An orbit foundation settlement evaluation method based on electrical digital data processing

Abstract

The present invention relates to the technical field of track evaluation, and specifically relates to a track foundation settlement evaluation method based on electronic digital data processing, which is as follows: collect and process the original data related to track foundation settlement, then store the processed data, and manually label the stored data; perform data enhancement, synthetic data, and sample diversity optimization operations based on the collected data, and then use the sample interpolation method based on boundary conditions to synthesize electronic digital data samples; establish a track foundation settlement evaluation prediction model, process the collected data and the synthesized data to train the model, obtain the evaluation prediction result, and then evaluate and improve the trained model; provide suggestions on settlement control and repair to users based on the evaluation prediction result of the track foundation settlement evaluation prediction model, and judge whether to give an early warning according to the evaluation prediction result. This method can improve the accuracy of model evaluation and enhance the robustness and generalization ability of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of track evaluation, and particularly to a method for evaluating track foundation settlement based on electrical digital data processing. Background Art

[0002] With the continuous expansion of urban rail systems and the increasing operating load, settlement of track foundations inevitably occurs. Traditional monitoring methods mostly rely on analog signals or simple digital acquisition methods, which results in deficiencies in data accuracy, real-time performance, and anti-interference ability, and cannot comprehensively and accurately reflect the actual settlement conditions. In the existing technology, single or localized sensors are often used in the data acquisition link, and the data preprocessing process is easily affected by environmental noise and interference. In addition, due to the non-uniform data format and the lack of effective time series processing, there are certain errors and discontinuities in the obtained settlement information.

[0003] On the other hand, traditional settlement evaluation methods often use algorithms such as classical neural networks, support vector machines, or decision trees in the model construction and prediction processes. These methods are prone to falling into local optimal states when facing high-dimensional, non-linear, and variable electrical digital data, and the parameter initialization and gradient update mechanisms lack sufficient flexibility, resulting in unsatisfactory prediction accuracy and stability of the models. At the same time, for the problems of insufficient data samples and uneven distribution, traditional sample synthesis methods mostly use simple linear interpolation, which fails to fully consider the boundary conditions and dynamic change trends existing in the data, resulting in the synthesized training data being unable to truly reflect the complex characteristics of track foundation settlement, thus affecting the generalization ability and practical application effect of the models.

[0004] The existing technology has the following deficiencies:

[0005] (1) Traditional neural networks, support vector machines, and decision trees are prone to falling into local optimal solutions when facing complex, high-dimensional, and non-linear electrical digital data, and cannot comprehensively capture global features, resulting in low classification accuracy and stability;

[0006] (2) Traditional machine learning models lack means for adaptive adjustment of global and local features during parameter initialization and gradient update, and are prone to gradient disappearance or explosion, affecting the convergence speed and training effect of the models;

[0007] (3) Traditional data synthesis and expansion mainly rely on simple linear interpolation, which is difficult to capture the boundary conditions and non-linear changes in the true data distribution, resulting in the synthesized new samples often deviating from the characteristics of actual track settlement and being unable to effectively solve the data imbalance problem.

[0008] Therefore, the present invention proposes a method for evaluating track foundation settlement based on electrical digital data processing to solve the above problems. Summary of the Invention

[0009] In view of the deficiencies of the prior art, the present invention develops a method for evaluating track foundation settlement based on electronic digital data processing. By means of a sample interpolation method based on boundary conditions and a fractional-order neural network based on path integral, the present invention can improve the accuracy of model evaluation and enhance the robustness and generalization ability of the model.

[0010] The technical solution for the present invention to solve the technical problem is a method for evaluating track foundation settlement based on electronic digital data processing, including the following steps:

[0011] S1. Data collection: Collect the original data related to track foundation settlement, then process it, store the processed data, and manually label the stored data;

[0012] S2. Sample synthesis: Perform data enhancement, synthetic data, and sample diversity optimization operations based on the collected data, and then use a sample interpolation method based on boundary conditions to synthesize electronic digital data samples;

[0013] S3. Learning and modeling: Establish a track foundation settlement evaluation and prediction model, preprocess the collected data and the synthetic data, extract key features according to the preprocessed data through feature engineering, then train the track foundation settlement evaluation and prediction model based on the extracted key features to obtain the evaluation and prediction results, and then evaluate and improve the trained model;

[0014] Among them, the fractional-order neural network based on path integral is adopted for the track foundation settlement evaluation and prediction model:

[0015] S4. Suggestions and warnings: Provide suggestions for settlement control and repair to users based on the evaluation and prediction results of the track foundation settlement evaluation and prediction model, and judge whether to issue a warning according to the evaluation and prediction results.

[0016] S1 is specifically as follows:

[0017] Automatically collect data through sensor devices. The sensors include settlement detectors, accelerometers, and seismic wave detectors. The sensor devices are arranged at key positions of the track foundation to collect the settlement state information of the track in real time. The data related to track foundation settlement collected includes ground settlement data, vibration data, and surrounding environment change data. After the collected data passes through various sensor devices and is digitally processed, it is transmitted in the form of discrete and accurate electronic digital data. The electronic digital data is in a digital format suitable for computer processing;

[0018] Then, perform time series processing and formatting processing on the data collected by various sensor devices, and remove the noise and interference in the data;

[0019] The processed data is transmitted to the storage system via wired or wireless communication, and manual annotation is performed for each piece of electrical digital data input to the storage system. A unique identifier is assigned to each piece of electrical digital data, and the identifier is the level of the settlement status of the track foundation and the annotation of data anomalies. The levels of the settlement status of the track foundation for the identifier are normal settlement, slight settlement, and severe settlement respectively.

[0020] S2 is as follows:

[0021] Based on the collected data, additional training data samples are synthesized, and the sample synthesis includes the following operations:

[0022] Data augmentation: Transform operations are performed on the collected data, and the transform operations include rotation, scaling, and translation;

[0023] Synthetic data: Based on the characteristics of the collected data, data augmentation methods are used to synthesize electrical digital data samples similar to the collected data;

[0024] Sample diversity optimization operation: By analyzing the collected data, the distribution of the synthesized electrical digital data samples is adjusted;

[0025] An electrical digital data sample is synthesized using a sample interpolation method based on boundary conditions: The minority-class electrical digital data samples are used as seed samples, and the sample selection is optimized by minimizing the energy function. At the same time, boundary condition analysis is performed on the seed samples, the distances and directions to their nearest neighbor samples are calculated, and the boundary conditions are dynamically identified.

[0026] In the specific implementation manner, the operation of synthesizing electrical digital data samples using the sample interpolation method based on boundary conditions is as follows:

[0027] (1) Select minority-class electrical digital data samples from the collected data as seed samples, and obtain an energy function for evaluating the quality of the selection of electrical digital data of the class by minimizing the energy function;

[0028] (2) By calculating the distances and directions between the seed samples and their nearest neighbor samples, dynamic boundary conditions are identified, boundary condition analysis is performed on each seed sample, and then the local structural characteristics of the distribution of electrical digital data samples are judged;

[0029] (3) Through the interpolation method and the dynamic adjustment vector based on boundary conditions, the interpolation path and the quantity and direction of the synthesized electrical digital data samples are dynamically adjusted;

[0030] (4) According to the influence of the newly synthesized point digital data samples on the overall distribution of electrical digital data samples, the synthesis strategy is corrected, and the calculation of the dynamic adjustment factor is performed during the synthesis process of electrical digital data samples;

[0031] Repeat the steps of sample interpolation method based on boundary conditions for multiple times to synthesize digital data samples until the number of the synthesized new sample data reaches a preset number threshold.

[0032] S3 is specifically as follows:

[0033] Construct an evaluation and prediction model for track foundation settlement, and input the collected data and the synthesized data into the evaluation and prediction model for track foundation settlement for training. The specific operations are as follows:

[0034] Data preprocessing: Read the digital data in the storage system and the digital data samples after sample synthesis, and perform preprocessing operations on the data. The preprocessing operations include filling in missing values, removing outliers, and data standardization;

[0035] Extract key features: Extract the key features in the process of track foundation settlement through feature engineering. The key features include foundation compaction degree, soil humidity, and temperature;

[0036] Training of the evaluation and prediction model for track foundation settlement: Train the model through a fractional-order neural network based on path integral. Specifically, simulate multiple possible decision paths through path integral, perform a global search for the parameters of the fractional-order neural network, and perform random perturbations during the training process with the help of the principle of Brownian motion, and finally obtain the evaluation and prediction results of track foundation settlement;

[0037] Evaluation and improvement of the evaluation and prediction model for track foundation settlement: Compare the training results with the actual results. The actual results are the manually marked identifiers, evaluate the accuracy of the evaluation and prediction model for track foundation settlement, and adjust and improve the evaluation and prediction model for track foundation settlement according to the evaluation results.

[0038] In the specific implementation manner, the training of the evaluation and prediction model for track foundation settlement is specifically as follows:

[0039] (1) Initialize the weights and biases of the fractional-order neural network through non-integer order derivatives in fractional calculus. The calculation formula is as follows:

[0040] ,

[0041] ,

[0042] ,

[0043] In the formula, represents the key feature of the th training data input to the fractional-order neural network, represents the key feature of the th training data input to the fractional-order neural network, Represents the total number of training data input to the fractional-order neural network in the current batch, Represents the regularization parameter of the fractional-order neural network, Represents the layer weights of the fractional-order neural network, Represents the layer bias of the fractional-order neural network. The fractional-order neural network has a total of layers, , Represents a symbol subject to a specific distribution, Represents the standard normal distribution, Represents the fractional-order control parameter, Is the standard deviation initialized for the fractional-order neural network;

[0044] (2) Adopt an integral path global search mechanism to optimize the fractional-order neural network in the global parameter space. Through dynamic path weighting, according to the similarity between different input training data and the decision boundary, adjust the contributions of different paths. The calculation formula of the integral path global search mechanism is as follows:

[0045] ,

[0046] ,

[0047] In the formula, Represents the key feature of the th training data input to the fractional-order neural network, Represents the calculation result of the integral path global search mechanism, Represents the mean of the training data input to the fractional-order neural network, Represents the standard deviation of the training data input to the fractional-order neural network, Represents the th path weighting coefficient corresponding to the training data input to the fractional-order neural network;

[0048] (3) Adjust the weights and biases of each layer through the fractional-order gradient update strategy, and then adjust the learning rate to adapt to the global and local learning requirements. Among them, the calculation formula for weight update is as follows:

[0049] ,

[0050] ,

[0051] In the formula, Represents the weight update amount of the th layer of the fractional-order neural network, Represents the gradient of the weight of the th layer of the fractional-order neural network, represents the fractional-order control parameter, represents the adjustment coefficient for gradient balance, represents the learning rate of the fractional-order neural network;

[0052] (4) Improve the global search ability and training stability through a stochastic perturbation method based on Brownian motion. The calculation formula is as follows:

[0053] ,

[0054] ,

[0055] ,

[0056] where, represents the parameter update operation, represents the weight update amount of the th layer of the updated fractional-order neural network, represents the Brownian motion perturbation intensity, represents the stochastic perturbation amount of the th layer of the fractional-order neural network, represents the stochastic perturbation coefficient, represents the particle motion intensity of the Brownian motion principle;

[0057] (5) Adjust the decision boundary using the historical path through sampling and updating of the integral path global search mechanism. The calculation formula is as follows:

[0058] ,

[0059] ,

[0060] In the formula, represents the path integral result of the th iteration update, represents the path update intensity control parameter of the th iteration, represents the path integral result of the th iteration, represents the standard deviation of the th path of the fractional-order neural network, represents the entropy value of the path complexity, represents the regularization parameter, represents the first path weighted adjustment parameter, represents the second path weighted adjustment parameter;

[0061] (6) Calculate the classification category by integrating the multi-path decision-making brought by the global search mechanism of the integral path and the weight update brought by the fractional gradient update strategy. The calculation formula is as follows:

[0062] ,

[0063] In the formula, represents the classification category result of the fractional-order neural network, that is, the evaluation prediction result, represents the category index corresponding to taking the maximum value, , the classification categories include normal settlement, slight settlement, and severe settlement, represents normal settlement, represents slight settlement, represents severe settlement, represents the weight of the th category, represents the bias of the th category, is the feature of the th training sample of the fractional-order neural network,

[0064] represents the path weighting coefficient;

[0064] (7) Repeat the above steps (1)-(6) iteratively, and set the stopping iteration condition for the track foundation settlement evaluation prediction model until the model is trained when the stopping iteration condition is met.

[0065] The effects provided in the invention content are only the effects of the embodiments, rather than all the effects of the invention. The above technical solutions have the following advantages or beneficial effects:

[0066] (1) By adopting the fractional-order neural network modeling based on path integral, the present invention utilizes the principle of fractional calculus to initialize and update the gradient of the neural network parameters through non-integer-order derivatives, and solves the problem of insufficient diversity in parameter initialization of traditional networks when dealing with data with large fluctuations and noise interference;

[0067] In addition, the present invention adopts the global search mechanism of path integral to simulate various possible decision paths, effectively broadening the global search ability of the model for high-dimensional and non-linear electrical digital data, avoiding falling into local optima, and thus greatly improving the accuracy and stability of classification decisions;

[0068] Moreover, the present invention combines the principle of Brownian motion and adds random perturbations during the gradient update process, which helps the model jump out of local minima, accelerate convergence, and ensure that the overall training process is more robust.

[0069] (2) The present invention adopts a sample synthesis technique based on boundary conditions. Aiming at the problems of minority class samples and uneven data distribution in electrical digital data, representative seed samples are selected by minimizing the energy function, and the interpolation direction and amplitude are dynamically adjusted according to the boundary condition analysis (calculating the nearest neighbor distance and direction), so that the synthesized new samples are closer to the physical laws of actual track settlement;

[0070] In addition, the present invention uses a dynamically adjusted factor to control the sample synthesis strategy, avoiding over-density or sparsity of data in some areas, thereby improving the balance of the overall samples and further enhancing the robustness and generalization ability of the model. Brief Description of the Drawings

[0071] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention, and do not constitute a limitation to the present invention.

[0072] Figure 1 It is a schematic flow chart of the method of the present invention.

[0073] Figure 2 It is a comparison chart of the accuracy rates of different methods in track settlement evaluation.

[0074] Figure 3 It is a comparison chart of the performance indicators of different methods.

[0075] Figure 4 It is a comparison chart of the error distributions of different models.

[0076] Figure 5 It is a comparison chart of the interpolation directions of different methods.

[0077] Figure 6 It is a comparison chart of the distribution density of the feature space. Specific Embodiments

[0078] In order to clearly illustrate the technical features of the present solution, the present invention will be elaborated in detail below through specific embodiments and in combination with its drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below.

[0079] Embodiment 1

[0080] A method for evaluating track foundation settlement based on electrical digital data processing includes the following steps:

[0081] S1. Data collection: Collect the original data related to track foundation settlement, then process it, store the processed data, and manually label the stored data;

[0082] S2. Sample Synthesis: Based on the collected data, perform data augmentation, synthetic data generation, and sample diversity optimization operations, and then use a sample interpolation method based on boundary conditions to synthesize electrical digital data samples;

[0083] S3. Learning and Modeling: Establish an evaluation and prediction model for track foundation settlement. Preprocess the collected data and the synthetic data. Through feature engineering, extract key features based on the preprocessed data. Then, train the evaluation and prediction model for track foundation settlement according to the extracted key features to obtain the evaluation and prediction results. Finally, evaluate and improve the trained model;

[0084] Among them, the evaluation and prediction model for track foundation settlement uses a fractional-order neural network based on path integral:

[0085] S4. Suggestions and Warnings: Provide suggestions for settlement control and repair to users based on the evaluation and prediction results of the evaluation and prediction model for track foundation settlement, and determine whether to issue a warning according to the evaluation and prediction results.

[0086] S1 is specifically as follows:

[0087] Automatically collect data through sensor devices. The sensors include settlement detectors, accelerometers, and seismic wave detectors. The sensor devices are arranged at key positions of the track foundation to collect the settlement status information of the track in real time. The collected data related to track foundation settlement includes ground settlement data, vibration data, and surrounding environment change data. After the collected data is digitally processed by various sensor devices, it is transmitted in the form of discrete and accurate electrical digital data. The electrical digital data is in a digital format suitable for computer processing, such as binary form, decimal form, etc.;

[0088] Then, perform temporal processing and formatting processing on the data collected by various sensor devices, and remove the noise and interference in the data;

[0089] The processed data is transmitted to the storage system through wired or wireless communication methods, and manual annotation is performed on each electrical digital data input to the storage system. A unique identifier is assigned to each electrical digital data. The identifier is the grade of the settlement status of the track foundation and the annotation of data anomalies. The grades of the settlement status of the track foundation for the identifier are normal settlement, slight settlement, and severe settlement, ensuring the uniqueness and traceability of each electrical digital data. The format of the electrical digital data input to the storage system is CSV, JSON, XML, etc.;

[0090] The data storage format includes but is not limited to key information such as timestamp, sensor ID, sampling value, etc.;

[0091] For example, the data related to track foundation settlement collected is as follows:

[0092] Ra is a timestamp indicating the exact time of data acquisition (usually a time value accurate to seconds or milliseconds);

[0093] Da is the sensor ID, uniquely identifying the sensor from which the data is sourced;

[0094] Pa is the sensor installation location, describing the geographical location or coordinates of the sensor (such as the position in a coordinate system, the position in an orbital segment, etc.);

[0095] Va is the sampled value, representing the numerical value of the measurement result. For example, the displacement sensor may collect the position change value, and the acceleration sensor may collect the acceleration value;

[0096] Ta is the ambient temperature, describing the temperature condition of the data acquisition environment, which affects the accuracy of the sensor;

[0097] Ha is the ambient humidity, describing the humidity condition of the data acquisition environment;

[0098] Sa is the voltage signal strength, the voltage amplitude at the input end of the sensor, which affects the measurement accuracy of the sensor;

[0099] Fa is the measurement error, representing the error caused by the sensor accuracy or environmental interference;

[0100] Ca is the calibration status, indicating whether the sensor is in the correct calibration state (such as the calibration error is less than a certain threshold);

[0101] Ua is the noise factor, describing the impact of the possible noise in the signal on the measurement result.

[0102] It should be noted that this embodiment is only to illustrate a data format and type of the present invention. In actual applications, the attributes of the data usually have more than 10 attributes, and the number of data attributes may reach dozens or even hundreds.

[0103] S2 is specifically as follows:

[0104] Based on the collected data, additional training data samples are synthesized. The sample synthesis includes the following operations:

[0105] Data augmentation: Perform transformation operations on the collected data. The transformation operations include rotation, scaling, and translation, thereby synthesizing more variant data and improving the adaptability of the track foundation settlement evaluation and prediction model to different scenarios;

[0106] Synthetic data: Based on the characteristics of the collected data, use data augmentation methods to synthesize digital data samples similar to the collected data, thereby enriching the data samples for training;

[0107] Sample diversity optimization operation: By analyzing the collected data, adjust the distribution of the synthesized digital data samples so that the track foundation settlement assessment and prediction model can effectively handle different categories and edge cases in the data, further improving the accuracy and reliability of the prediction;

[0108] Adopt a sample interpolation method based on boundary conditions to synthesize digital data samples: Use the minority-class digital data samples as seed samples, optimize the sample selection by minimizing the energy function, making the selected samples more representative. At the same time, analyze the boundary conditions of the seed samples, calculate the distance and direction between them and the nearest neighbor samples, dynamically identify the boundary conditions, and precisely control the interpolation direction and amplitude of the subsequent digital data sample synthesis, making the synthesized digital data samples more representative and diverse, thereby improving the quality of digital data and the generalization ability of the model.

[0109] In the specific implementation manner, the operation of synthesizing digital data samples by the sample interpolation method based on boundary conditions is as follows:

[0110] (1) Select minority-class digital data samples as seed samples from the collected data according to the manually labeled types, and obtain the energy function for evaluating the quality of class digital data selection by minimizing the energy function. The calculation formula is as follows:

[0111] ,

[0112] In the formula, represents the energy function for evaluating the quality of class digital data selection, represents the set of selected minority-class digital data samples, represents the total number of digital data samples, represents the th feature vector of the digital data sample, represents the mean vector of the digital data sample features, represents the distribution density adjustment weight, set , represents the L2 norm, represents the boundary complexity adjustment weight, set , represents the th average distance from the feature vector of the digital data sample to its nearest neighbor sample set, represents the nearest neighbor sample set;

[0113] Among them, the energy function comprehensively considers the distribution density and boundary complexity of the electro-digital data samples in the feature space, enabling the selected minority-class electro-digital data samples to represent key regions, thereby improving the diversity of the subsequently synthesized electro-digital data samples. For example, the signals from ground displacement sensors may be less in the tiny deformation regions of the track foundation, while the sampling data of strain sensors and accelerometers are more in the severe settlement regions. Traditional methods are prone to ignoring these low-density regions during data augmentation. However, this method ensures that the selected seed samples cover different states of the track foundation (such as micro-settlement, stability, and severe settlement) by minimizing the energy function, thereby improving the effectiveness of data augmentation and ensuring a more uniform sample synthesis effect even in the case of unbalanced electro-digital data distribution, thus enhancing the robustness of the classification model.

[0114] (2) By calculating the distance and direction between the seed samples and their nearest neighbor samples, identifying the dynamic boundary conditions, and conducting boundary condition analysis on each seed sample, the local structural characteristics of the electro-digital data sample distribution are further determined. The calculation formula is as follows:

[0115] ,

[0116] ,

[0117] In the formula, represents the dynamic adjustment vector based on the boundary conditions, represents the feature vector of the th electro-digital data sample, represents the set of nearest neighbor samples of the th electro-digital data sample, represents the weight between the th electro-digital data sample and the th electro-digital data sample, which is used to measure the importance between the th electro-digital data sample and its neighboring electro-digital data samples, represents the attenuation parameter, set , represents the feature vector of the th electro-digital data sample, , , represent the total number of electro-digital data samples three different indices;

[0118] By performing boundary condition analysis on each seed sample, it is convenient to precisely control the interpolation direction and interpolation amplitude in subsequent synthesis of electro-digital data samples. For example, the displacement signal of a ground displacement sensor may change smoothly, while the signals of strain sensors and accelerometers may exhibit sudden changes or periodic oscillations. Traditional data augmentation methods often use simple linear interpolation, which easily leads to the deviation of the synthesized data distribution from the true track state. By calculating the dynamic boundary adjustment vector of the seed sample, the present invention can capture the change trends of track settlement data in different regions. In the region where the track foundation is stable, the displacement signal is relatively gentle, and the interpolation amplitude should be small. While in the region of severe settlement, the interpolation amplitude should be larger to ensure that the synthesized data can reflect the true characteristics of the track state and avoid the problem that the samples synthesized by traditional interpolation methods do not conform to the true track settlement pattern, thereby improving the rationality of data augmentation.

[0119] (3) Through the interpolation method and the dynamic adjustment vector based on boundary conditions, dynamically adjust the interpolation path and the quantity and direction of the synthesized electro-digital data samples. The calculation formula is as follows:

[0120] ,

[0121] ,

[0122] In the formula, represents the newly synthesized electro-digital data sample, represents the interpolation coefficient, represents a random number with a value between 0 and 1, represents the dynamic adjustment factor;

[0123] Through dynamic adjustment, the limitations of traditional linear interpolation methods can be improved, and the local electro-digital data distribution can be adaptively augmented. For example, in track settlement monitoring, the data distribution of accelerometers usually has a large discreteness, and simple interpolation may cause the synthesized samples to deviate from the actual track state. This method ensures that the augmented data can more realistically reflect the track settlement changes by dynamically adjusting the interpolation direction and amplitude. If the data points collected by a ground displacement sensor show that the track foundation has a large deformation in a certain area, while the change trend of the nearest neighbor samples is relatively slow, then the newly synthesized interpolation samples will be interpolated along the main direction of track deformation according to the distribution characteristics of their neighboring samples, without deviating from the true track state, ensuring the rationality and diversity of data augmentation.

[0124] (4) Modify the synthesis strategy according to the influence of the newly synthesized point digital data samples on the overall electro-digital data sample distribution, and calculate the dynamic adjustment factor during the synthesis process of electro-digital data samples. The calculation formula is as follows:

[0125] ,

[0126] In the formula, represents the dynamic adjustment factor, represents the adjustment rate parameter, represents the number of dimensions of the electrical digital data sample, represents the current synthesized electrical digital data sample at the th dimension of the distribution probability, represents the target electrical digital data sample at the th dimension of the distribution probability, represents the operation of taking the minimum value;

[0127] By calculating the dynamic adjustment factor, it is possible to avoid excessive or insufficient expansion in certain regions and maintain the balance of the electrical digital data distribution. A key problem with traditional data augmentation methods is that they may lead to too high a data density in some regions and too little data in other regions, thus affecting the learning effect of the model. For example, if there are few abnormal track foundation settlements in a certain region of the electrical digital data, and the traditional data augmentation method simply copies the existing samples, then the model may be affected by the unbalanced data distribution during training, resulting in an increase in the misclassification rate. This method calculates the deviation between the newly synthesized samples and the target distribution, dynamically adjusts the interpolation strategy, so that the new samples can be more evenly distributed in the entire feature space. If the settlement data in some track regions is overly concentrated during sample synthesis, then by reducing the interpolation amplitude in this region or increasing the interpolation ratio in other regions, the data distribution can be made more balanced and the generalization ability of the model can be improved;

[0128] Finally, repeat the steps of sample interpolation method based on boundary conditions for electrical digital data sample synthesis multiple times until the number of newly synthesized sample data reaches the preset number threshold.

[0129] S3 is specifically as follows:

[0130] Construct an evaluation and prediction model for track foundation settlement, and input the collected data and the synthesized data into the evaluation and prediction model for track foundation settlement for training. The specific operations are as follows:

[0131] Data preprocessing: Read the electrical digital data in the storage system and the electrical digital data samples after sample synthesis, and perform preprocessing operations on the data. The preprocessing operations include filling in missing values, removing outliers, and data standardization;

[0132] Extract key features: Extract the key features during the track foundation settlement process through feature engineering. The key features include foundation compaction degree, soil moisture, and temperature;

[0133] Training of the track foundation settlement evaluation and prediction model: The model is trained through a fractional-order neural network based on path integration. Specifically, various possible decision paths are simulated through path integration to conduct a global search for the parameters of the fractional-order neural network, and random perturbations are carried out during the training process with the help of the Brownian motion principle. Finally, the evaluation and prediction results of the track foundation settlement are obtained;

[0134] Evaluation and improvement of the track foundation settlement evaluation and prediction model: Compare the training results with the actual results, where the actual results are the manually labeled identifiers, evaluate the accuracy of the track foundation settlement evaluation and prediction model, and adjust and improve the track foundation settlement evaluation and prediction model according to the evaluation results.

[0135] In the specific implementation manner, the training of the track foundation settlement evaluation and prediction model is as follows:

[0136] (1) Initialize the weights and biases of the fractional-order neural network through non-integer-order derivatives in fractional calculus, which can solve the problem of insufficient diversity in parameter initialization of traditional fractional-order neural networks, especially when there are large fluctuations and noises in the data (such as sudden changes in track settlement), and better capture the complexity of track settlement data. The calculation formula is as follows:

[0137] ,

[0138] ,

[0139] ,

[0140] In the formula, represents the key feature of the th training data input to the fractional-order neural network, represents the key feature of the th training data input to the fractional-order neural network, represents the total number of training data input to the fractional-order neural network in the current batch, represents the regularization parameter of the fractional-order neural network, set , represents the th layer weight of the fractional-order neural network, represents the th layer bias of the fractional-order neural network. The fractional-order neural network has a total of layers, , represents a symbol subject to a specific distribution, represents the standard normal distribution, represents the fractional-order control parameter, is the standard deviation for the initialization of the fractional-order neural network, set to ;

[0141] Among them, several data included in the key features, such as the sensor installation position Pa, the sampling value Va, the ambient temperature Ta, the ambient humidity Ha, the voltage signal strength Sa, the measurement error Fa, the calibration state Ca, the noise factor Ua, etc.;

[0142] (2) Adopt the integral path global search mechanism to optimize the fractional-order neural network in the global parameter space. Through dynamic path weighting, according to the similarity between different input training data and the decision boundary, adjust the contributions of different paths. The calculation formula of the integral path global search mechanism is as follows:

[0143] ,

[0144] ,

[0145] In the formula, represents the key feature of the th training data input into the fractional-order neural network, represents the calculation result of the integral path global search mechanism, represents the mean value of the training data input into the fractional-order neural network, represents the standard deviation of the training data input into the fractional-order neural network, represents the th path weighting coefficient corresponding to the training data input into the fractional-order neural network;

[0146] By adopting the integral path global search mechanism, the global search ability for non-linear and high-dimensional electrical digital data is enhanced through the simulation of multiple paths, solving the problem of being easily trapped in local optima, enabling the fractional-order neural network to optimize in the global parameter space. Electrical digital data usually exhibits complex non-linear relationships. For example, the track settlement process is affected by multiple factors, and the signals provided by each sensor are also high-dimensional. The integral path global search mechanism can prevent the model from falling into local optima and help the neural network find a more suitable solution in the global parameter space. In this case, through dynamic path weighting, the model can adjust the contributions of different paths according to the similarity between different data samples and the decision boundary, thus more accurately reflecting the true state of track settlement;

[0147] (3) Adjust the weights and biases of each layer through the fractional-order gradient update strategy, and then adjust the learning rate to adapt to the global and local learning requirements. Among them, the calculation formula for weight update is as follows:

[0148] ,

[0149] ,

[0150] In the formula, represents the weight update amount of the th layer of the fractional-order neural network, represents the gradient of the weight of the th layer of the fractional-order neural network, represents the fractional-order control parameter, represents the adjustment coefficient for gradient balance, set , represents the learning rate of the fractional-order neural network, set = 0.3;

[0151] Adopt the fractional-order gradient update strategy, adjust the weights and biases of each layer through fractional-order differentiation, avoid the problem that it is difficult to balance the global and local fineness of the learning rate when the fractional-order neural network processes complex non-linear digital data, flexibly adjust the convergence speed under diverse distributions. In digital data, different data features of sensors (such as acceleration changes or irregular displacement fluctuations) may require different learning rates to adapt to the global and local learning needs. The fractional-order gradient update can flexibly adjust the learning rate, avoid the problems of gradient disappearance or explosion that may occur in traditional methods, and accelerate the convergence speed of the model;

[0152] (4) Improve the global search ability and training stability through the random perturbation method based on Brownian motion. The calculation formula is as follows:

[0153] ,

[0154] ,

[0155] ,

[0156] Among them, represents the parameter update operation, represents the weight update amount of the th layer of the updated fractional-order neural network, represents the Brownian motion perturbation intensity, set , represents the random perturbation amount of the th layer of the fractional-order neural network, represents the random perturbation coefficient, set , represents the particle motion intensity of the Brownian motion principle, ;

[0157] Adopt a random perturbation method based on Brownian motion. Brownian motion refers to the incessant random motion of particles suspended in a liquid or gas. Referring to this principle, the weights are randomly adjusted after each gradient update to solve the problems of getting stuck in local minima and unstable gradients, improve the global search ability and training stability. In the task of track settlement assessment, digital data may be affected by noise. Especially during real-time acquisition, the accuracy of sensor data may be interfered by the external environment. The Brownian motion method can help the model escape from local minima through random perturbation, improve its global search ability, and enable the model to train more stably when facing complex and irregular settlement patterns;

[0158] (5)Through the sampling and update of the integral path global search mechanism, use the historical path to adjust the decision boundary. The calculation formula is as follows:

[0159] ,

[0160] ,

[0161] In the formula, represents the path integral result of the th iteration update, represents the path update intensity control parameter of the th iteration, represents the path integral result of the th iteration, represents the standard deviation of the th path of the fractional-order neural network, represents the entropy value of the path complexity, represents the regularization parameter, represents the first path weighted adjustment parameter, set , represents the second path weighted adjustment parameter, set ;

[0162] Through the sampling and update of path integration, use the historical path to adjust the decision boundary, solve the balance problem of classification decision in global optimization and local adaptability, continuously correct the classification boundary, and improve the overall robustness. For different types of track settlement, the model can perform multi-path adaptive adjustment based on the input data of the sensor, not only considering global information, but also dynamically adjusting the local classification boundary, making the evaluation results more accurate;

[0163] (6)Integrate the multi-path decision-making brought by the integral path global search mechanism and the weight update brought by the fractional-order gradient update strategy to calculate the classification category. The calculation formula is as follows:

[0164] ,

[0165] In the formula, represents the classification category result of the fractional-order neural network, that is, the evaluation prediction result, represents the category index corresponding to taking the maximum value, The classification categories include normal settlement, slight settlement, and severe settlement, represents normal settlement, represents slight settlement, represents severe settlement, represents the weight of the th category, represents the bias of the th category, is the feature of the th training sample of the fractional-order neural network, ;

[0166] Combining the multi-path decision-making brought by path integration and the flexible weight update brought by fractional calculus, the calculation of classification categories is carried out. In the task of track foundation settlement evaluation, the digital data of different positions and different sensors may show different distributions. Traditional classifiers may have difficulty dealing with these complex feature distributions. The multi-path decision-making mechanism of path integration enables the model to dynamically adjust the classification decision boundary, improving the overall classification accuracy and robustness in the face of complex or extreme settlement situations;

[0167] (7) Repeat the above steps (1)-(6) iteratively, and set the stop iteration condition for the track foundation settlement evaluation prediction model until the model completes training when the stop iteration condition is met;

[0168] The preset stop iteration condition is to reach the preset maximum number of iterations. In the present invention, the preset maximum number of iterations is set to 1000 times.

[0169] Embodiment 2

[0170] As Figure 2As shown in the figure, to verify the effectiveness of the technical solution of the present invention and evaluate the accuracy performance of different methods in the task of track settlement evaluation, accuracy is the core index to measure the performance of the classification model, which directly reflects the model's ability to identify the track settlement state. During the experiment, the accuracies of different methods in multiple test rounds were sampled and a line chart was plotted. The experimental results show that the accuracy of the fractional-order neural network based on path integral always stabilizes between 92% and 95%, while the accuracies of traditional neural network, support vector machine and decision tree are between 85% - 88%, 82% - 86% and 80% - 85% respectively, and show large fluctuations, indicating that this technology is not only significantly superior to traditional methods in terms of accuracy, but also has stronger stability and can classify the track settlement state more reliably. The reason is that the path integral method can simulate multiple possible decision paths, thus avoiding the local optimum problem caused by the non-linear characteristics of data and making the classification decision more comprehensive and accurate.

[0171] Example 3

[0172] As Figure 3 shown, to further verify the overall performance of the model, the advantages and disadvantages of different methods were measured by the relationship between accuracy and error. The experiment used a bar chart to show the accuracy and error distributions of different methods respectively for intuitive comparison. The experimental results show that the method of the present invention has the highest accuracy (0.95) and the lowest error (0.05); in contrast, the accuracy of the traditional neural network is 0.88 and the error is 0.12, the accuracy of the support vector machine is 0.85 and the error is 0.15, and the accuracy of the decision tree is the lowest, 0.83, while the error is the largest, 0.17, further indicating that the method of the present invention can not only evaluate the track settlement state more accurately, but also effectively control the classification error and more accurately adapt to different modes of track settlement.

[0173] Example 4

[0174] As Figure 4As shown, to further analyze the distribution of model errors and verify the stability of different methods when facing real data, traditional error analysis methods may only focus on the average error. However, in the task of track settlement assessment, the fluctuation range of errors is equally important. The KDE kernel density estimation graph is used to visualize the error distributions of different methods and observe the degree of error concentration. The experimental results show that the error distribution of the fractional-order neural network based on the integral path adopted in the present invention is the most concentrated, with most error values concentrated around 0.05, while the error distributions of other methods are significantly more dispersed. The average error of the traditional neural network is about 0.12, the average error of the support vector machine is 0.15, and the error distribution of the decision tree is the most discrete, with an average error reaching 0.17, proving that the method of the present invention can provide a more stable prediction effect when facing complex and high-dimensional electrical digital data and avoid the problem of excessive classification uncertainty.

[0175] Example 5

[0176] As Figure 5 shown, to verify that the effect of the dynamic adjustment interpolation method of the present invention is superior to the traditional linear interpolation method, an experiment is conducted. From the experimental results, it can be seen that the interpolation direction of the traditional method randomly diverges, while the present invention makes the interpolation direction (blue arrow) consistent with the main direction of real track deformation through boundary condition analysis, ensuring that the synthesized samples conform to the actual physical laws. Therefore, it is better to adopt the dynamic adjustment interpolation method.

[0177] Example 6

[0178] As Figure 6 shown, the present invention also conducts a comparative experiment on the distribution density of different methods in the feature space. The experimental results show that the synthesized samples of the traditional method are concentrated in the middle of the feature space, while the present invention successfully expands the sample coverage in the low-density regions (both sides) while maintaining the main distribution, verifying the effectiveness of the boundary condition analysis.

[0179] Although the specific implementation manners of the invention are described above in conjunction with the drawings, it is not a limitation to the protection scope of the present invention. Based on the technical solutions of the present invention, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present invention.

Claims

1. A method for evaluating track foundation settlement based on electronic digital data processing, characterized in that: S1. Data collection: collect the original data related to track foundation settlement, process it, store it, and manually annotate it; S2, sample synthesis: based on the collected data, data enhancement, data synthesis, sample diversity optimization operations are performed, and then the sample interpolation method based on boundary conditions is used to synthesize the electrical digital data samples; S3. Learning modeling: Establish a track foundation settlement assessment prediction model, pre-process the collected data and synthesized data, extract key features from the pre-processed data through feature engineering, and then train the track foundation settlement assessment prediction model based on the extracted key features to obtain the assessment prediction results, and then evaluate and improve the trained model; Among them, the track foundation settlement assessment prediction model adopts a fractional-order neural network based on path integral; S3 is as follows: Construct a track foundation settlement assessment prediction model, input the collected data and synthesized data into the track foundation settlement assessment prediction model for training. The specific operations are as follows: Data preprocessing: Read the electrical digital data in the storage system and the electrical digital data samples synthesized by the samples, and perform preprocessing operations on the data. The preprocessing operations include supplementing missing values, removing outliers, and standardizing data; Extract key features: Extract key features of the track foundation settlement process through feature engineering, including foundation compaction, soil moisture, and temperature; Training of the track foundation settlement assessment prediction model: The model is trained through a fractional-order neural network based on path integrals. Specifically, multiple possible decision paths are simulated through path integrals, and a global search of the fractional-order neural network parameters is performed. Random perturbations are performed during the training process using the principle of Brownian motion, and finally the assessment and prediction results of the track foundation settlement are obtained. Evaluation and improvement of the track foundation settlement assessment prediction model: Compare the training results with the actual results, where the actual results are the manually labeled identifiers, evaluate the accuracy of the track foundation settlement assessment prediction model, and adjust and improve the track foundation settlement assessment prediction model based on the evaluation results; S4. Recommendations and warnings: Based on the evaluation and prediction results of the track foundation settlement evaluation and prediction model, provide users with suggestions on settlement control and repair, and determine whether to issue a warning based on the evaluation and prediction results; The operation of synthesizing electrical digital data samples by the sample interpolation method based on boundary conditions is as follows: (1) Select a few quasi-electric digital data samples from the collected data as seed samples, and minimize the energy function to obtain an energy function for evaluating the quality of quasi-electric digital data selection; (2) By calculating the distance and direction between the seed sample and the nearest neighbor sample, the dynamic boundary conditions are identified, and the boundary condition analysis is performed on each seed sample to determine the local structural characteristics of the distribution of the electrical digital data sample; (3) Dynamically adjust the number and direction of the interpolation path and the synthesized electrical digital data samples through interpolation methods and dynamic adjustment vectors based on boundary conditions; (4) modifying the synthesis strategy according to the influence of the newly synthesized electrical digital data samples on the overall electrical digital data sample distribution, and calculating the dynamic adjustment factor during the electrical digital data sample synthesis process; The step of synthesizing the electrical digital data samples by the sample interpolation method based on the boundary condition is repeated multiple times until the number of synthesized new sample data reaches a preset number threshold.

2. A track foundation settlement assessment method based on electronic digital data processing according to claim 1, characterized in that: S1 is as follows: Automatic data collection is performed through sensor equipment. The sensors include settlement detectors, accelerometers and seismic wave detectors. The sensor equipment is set at key positions of the track foundation to collect track settlement status information in real time. The collected track foundation settlement related data include ground settlement data, vibration data and surrounding environment change data. The collected data is digitally processed when passing through various sensor equipment and transmitted in the form of discrete and accurate electrical digital data. The electrical digital data is in a digital format suitable for computer processing; Then, the data collected by various sensor devices are processed in time series and formatted, and noise and interference in the data are removed; The processed data is transmitted to the storage system via wired or wireless communication, and each piece of electrical digital data input into the storage system is manually labeled, and a unique identifier is assigned to each piece of electrical digital data. The identifier is the level of settlement status of the track foundation and the data abnormality label. The identifier is the level of settlement status of the track foundation, which is normal settlement, slight settlement and severe settlement.

3. A track foundation settlement assessment method based on electronic digital data processing according to claim 2, characterized in that: S2 is as follows: Based on the collected data, additional training data samples are synthesized. Sample synthesis includes the following operations: Data enhancement: Transform the collected data, including rotation, scaling, and translation; Synthetic data: Based on the characteristics of the collected data, the data augmentation method is used to synthesize electrical digital data samples similar to the collected data; Sample diversity optimization operation: by analyzing the collected data, the distribution of synthesized electrical digital data samples is adjusted; A sample interpolation method based on boundary conditions is used to synthesize electrical digital data samples: a minority class of electrical digital data samples is used as seed samples, and sample selection is optimized by minimizing the energy function. At the same time, boundary condition analysis is performed on the seed samples, and the distance and direction between the seed samples and the nearest neighbor samples are calculated to dynamically identify boundary conditions.

4. A track foundation settlement assessment method based on electronic digital data processing according to claim 3, characterized in that: The training of the track foundation settlement assessment prediction model is as follows: (1) The weights and biases of the fractional neural network are initialized by using non-integer derivatives in fractional calculus. The calculation formula is as follows: ， ， ， In the formula, represents the first The key features of the training data are represents the first The key features of the training data are Represents the total number of training data input to the fractional-order neural network in the current batch, represents the regularization parameter of the fractional-order neural network, represents the fractional-order neural network Layer weights, represents the fractional-order neural network Layer bias, fractional order neural network layer, , The symbol for a particular distribution is represented by represents the standard normal distribution, represents the fractional-order control parameter, The standard deviation for initializing the fractional neural network; (2) The integral path global search mechanism is used to optimize the fractional-order neural network in the global parameter space. Through dynamic path weighting, the contribution of different paths is adjusted according to the similarity between different input training data and the decision boundary. The calculation formula of the integral path global search mechanism is as follows: ， ， In the formula, represents the first The key features of the training data are represents the calculation result of the global search mechanism of the integral path, represents the mean of the training data input to the fractional neural network, represents the standard deviation of the training data input to the fractional neural network, represents the first The path weight coefficient corresponding to the training data; (3) The weights and biases of each layer are adjusted through the fractional gradient update strategy, and then the learning rate is adjusted to adapt to the global and global learning needs. The calculation formula for weight update is as follows: ， ， In the formula, Represents the fractional order neural network The weight update amount of the layer, Represents the fractional order neural network The gradient of the layer's weights, represents the fractional-order control parameter, represents the adjustment coefficient of gradient balance, Represents the learning rate of the fractional-order neural network; (4) The global search capability and training stability are improved by using a random perturbation method based on Brownian motion. The calculation formula is as follows: ， ， ， in, Indicates a parameter update operation. represents the updated fractional-order neural network The weight update amount of the layer, represents the Brownian motion disturbance intensity, Represents the fractional order neural network The random perturbation of the layer, represents the random perturbation coefficient, The intensity of particle motion that represents the Brownian motion principle; (5) Through sampling and updating of the integral path global search mechanism, the decision boundary is adjusted using the historical path. The calculation formula is as follows: ， ， In the formula, Indicates The path integral result of the iteration update is: Indicates The path update intensity control parameter for the iteration, Indicates The path integral result of the iteration is: represents the fractional-order neural network The standard deviation of the paths, The entropy value representing the complexity of the path, represents the regularization parameter, represents the weighted adjustment parameter of the first path, represents the weighted adjustment parameter of the second path; (6) The multi-path decision brought by the global search mechanism of the integrated path and the weight update brought by the fractional gradient update strategy are used to calculate the classification category. The calculation formula is as follows: ， In the formula, Represents the classification category result of the fractional-order neural network, that is, the evaluation prediction result, Indicates the category index corresponding to the maximum value. , the classification categories include normal settlement, slight settlement and severe settlement, Indicates normal settlement. Indicates slight subsidence. Indicates severe subsidence. Indicates The weight of each category, Indicates The bias of the categories, is the fractional order neural network The features of the training samples, represents the path weight coefficient; (7) Repeat the above steps (1) to (6) and set a stop iteration condition for the track foundation settlement assessment prediction model until the model training is completed when the stop iteration condition is met.

Images