noise resistance in existing railway subgrade frost damage diagnosis technologies, it significantly improves diagnostic accuracy and efficiency through the following innovative methods:

[0075] 1. Data Augmentation Combining PS-InSAR Technology and FDTW: This invention acquires high-precision settlement time-series data using PS-InSAR technology and augments the time-series data using the FDTW (Fast Dynamic Time Warping) algorithm. The FDTW algorithm effectively solves the misalignment problem in time series data and generates high-quality synthetic samples, thereby alleviating data imbalance and enhancing data representativeness. Compared to traditional methods, the application of FDTW improves data diversity and balance, providing higher-quality input data for subsequent model training.

[0076] 2. Innovative SCNN-BN Deep Learning Model: This invention designs a deep learning model (SCNN-BN) that combines Sequential Convolutional Neural Network (SCNN) and Batch Normalization (BN) techniques. This model overcomes the limitations of traditional feature engineering by automatically learning temporal features from the data, and accelerates the training process through BN, significantly improving the model's training efficiency and stability. Compared to a single convolutional neural network model, SCNN-BN can extract complex temporal features more accurately, effectively improving the accuracy of frost damage diagnosis.

[0077] 3. Enhanced Anti-interference Capability: This invention incorporates noise into the test set for interference testing and uses the signal-to-noise ratio (SNR) metric to verify the robustness of the SCNN-BN model in noisy environments. This method effectively addresses data interference caused by environmental changes or other external factors, enhancing the model's stability and reliability under complex conditions.

[0078] 4. Real-time and efficient frost damage diagnosis: The railway subgrade frost damage diagnosis method of this invention can achieve real-time and efficient frost damage identification, avoiding the high cost and low efficiency problems of traditional monitoring methods. By combining PS-InSAR technology with the SCNN-BN model, this invention realizes large-scale, real-time monitoring of railway subgrade frost damage, providing a scientific basis for railway operation departments and ensuring the safety and sustainable development of railway infrastructure.

[0079] In summary, the railway subgrade frost damage diagnosis method based on satellite remote sensing data augmentation and SCNN-BN model proposed in this invention can overcome the shortcomings of existing technologies, significantly improve the accuracy and efficiency of frost damage diagnosis, reduce costs, enhance anti-interference capabilities, and provide railway transportation departments with a more reliable and efficient subgrade monitoring and maintenance solution. Attached Figure Description

[0080] Figure 1 Flowchart of the method of this invention;

[0081] Figure 2 PS = Deformation rate map of InSAR satellite remote sensing data acquisition;

[0082] Figure 3 Flowchart of FDTW-based satellite remote sensing data augmentation algorithm;

[0083] Figure 4 SCNN-BN model performance confusion matrix diagram Figure 4 In the middle, (a) is the confusion matrix with the highest accuracy of 97.52%, and (b) is the confusion matrix with the lowest accuracy of 94.98%.

[0084] Figure 5Comparison chart of model anti-interference capabilities (comparison results with 10 intelligent diagnostic methods under different SNRs), Figure 5 In the diagram, (a) has an SNR of -2dB, (b) has an SNR of 0dB, (c) has an SNR of 4dB, (d) has an SNR of 4dB, (e) has an SNR of 6dB, and (f) has an SNR of 8dB. Detailed Implementation

[0085] This invention provides a method for diagnosing railway subgrade frost damage based on satellite remote sensing data augmentation and the SCNN-BN model. The specific implementation of this invention will be described in detail below, explaining how each step is implemented and how to optimize the accuracy, efficiency, and robustness of railway subgrade frost damage diagnosis. The following is a further detailed description of the implementation scheme of this invention with reference to the accompanying drawings, and the system flowchart is as follows. Figure 1 As shown.

[0086] S1. Data Acquisition and Preprocessing: In implementing this invention, the settlement time series data of the railway subgrade is first acquired using PS-InSAR technology. PS-InSAR technology utilizes radar signals emitted by satellites to analyze ground deformation, and is particularly suitable for large-scale and long-term monitoring of railway subgrades.

[0087] S1.1 Study Area Selection: The study area is located near Shihezi Railway Station, along the Urumqi-Shihezi section of the Lanzhou-Xinjiang High-Speed ​​Railway. This region belongs to the temperate continental arid climate zone, characterized by severe winters and hot summers, with scattered but concentrated rainfall. The local soil is predominantly seasonally frozen soil, with a loose texture and high water absorption. Therefore, the roadbed is extremely susceptible to severe frost damage and deformation.

[0088] S1.2 PS-InSAR Data Acquisition: Permanent scatterer interferometric synthetic aperture radar (PS-InSAR) technology, with its millimeter-level accuracy, is widely used in deformation monitoring. This invention utilizes data from the Sentinel-1 satellite launched by the European Space Agency (parameter settings are shown in Table 1). This satellite carries a C-band radar imaging system, which can effectively cover the study area and provide continuous image data. Based on the annual time-series deformation rate of the area (e.g., ... Figure 2 As shown, the image acquired on September 30, 2021, was selected as the primary reference image, and the remaining images were used as secondary images. Differential interferometry processing was performed using the primary image, secondary images, and a digital elevation model (DEM), which was generated based on SRTM 1-30 meter elevation data provided by the U.S. Geological Survey. Atmospheric errors were corrected to obtain accurate deformation data for the observation points. This high-precision deformation data provided a solid foundation for subsequent classification and analysis of foundation deformation characteristics.

[0089] Table 1. Specific parameter settings for Sentinel-1A data sampling

[0090] parameter describe Data pattern IW Time range 2022-09-01--2023-08-28 Data volume 53 scenic spots Track direction Ascend Width / km 250 Spatial resolution / m2 5*20 Revisit cycle / d 12 Angle of incidence (degree) 37 polarization mode Vv

[0091] S2. Classification of Frost Damage Types and Intervention Measures: By combining the on-site records of railway staff with the time deformation rate map of the study area, the subgrade deformation can be divided into four types. The frost damage level, characteristics, impact on train operation, and intervention measures for the four types of subgrade deformation are shown in Table 2.

[0092] (1) Seasonal deformation: This is usually related to seasonal freeze-thaw cycles and manifests as periodic uplift and subsidence. This type of deformation can be represented as a periodic function:

[0093] D seasonal (t)=A·sin(ωt+φ)+B (1)

[0094] Where: D seasonal (t) represents the seasonal deformation function; A is the settlement amplitude, representing the maximum change in settlement or uplift during freeze-thaw cycles; ω is the periodic angular frequency, controlling the periodicity of seasonal changes; t is the time variable; φ is the phase angle, used to adjust the starting point of the cycle; B is the baseline settlement value, representing the settlement value when no frost damage has occurred.

[0095] (2) Non-seasonal deformations are usually caused by sudden or extreme changes and have a large amplitude. Such deformations can be represented by a step function and are characterized by their suddenness.

[0096]

[0097] Where: D non-seasonal (t) represents the non-seasonal deformation function; t0 is the time point when the deformation occurs, indicating the moment when the abrupt change occurs; C is the settlement value after the abrupt change, which is usually large and does not change further in subsequent time steps.

[0098] (3) Continuous uplift deformation refers to the continuous upward uplift of the ground, usually caused by hydrological changes, soil problems, or temperature fluctuations. Its changes can be represented by a linear growth function, indicating the continuous upward trend of the ground:

[0099] D uplift (t)=α·t+D0 (3)

[0100] Where: D uplift (t) represents the continuous uplift deformation function; α is the rate of settlement change, with a positive value indicating the speed of uplift; D0 is the initial settlement value, representing the ground position before frost damage occurs; t is the time variable, representing the passage of time during the deformation process.

[0101] (4) Continuous settlement deformation refers to the continuous subsidence of the ground, usually caused by foundation problems or water loss. This type of deformation can be represented by an exponential decay model, indicating that the rate of settlement gradually slows down over time:

[0102] D subsidence (t)=D0·e -β (4)

[0103] Where: D0 is the initial settlement value, which is usually a large positive value, representing the height at which settlement begins; β is the decay rate, representing the acceleration of settlement, which is usually a positive value; t is a time variable, representing the duration of settlement.

[0104] Table 2 Summary of Frost Damage Types and Characteristics

[0105]

[0106]

[0107] Table 2 summarizes the characteristics, causes, intervention needs, and specific intervention measures for each deformation type. Further analysis shows that seasonal frost heave deformation is a normal subgrade deformation, while the other three types are abnormal. This classification provides an important basis for the diagnosis and prevention strategies of railway subgrade frost damage.

[0108] S3. FDTW Satellite Remote Sensing Time Series Data Augmentation: The FDTW algorithm generates multiple time series of the same type based on two existing time series to achieve data augmentation. The core idea of ​​this algorithm is to perform data augmentation by calculating the optimal alignment path between two time series data. The advantage of this process is that it can overcome time series misalignment problems caused by seasonal variations, environmental factors, etc. FDTW not only preserves the characteristics of the original time series data but also enhances the data quality through time series alignment, especially improving the model's diagnostic capabilities under imbalanced data conditions. The data augmentation process of the FDTW algorithm is as follows: Figure 3 As shown. The specific steps of the FDTW satellite remote sensing time series data augmentation algorithm are as follows:

[0109] S3.1 Initial Time Series Definition: Let the time series S1 = {s 11 s 12 ...s 1m} represents the deformation observation sequence of a certain region, S2={s 21 s 22 ...s 2n} represents the deformation sequence of adjacent regions. The lengths of the two sequences are m and n, respectively, and the corresponding time steps are T1 = {t}. 11 , t 12 ...t 1m} and T2={t 21, t 22 ...t 2n The alignment path between S1 and S2 is calculated using the FDTW algorithm, and a synthetic sample is generated based on the alignment result.

[0110] S3.2 calculates the alignment path between S1 and S2 using the improved DTW algorithm, FDTW: It constructs the optimal alignment path between the two time series data points by calculating the minimum distance between each pair of time series points. Specifically, the FDTW algorithm updates the cost matrix recursively, with the cost of each pair of time series points being the Euclidean distance between them. To ensure computational efficiency, the FDTW algorithm limits the search window and uses a multi-resolution hierarchical mechanism, reducing the computational complexity from the traditional O(n^2)^2. 2 The calculation method is reduced to O(n). The formula for calculating the alignment path is:

[0111]

[0112] Where D(i,j) is the calculated cumulative cost, representing the cost from time point [i,j]. arrive The minimum cost between; Let be the Euclidean distance between time steps i and j; D(i-1,j-1), D(i,j-1), and D(i-1,j) are the costs of candidate paths. The path with the minimum cost is selected to find the optimal alignment path. Based on the above process, FDTW generates an alignment path:

[0113] path = {(w i1 ,w j1 ),(w i2 ,w j2 ),…,(w iL ,w jL (6)

[0114] Where w iL and w jL These represent the synchronization points S1 and S2 in the sequence alignment path, respectively.

[0115] S3.3 Synthetic Sample Generation and Interpolation: New synthetic samples can be generated using the optimal alignment path calculated by FDTW. Specifically, the FDTW algorithm calculates the difference between each pair of time points through interpolation, thereby generating new augmented data. These synthetic samples not only possess the statistical characteristics of the original data but also improve the generalization ability of the training model by increasing the diversity of the data. The generated synthetic samples are represented by the following formula;

[0116]

[0117] Where λ is a random interpolation factor satisfying 0 ≤ λ ≤ 1, spl and t pl These are the synthetic sample values ​​and their corresponding time steps; This represents the deformation value of the i-th time series in the time series data at step k. This represents the deformation value of the j-th time series sequence at step k in the time series data. This represents the time value of the i-th time series sequence in the time series data at step k. This represents the time value of the j-th time sequence in the time series at step k.

[0118] S3.4 Time Consistency Adjustment: Time consistency adjustment is performed on the generated synthetic samples to ensure that the time step of the synthetic samples is consistent with that of the original data. The adjustment rules are as follows:

[0119] (a) If t j ∈T p , where t j T represents the time value corresponding to time step j after adjustment. p If the original set of time steps is represented, then the numerical value is taken directly:

[0120] s j =s pj ,t j =t pj (9)

[0121] (b) If Then use linear interpolation:

[0122]

[0123] Among them, s j s represents the deformation value of the j-th sample in the adjusted time series data. pj t represents the deformation value of the j-th sample in the synthetic sample. pj s represents the time value of the j-th sample in the synthetic sample; j This represents the settlement value of the j-th sample in the adjusted time series data. Indicates the closest sample to t in the synthetic sample. j The previous time step, Indicates the closest sample to t in the synthetic sample. j The next time step, Indicates the time step in the synthetic sample The corresponding deformation value, Indicates the time step in the synthetic sample The corresponding deformation value.

[0124] S3.5 Generate different samples. By randomly selecting λ and repeating the above steps, a diverse set of augmented samples {S} is generated. p1 ,Sp2 ,…,S pN These synthetic samples were subsequently used for foundation deformation analysis and diagnosis.

[0125] S4. SCNN-BN Model Design and Training: A sequential CNN-BN model (SCNN-BN) is adopted. Specifically, the SCNN-BN model uses modules combining multiple CNN networks and BN networks. These modules are trained sequentially from front to back. First, the first CNN-BN module is trained until it reaches its optimum, then its parameters are fixed. Next, the combined network of the first and second modules is trained to determine the parameters of the second network module, and so on, until a global optimum is reached. The trained network is used for frost damage diagnosis of railway subgrade. Specific network information is as follows:

[0126] S4.1 CNN-BN Module: The model parameter list is shown in Table 3. In the railway frost damage diagnosis task, convolutional neural networks (CNNs) can effectively extract local temporal features from time-series data, such as periodic changes and abrupt changes that occur when frost damage occurs. The advantage of CNNs in this task lies in their ability to adaptively learn local patterns in time-series data, which reduces the need for manual feature engineering and significantly improves classification accuracy. Assume the input data is... Where T is the number of time steps, F is the feature dimension, and the output after convolution is given by the following formula:

[0127] Y = Conv(X; W, b) (11)

[0128] Where Y represents the convolution output, W represents the convolution kernel, b is the bias term, and Conv represents the convolution operation.

[0129] Batch normalization (BN) is a technique used to address common problems in deep neural network training, such as vanishing and exploding gradients. This technique normalizes the input data of each layer, keeping its mean zero and variance 1. This reduces the difficulty of parameter updates during training and accelerates network convergence. Furthermore, BN has a regularization effect, which helps improve the model's generalization ability. In this model, BN is inserted after each convolutional layer to stabilize the training process and enhance model performance. By normalizing the output of convolutional layers, BN reduces internal covariate bias and makes training subsequent layers more efficient. The mathematical expression for batch normalization is:

[0130]

[0131] Where μ and σ 2 represents the mean and variance of the current mini-batch, respectively, while ∈ is a small constant used to prevent division by zero.

[0132] In multi-class classification tasks, the Sparse Classification Cross-Entropy (SCCE) loss function is widely used to handle integer label classification problems. This loss function measures the model's classification performance by calculating the cross-entropy between the predicted probability distribution of each class and the true label. This loss function is particularly suitable for scenarios with sparse class labels, meaning that when the number of sample classes is small, the model can more accurately identify different classes. The mathematical expression for the Sparse Classification Cross-Entropy loss function is:

[0133]

[0134] Where y is the true category label (represented by an integer), This is the predicted class probability distribution, where N is the total number of samples.

[0135] Table 3 Summary of SCNN-BN Model Parameters

[0136]

[0137] S4.2 SCNN-BN Railway Subgrade Frost Damage Diagnostic Model: The purpose of the sequential training structure is to progressively optimize the model, avoiding problems such as gradient vanishing and gradient explosion in the early stages of training, while accelerating convergence. This structure employs methods such as staged optimization, learning rate adjustment, and regularization to improve model stability, enhance generalization ability, and reduce the risk of overfitting.

[0138] Step 1: Data Preprocessing and Initial Training. In the first stage, the labeled dataset in S2 is standardized and divided into training and test sets. A low learning rate is used in the early layers of the network (such as convolutional and pooling layers) to allow the model to learn low-level features, such as local patterns and simple spatial relationships. This progressive learning method avoids the instability caused by using a high learning rate initially and lays the foundation for learning higher-level features later. The goal of the first stage is to optimize the initial model parameters θ0.

[0139]

[0140] Where L1(θ0) is the loss function for the first stage. These are the optimization parameters obtained from the first stage.

[0141] Step 2: Add more complex network layers and adjust the learning rate. As training progresses, more complex network structures, such as fully connected layers, are gradually introduced, and the learning rate is appropriately increased to accelerate the training process. This stage ensures that the model can capture higher-level features after completing the learning of lower-level features by gradually adjusting the parameters. Adjusting the learning rate can accelerate convergence and avoid the gradient vanishing problem when training complex layers. The goal of the second stage is to optimize the parameters θ1 of the new model:

[0142]

[0143] Where L2(θ1) is the loss function for the second stage. These are the optimization parameters obtained from the second stage.

[0144] Step 3: Third Stage - Comprehensive Training and Fine-tuning. In this final stage, the entire network (including all convolutional, pooling, and fully connected layers) is jointly trained, accelerating convergence to the optimal solution by gradually increasing the learning rate. Dynamically adjusting regularization terms (such as Dropout) and the learning rate ensures the model maintains stability and generalization ability in later training phases. The goal of this stage is to optimize all network parameters θ*:

[0145]

[0146] Where L(θ) is the loss function in the final stage, θ * These are the optimal parameters obtained after the final stage of training.

[0147] S5. Model Evaluation and Validation: The railway subgrade frost damage dataset used includes both observational and simulated data, comprehensively presenting various typical scenarios of railway subgrade settlement characteristics. The original observational data was collected by China Railway Urumqi Group Co., Ltd. using PS-InSAR technology from the K2165-K2265 section of the Lanzhou-Xinjiang Railway. This area is located in the seasonal permafrost zone and is a high-incidence area for railway subgrade frost damage, possessing significant research value. The original dataset contains samples from four categories, with a category ratio of approximately 5:1:2:5, exhibiting a clear imbalance.

[0148] To address the data imbalance problem, this invention employs the proposed FDTW data augmentation technique to generate a balanced simulated dataset. After augmentation, the dataset size reaches 2156×50, with each sample containing 50 time-series features. To ensure the scientific validity and robustness of model training and evaluation, the dataset is divided into training and test sets in an 8:2 ratio. Detailed dataset information is shown in Table 4.

[0149] Table 4 Summary of Railway Subgrade Deformation Samples Due to Freezing Damage

[0150]

[0151] The model's performance was validated using evaluation metrics such as accuracy, precision, recall, and F1 score. Furthermore, a confusion matrix was used to analyze the model's predictive ability for different types of frost damage. Cross-validation was employed to further verify the model's generalization ability and ensure its stability under various environmental conditions. The specific formulas for the evaluation metrics are as follows:

[0152]

[0153] S5.1, FDTW Data Augmentation Validation: To validate the effectiveness of the FDTW data augmentation algorithm, this section analyzes it from two dimensions: data distribution visualization and diagnostic experiments. First, the PCA-KMEANS method is used to perform three-dimensional visualization of the sample distribution before and after augmentation, intuitively presenting the effect of data augmentation. Second, the FDTW-KNN algorithm is used to conduct frost damage diagnosis, evaluating the performance improvement effect brought by the augmented data.

[0154] Table 5 presents the diagnostic experimental results based on the FDTW-KNN algorithm, with accuracy and F1 score as the evaluation metrics. On the original dataset, when k=3, the FDTW-KNN model achieved the highest accuracy of 69.58% and an F1 score of 69.73%. However, the severe imbalance in class distribution significantly limited the model's performance. In contrast, on the data-augmented dataset, when k=1, the FDTW-KNN model achieved an accuracy of 87.50% and an F1 score of 87.46%, representing improvements of 17.92% and 17.73% respectively compared to the original dataset. These results demonstrate that the FDTW-based data augmentation algorithm can effectively balance class distribution and capture intra-class variation features in time-series data, thereby significantly improving the model's performance in frost damage diagnosis. The results strongly prove that the FDTW data augmentation algorithm has significant advantages in processing time-series data. On the one hand, the augmented dataset effectively alleviates the class imbalance problem, providing more balanced input data for the diagnostic model; on the other hand, the algorithm can accurately capture the variation patterns within the same class, significantly improving the diagnostic accuracy of roadbed frost damage tasks. These results validate the effectiveness of the proposed method, making it a reliable and efficient solution in the field of time-series data augmentation for railway roadbed frost damage.

[0155] Table 5. Classification performance of FDTW-KNN on the original and augmented datasets.

[0156]

[0157]

[0158] Experimental Results of S5.2 SCNN-BN Model: Table 6 shows the average accuracy, precision, recall, and F1 score of the proposed SCNN-BN model in 10 repeated runs. The average accuracy of the SCNN-BN model reaches 96%. The precision reaches 76%, with a maximum of 97.52%. The F1 score for classes 0 and 3 exceeds 93%, while the F1 score for classes 1 and 2 even breaks the 99% mark. Figure 4The comparison of confusion matrices with the highest (97.52%) and lowest (94.98%) accuracy is presented. The lower accuracy is due to the high similarity of frost damage features in categories 0 and 3, which significantly increases the difficulty of classification, especially in noisy environments. However, the overall fault diagnosis results show that the SCNN-BN framework proposed in this invention can effectively identify various frost damage conditions in railway subgrades.

[0159] Table 6. Diagnostic results of SCNN-BN algorithm for frost damage to railway subgrade.

[0160]

[0161] S6. Anti-interference Capability Verification: To verify the model's anti-interference capability, noise was added to the test set for interference testing. The signal-to-noise ratio (SNR) was calculated to evaluate the model's performance under different noise interference conditions, further verifying its robustness. In real industrial environments, the raw signals acquired by sensors typically contain a large amount of background noise. To evaluate the robustness of the SCNN-BN model under such conditions, Gaussian white noise was added to the raw signal in the experiment. The robustness was verified by testing the model's performance under different noise levels. The formula for calculating the signal-to-noise ratio (SNR) is:

[0162]

[0163] Where P s P represents signal power. n This represents the noise power. When the SNR is less than 0, the energy of the noise exceeds the energy of the original signal, making it particularly difficult to extract meaningful features from the noisy composite signal.

[0164] Table 7. Test results of nine different models under different noise environments (all indicators are expressed as a percentage).

[0165]

[0166] To verify the model's robustness against interference, the proposed SCNN-BN model was compared with 10 cutting-edge intelligent diagnostic methods, including Sequential CNN (SCNN), Artificial Neural Network (ANN), Long Short-Term Memory Network (LSTM), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and Gradient Boosting Classifier (GBC). The test noise levels covered the range of -2 to 8 dB. All experiments were repeated 10 times, with each model using the same structure and parameter settings. Table 7 shows the average accuracy and standard deviation of each model. Figure 5The results present the average accuracy performance of these 10 methods under different noise conditions. Experimental results show that the SCNN-BN model maintains the highest accuracy across all noise levels. When the signal-to-noise ratio (SNR) reaches 8 dB, the model achieves a peak accuracy of 94.91%. As the SNR decreases, the accuracy of each method declines. When the SNR drops to 4 dB, the accuracy of the SCNN-BN model is closest to that of the other methods. However, when the SNR further decreases to 0 dB and -2 dB, the SCNN-BN model exhibits a significant accuracy advantage over the other methods.

Claims

1. A method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN, characterized in that, The method includes the following steps: S1. Data Acquisition and Preprocessing: The railway subgrade is monitored using PS-InSAR technology. Time series data of railway subgrade frost damage settlement are collected as the raw dataset. The collected raw dataset is then preprocessed. The data preprocessing steps include removing missing values, removing outliers, and normalizing or standardizing the data. S2. Classification of Frost Damage Types and Intervention Measures: The railway subgrade frost damage settlement time series data after preprocessing in step S1 is labeled to form a labeled dataset. The labeled dataset is divided into four types of frost damage: seasonal deformation, non-seasonal deformation, continuous uplift deformation, and continuous settlement deformation. S3, FDTW satellite remote sensing time series data enhancement: The Dynamic Time Warping (DTW) algorithm (FDTW for short) is improved by using a limited search window and a multi-resolution hierarchical mechanism to enhance the labeled dataset obtained in step S2, generating an enhanced dataset. The enhancement specifically includes aligning the time series data of adjacent regions to generate synthetic samples. FDTW generates enhanced samples by calculating the optimal alignment path between two time series data and interpolating the time series points. S4. SCNN-BN Model Design and Training: Based on the augmented dataset obtained in step S3, an SCNN-BN model based on Convolutional Neural Network (CNN) and Batch Normalization (BN) technology is designed and trained. The SCNN-BN model automatically extracts temporal features from the settlement data through CNN and accelerates network training through BN. The training of the SCNN-BN model improves the accuracy of frost damage diagnosis by optimizing network parameters, including convolution kernel, bias, and learning rate. S5 and SCNN-BN Model Evaluation and Validation: The performance of the SCNN-BN model was validated using accuracy, precision, recall, and F1 score evaluation metrics. In addition, the confusion matrix was used to analyze the predictive ability of the SCNN-BN model for different types of frost damage. Cross-validation was used to further verify the generalization ability of the SCNN-BN model and ensure its stability under different environmental conditions. S6. Anti-interference capability verification: In order to verify the anti-interference capability of the SCNN-BN model, noise was added to the test set for interference testing. The performance of the SCNN-BN model under different noise interference conditions was evaluated by calculating the signal-to-noise ratio (SNR). S7. The SCNN-BN model, which meets the evaluation criteria and passes the anti-interference capability verification after training, is applied to the diagnosis of frost damage in railway subgrade.

2. The method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN according to claim 1, characterized in that, The labeled dataset mentioned in step S2 is divided into four types of frost damage: seasonal deformation, non-seasonal deformation, continuous uplift deformation, and continuous subsidence deformation. The characteristics of these four types of frost damage are specifically described by mathematical formulas, as shown below: Seasonal deformation is related to seasonal freeze-thaw cycles, manifesting as periodic uplift and subsidence. The types of seasonal deformation can be represented as periodic functions. D seasonal (t)=A·sin(ωt+φ)+B (1) Where: D seasonal (t) represents the seasonal deformation function; A is the settlement amplitude, representing the maximum change in settlement or uplift during the freeze-thaw process; ω is the periodic angular frequency, controlling the periodicity of seasonal changes; t is the time variable; φ is the phase angle, used to adjust the starting point of the cycle; B is the baseline settlement value, representing the settlement value when no frost damage has occurred. Non-seasonal deformation is caused by sudden or extreme changes, and the magnitude of these changes is greater than that of seasonal deformation. Non-seasonal deformation is represented by a step function and is characterized by its suddenness. Where: D non-seasonal (t) represents the non-seasonal deformation function; t0 is the time point when the deformation occurs, indicating the moment when the abrupt change occurs; C is the settlement value after the abrupt change; Continuous uplift deformation refers to the continuous upward bulge of the ground, caused by hydrological changes, soil problems, or temperature fluctuations. Continuous uplift deformation is represented by a linear growth function, indicating the continuous upward trend of the ground. D uplift (t)=α·t+D0 (3) Where: D uplift (t) represents the deformation function of continuous uplift; α is the rate of settlement change, with a positive value indicating the speed of uplift; D0 is the initial settlement value, representing the ground position before frost damage; t is the time variable; Continuous settlement deformation refers to the continuous subsidence of the ground, usually caused by foundation problems or water loss. This type of deformation can be represented by an exponential decay model, indicating that the rate of settlement gradually slows down over time. D subsidence (t)=D0·e -βt (4) Where: D0 is the initial settlement value, which is usually a large positive value, representing the height at which settlement begins; β is the decay rate, representing the acceleration of settlement, which is usually a positive value; t is a time variable, representing the duration of settlement.

3. The method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN according to claim 1, characterized in that, The FDTW satellite remote sensing time-series data augmentation in step S3 involves the following steps: S3.1 Initial timing sequence definition: Let S1 = {s 11 s 12 ...s 1m } represents the deformation observation sequence of a certain region, S2={s 21 s 22 ...s 2n } represents the deformation sequence of the region adjacent to this region; the lengths of the two sequences are m and n, respectively, representing the lengths of sequences S1 and S2, and the time steps corresponding to sequences S1 and S2 are T1={t 11 , t 12 ...t 1m } and T2={t 21 , t 22 ...t 2n }; S3.2 The alignment path between S1 and S2 is calculated using the improved DTW algorithm, FDTW: The optimal alignment path between the two time series data points is constructed by calculating the minimum distance between each pair of time series points. The FDTW algorithm updates the cost matrix recursively, with the cost of each pair of time series points being the Euclidean distance between them. By limiting the search window and using a multi-resolution hierarchical mechanism, the FDTW algorithm reduces the computational complexity from the traditional O(n log n) to O(n log n). 2 The computation time is reduced to O(n), and the formula for calculating the alignment path is: Where D(i,j) is the calculated cumulative cost, representing the cost from time point [i,j]. arrive The minimum cost between; Let be the Euclidean distance between time steps i and j; D(i-1,j-1), D(i,i-1), and D(j-1,j) are the costs of the candidate paths. The path with the minimum cost is selected to find the optimal alignment path. An alignment path is generated based on the FDTW algorithm. path={(w i1 ,w j1 ),(w i2 ,w j2 ),…,(w iL ,w jL )} (6) Where w iL and w jL These represent the synchronization points S1 and S2 in the sequence alignment path, respectively; S3.3 Synthetic Sample Generation and Interpolation: New synthetic samples are generated using the optimal alignment path calculated by FDTW. Specifically, the FDTW algorithm calculates the difference between each pair of time points through interpolation, thereby generating new synthetic samples. The generated synthetic samples are represented by the following formula. Where λ is a random interpolation factor satisfying 0 ≤ λ ≤ 1, s pl and t pl These are the synthetic sample values ​​and their corresponding time steps; This represents the deformation value of the i-th time series in the time series data at step k. This represents the deformation value of the j-th time series sequence at step k in the time series data. This represents the time value of the i-th time series sequence in the time series data at step k. This represents the time value of the j-th time sequence in the time series at step k. S3.4 Time Consistency Adjustment: Perform time consistency adjustment on the synthesized samples generated in step S3.3 to ensure that the time step of the synthesized samples is consistent with the original data. The adjustment rules are as follows: (a) If t j ∈T p , where t j T represents the time value corresponding to time step j after adjustment. p If the original set of time steps is represented, then the numerical value is taken directly: s j =s pj ,t j =t pj (9) (b) If Then use linear interpolation: Among them, s j s represents the deformation value of the j-th sample in the adjusted time series data. pj t represents the deformation value of the j-th sample in the synthetic sample. pj s represents the time value of the j-th sample in the synthetic sample; j This represents the settlement value of the j-th sample in the adjusted time series data. Indicates the closest sample to t in the synthetic sample. j The previous time step, Indicates the closest sample to t in the synthetic sample. j The next time step, Indicates the time step in the synthetic sample The corresponding deformation value, Indicates the time step in the synthetic sample The corresponding deformation value; S3.5 Generates different samples by randomly selecting a random interpolation factor λ and repeating steps S3.1-S3.4 to generate a diverse set of augmented samples {S}. p1 ,S p2 ,…,S pN These synthetic samples were subsequently used for foundation deformation analysis and diagnosis.

4. The method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN according to claim 1, characterized in that, The SCNN-BN model described in step S4 employs multiple modules combining CNN and BN networks. These modules are trained sequentially from front to back: first, the first CNN-BN module is trained until it reaches its optimal state, then its parameters are fixed. Next, the combined network of the first and second modules is trained to determine the parameters of the second module, and so on, until a global optimum is reached. The trained network is used for frost damage diagnosis of railway subgrade. Specific information about the SCNN-BN model is as follows: S4.1 Constructing the CNN-BN Module: In the railway frost damage diagnosis task, the Convolutional Neural Network (CNN) extracts local temporal features from the time-series data, including periodic changes and abrupt changes that occur when frost damage occurs. Assuming the input data to the CNN is... Where T is the number of time steps, F is the feature dimension, and the output after convolution is given by the following formula: Y = Conv(X; W, b) (11) Where Y represents the convolution output, W represents the convolution kernel, b is the bias term, and Conv represents the convolution operation; The mathematical expression for batch normalized BN is: Where μ and σ 2 These represent the mean and variance of the current mini-batch, respectively, while ∈ is a small constant used to prevent division by zero; The mathematical expression for the sparse classification cross-entropy (SCCE) loss function is: Where y is the true category label (represented by an integer), This is the predicted class probability distribution, where N is the total number of samples. S4.2 Constructing an SCNN-BN model for railway subgrade frost damage diagnosis, as detailed below: Step 1: In the first stage, the labeled dataset in S2 is standardized and divided into training and testing sets. The goal of the first stage of training is to optimize the initial model parameters θ0. Where L1(θ0) is the loss function for the first stage. These are the optimization parameters obtained from the first stage; Step 2: Add more complex network layers and adjust the learning rate. The goal of the second stage is to optimize the parameters θ1 of the new model. Where L2(θ1) is the loss function for the second stage. These are the optimization parameters obtained from the second stage; Step 3: The third stage involves comprehensive training and fine-tuning. The entire network, including all convolutional layers, pooling layers, and fully connected layers, will be jointly trained. The learning rate is gradually increased to accelerate convergence to the optimal solution. The regularization term and learning rate are dynamically adjusted to ensure the model maintains stability and generalization ability in later training stages. The goal of this stage is to optimize all network parameters θ*. Where L(θ) is the loss function in the final stage, and θ* is the optimal parameter obtained after the final stage of training.

5. The method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN according to claim 1, characterized in that, The specific formula for the evaluation index in step S5, model evaluation and validation, is as follows: Wherein, Accuracy represents accuracy, Precision represents precision, Recall represents recall, F1 represents F1-Score, TP represents the number of actual positive samples that were predicted as positive samples, TN represents the number of actual negative samples that were predicted as negative samples, FP represents the number of actual negative samples that were predicted as positive samples, and FN represents the number of actual positive and negative samples that were predicted as negative samples.

6. The method for diagnosing roadbed frost damage using satellite remote sensing data augmentation and SCNN-BN according to claim 1, characterized in that, The anti-interference capability verification in step S6 first verifies its robustness by testing the model performance under different noise levels. The signal-to-noise ratio (SNR) is calculated using the following formula: Where P s P represents signal power. n Indicates noise power; When the SNR is less than 0, the energy of the noise exceeds the energy of the original signal, which makes it particularly difficult to extract meaningful features from the noisy composite signal.

Images