Intelligent railway ballast track mud pumping judgment method

Abstract

The invention discloses an intelligent railway ballast track frost boiling and mud rising distinguishing method, and relates to the technical field of railway intelligent detection.The method includes the steps that multi-source data parameters are monitored, a high-dimensional feature space is built, intelligent distinguishing is conducted through a random forest, firstly, test data are collected, normalization and time sequence analysis preprocessing are conducted, and a high-dimensional feature space is obtained; and screening key features according to importance of model features, training a classification model by using annotation data, optimizing the model through multiple methods, and inputting real-time monitoring data to obtain the frost boiling and mud pumping state and possibility distribution. Compared with a traditional method, the method is high in precision, high in robustness and higher in environmental adaptability; the method can also be expanded to other engineering disease identification, a new idea is developed for intelligent monitoring of the railway ballast bed, technical support is provided, the railway operation and maintenance level and safety are powerfully improved, and the method has wide application prospects.

Description

Technical Field

[0001] The invention belongs to the cross-technical field of railway engineering and artificial intelligence, and in particular relates to an intelligent method for distinguishing mud bubbling on ballasted railway tracks. Background Art

[0002] Railway ballasted tracks are prone to mud and slurry due to long-term exposure to train loads and environmental factors. This phenomenon not only leads to a decline in the performance of the ballast track structure, but also poses a serious threat to the safety of train operation. The formation mechanism of mud and slurry is complex, involving multiple dynamic processes such as soil moisture migration, pore water pressure changes, and mud intrusion. Its influencing factors include dynamic loads, moisture content, freeze-thaw cycles, etc. However, the existing methods have the following problems in monitoring and distinguishing mud and slurry: First, the phenomenon is not easy to observe due to ballast covering. The ballast layer covers the surface of the roadbed, making it difficult to directly monitor the mud and slurry process, especially in the initial stage. It is difficult to find problems in time, which brings hidden dangers to track maintenance; second, a single indicator is difficult to effectively distinguish. The existing technologies mostly rely on a single indicator for distinction, but the formation of mud and slurry is affected by a combination of factors. A single indicator is not enough to accurately reflect the complex mechanism, resulting in misjudgment or missed judgment; third, there is a lack of real-time observation means. Most of the existing technologies are based on offline data analysis, which cannot realize real-time monitoring and dynamic distinction of the mud and slurry process, and it is difficult to meet the requirements of rapid response in railway operations.

[0003] Chinese patent CN114818998A discloses a method for distinguishing mud and slurry. For a ballastless track line that has suffered from mud and slurry, the location of the mud and slurry on the line is marked and the disease level is divided; a track state detection operation is carried out on the line by a high-speed comprehensive detection train to obtain track state data with mud and slurry location, different mud and slurry levels and development trend of low-level mud and slurry, and the track height and vehicle body vertical acceleration data most affected by the mud and slurry are selected as training data; a ballastless track mud and slurry state discrimination network is constructed.

[0004] However, the above disclosed prior art still has many shortcomings. Summary of the invention

[0005] The problem to be solved by the present invention is how to provide a method for distinguishing mud bubbling on railway ballasted track more quickly, accurately and adapting to more complex situations.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: an intelligent railway ballasted track mud bubbling identification method, the key technology of which is that it includes the following steps:

[0007] 1) Data collection: Use the automated data collection equipment installed in the track bed and subgrade to monitor the data in real time, and transmit the collected data to the cloud processing platform in real time through the wireless communication module;

[0008] 2) Data preprocessing: The cloud processing platform preprocesses the collected raw data and performs feature extraction;

[0009] 3) Model training: Use historical test data and annotated mud-swelling states to build a data set to train a random forest classification model, and then optimize the model;

[0010] 4) Real-time discrimination: The real-time collected monitoring data is input into the trained random forest model, the mud-bubbling state and its probability distribution are output, and the model is optimized through a real-time feedback mechanism.

[0011] Preferably, the data collection step includes the following contents:

[0012] Preferably, the status information of the railway ballasted track is collected by arranging sensors at key positions of the ballast bed and the subgrade, including moisture content sensors, pore water pressure sensors, soil pressure gauges, temperature sensors and smart particles; the sampling frequency of the sensors is set to 0.01 Hz to 10 Hz; the collected information includes moisture content, pore water pressure, soil pressure, and the movement state and vibration response of ballast particles.

[0013] Preferably, the data preprocessing step first performs outlier detection, missing value filling and normalization on the original data, and then uses an algorithm to extract features, including introducing the liquid and plastic limits of the soil, and further calculating the volume moisture content of the soil to obtain features, which are specifically manifested as:

[0014]

[0015] Among them, W t is the input feature, W is the monitoring data, W L is the soil liquid limit, W P is the plastic limit of the soil; the gradient of the pore water pressure information is calculated and the gradient is normalized as a feature, which is specifically expressed as follows:

[0016]

[0017] Among them, Δu is the rate of change of pore water pressure, u(t) and u(t+Δt) represent the pore water pressure at time t and t+Δt respectively, and Δt is the sampling interval. The effective stress is further obtained from the pore water pressure and soil pressure as a feature. The specific process is as follows:

[0018]

[0019] Among them, Gu is the normalized pore water pressure gradient, Δu is the change of pore water pressure in the height direction, and Δz is the corresponding height change; the original value of the monitored temperature is used as the feature; the sliding window method is used to process the original data collected by the intelligent particles to obtain the vibration acceleration characteristics of the ballast.

[0020] Preferably, the data set constructed in the model training, wherein the training set and the test set account for 70% and 30% of the total data respectively;

[0021] Preferably, features are screened during the model training, multiple decision trees are constructed through a random forest algorithm to cover multi-dimensional data for centralized learning, importance scores are performed on features involved in decision-making during training, and key features are screened out for discriminant model training.

[0022] Preferably, the hyperparameters of the random forest model, including the number of decision trees, the maximum tree depth, and the minimum number of split samples, are optimized by using a grid search method; and the classification accuracy, recall rate, and F1 value are calculated by cross-validation.

[0023] Preferably, in the real-time discrimination step, the model outputs two states of mud bubbling and bubbling: normal and mud bubbling, and is accompanied by a state possibility distribution.

[0024] Preferably, the method can compare the real state based on the real-time output result and perform feedback adjustment in combination with the monitored data.

[0025] Compared with the prior art, the present invention is beneficial in that:

[0026] The present invention solves the observation problem caused by ballast coverage through multi-source sensor data collection, and fully reflects the multi-dimensional characteristics of mud bubbling;

[0027] The present invention uses a random forest model combined with multi-feature data analysis to integrate the features of multi-source data and uses multi-dimensional features for discrimination, which significantly improves the accuracy of state discrimination.

[0028] The state classification and probability distribution output of the present invention provide a basis for subsequent maintenance classification, avoid the risk of manpower waste or untimely maintenance, and provide data support for the prevention of abnormal conditions;

[0029] The present invention provides support for the present invention to complete the real-time monitoring and prediction tasks under the action of freeze-thaw cycles, rainfall and dynamic loads through the model self-optimization mechanism, and provides the possibility of expansion to the identification of other disease types under complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A specific flow chart of an intelligent railway ballasted track mud bubbling identification method proposed by the present invention;

[0031] Figure 2 This is a schematic diagram of the arrangement of sensors in the data collection method proposed by the present invention;

[0032] Figure 3 This is a flow chart of the ballast particle vibration acceleration feature extraction proposed in the present invention.

[0033] Specific Examples

[0034] In order to make the features and advantages of the present invention more obvious and easy to understand, the technical solution will be clearly and completely described below in combination with the drawings and specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0035] The present invention proposes an intelligent railway ballasted track mud bursting and mud emitting identification method.

[0036] like Figure 1 First, multi-source data is collected on the ballasted track, and multi-source sensors are arranged at key positions of the roadbed and subgrade, including moisture content, pore water pressure and other types. The sampling frequency is set at 0.01Hz-10Hz according to the dynamics of the train and the environment, and the data is transmitted to the cloud in real time via wireless communication to build a data perception foundation.

[0037] After preprocessing multi-source data and receiving the data on the cloud, the data is regularized through outlier detection, missing value filling and normalization. Then, key features such as pore water pressure gradient normalization are automatically extracted based on the physical characteristics of mud bubbling, thereby improving data quality and discrimination value.

[0038] The discriminant model was trained with 70% of the historical data as the training set and 30% as the test set to train the random forest model. Grid search was used to optimize the number of decision trees, depth and other hyperparameters. Based on the random forest algorithm, at least 5 key features were selected from multidimensional data according to feature importance, such as the rate of change of pore water pressure, etc. This provided core elements for model training and enhanced model performance. Cross-validation was used to adjust the model according to indicators such as accuracy to ensure accurate model discrimination.

[0039] Real-time identification: real-time data is input into the trained model to output the mud-bubbling status and possibility distribution. The model is optimized based on real-time feedback of deviation to adaptively improve the identification accuracy.

[0040] In the implementation process of the present invention, the real-time collection of key data in the railway ballasted track is realized by reasonably deploying sensors, providing an accurate data basis for the subsequent identification of mud bubbling and mud bubbling. Figure 2As shown in the figure, the volumetric moisture content and temperature sensors are arranged at different depths of the test soil samples of the ballast bed to monitor the moisture and temperature changes of the soil at different depths. The sensors need to be in close contact with the soil to avoid data deviation; the pore water pressure and soil pressure sensors are arranged at the same depth as the moisture content sensor to monitor the pore water pressure and soil pressure changes inside the soil in real time to ensure the correspondence and uniformity of the data. The smart particles are arranged at the bottom and middle of the ballast layer to monitor the movement state and vibration response of the ballast particles.

[0041] All sensors are connected to the data acquisition module. The sampling frequency of the sensor is set to 0.01Hz to 10Hz. The data acquisition module must have a multi-channel input function and be connected to the cloud processing platform to achieve high-frequency sampling and storage of multi-source data to provide data support for subsequent operations.

[0042] The data in the data acquisition module is transmitted to the cloud processing platform in real time through the wireless communication module, and stored by time classification to form time series data, providing support for subsequent preprocessing and model input.

[0043] The data preprocessing of the present invention is an important step to ensure that the model can accurately identify the state of mud boiling. By introducing preprocessing methods related to physical mechanisms, including calculation of liquid and plastic limits, time series analysis, data normalization and other processing methods, the present invention can effectively improve the data quality and further extract meaningful features from the data in the sensor. The following are the detailed implementation steps of data preprocessing and feature extraction:

[0044] (1) In the present invention, since temperature itself is an important environmental factor affecting the mud and oozing phenomenon, the temperature data at different depths directly use their original values ​​as input features. This data reflects the influence of external thermal changes on the soil and can provide useful feature data for the model without further complex processing, which can be used as a feature for mud and oozing identification.

[0045] (2) In the present invention, the liquid and plastic limits are introduced as input features for the water content processing at different depths. During the occurrence of mud bubbling, the liquid and plastic limits can effectively indicate the mud bubbling behavior of the soil under different wet conditions. Therefore, the water content data is further calculated to obtain the liquid and plastic limits as feature inputs, and the equation is:

[0046]

[0047] Where Wt is the input feature, W is the monitoring data, WL is the liquid limit of soil, and WP is the plastic limit of soil.

[0048] (3) In the implementation process of the present invention, the collected raw data of pore water pressure and soil pressure at different depths need to be further processed to extract the rate of change of pore water pressure that can accurately reflect the dynamic response of the soil. The accumulation and dissipation of pore water pressure has a great influence on the pumping of mud during the mud pumping process. The rate of change of pore water pressure is used to characterize the dynamic change characteristics of pore water pressure over time, and its calculation formula is:

[0049]

[0050] Where Δu is the rate of change of pore water pressure, u(t) and u(t+Δt) represent the pore water pressure at time t and t+Δt, respectively, and Δt is the sampling interval.

[0051] (4) The pore water pressure gradient is used to reflect the distribution law of the pore water pressure inside the soil along the height direction. It can effectively characterize the hydraulic gradient in the soil and its influence on the migration of fine particles, and provide a scientific explanation for mud intrusion. In order to eliminate the influence of dimension, the normalization method is used to calculate the normalized value of the pore water pressure gradient. The calculation formula is:

[0052]

[0053] Where Gu is the normalized pore water pressure gradient, Δu is the change in pore water pressure in the height direction, and Δz is the corresponding height change.

[0054] (5) Effective stress is an important control parameter of soil strength and deformation. It can reflect the coupling relationship between pore water pressure and soil pressure and reveal the dynamic changes of soil mechanical properties. Its calculation formula is:

[0055] σ , =σ-u (4)

[0056] Where, σ is the effective stress, σ is the total stress (measured by the soil pressure sensor), and u is the pore water pressure.

[0057] (5) The equal pore water pressure ratio is an important indicator to measure the pore water pressure reaching the limit state, which can identify the key process of mud formation and migration inside the soil. Its calculation formula is:

[0058]

[0059] Where ru is the equal pore water pressure ratio.

[0060] (6) In a specific embodiment of the present invention, the ballast vibration acceleration response information is obtained by an intelligent particle sensor, and the vibration characteristics are extracted based on the time series analysis method for input into the random forest model, such as Figure 3As shown in the figure. First, intelligent particle sensors are deployed in the railway ballasted track to collect the raw data of ballast vibration acceleration in real time, and preprocess it, including denoising and normalization, to improve the data quality. Then, the preprocessed vibration acceleration time series is divided into several time windows at fixed time intervals. The data in each window represents the vibration characteristics of the ballast in a specific time period. Further, by drawing autocorrelation diagrams and partial autocorrelation diagrams, the vibration data characteristics in the time window are analyzed to determine the time series model category, such as AR, MA, ARMA and ARIMA models. According to the model category, the time series model of the time window is further constructed, and the autocorrelation coefficient of the model is calculated. These autocorrelation coefficients are divided into several subgroups, and the X-bar control chart is used to describe their fluctuation characteristics, and key statistical features are extracted, including mean, fluctuation range, etc. Finally, these features are input into the random forest model as key indicators of ballast vibration for training and identification, so as to achieve efficient and accurate discrimination of mud bubbling state.

[0061] After the data preprocessing stage is completed, the present invention completes the training and optimization of the model through the following specific implementation methods.

[0062] The present invention adopts the random forest model. The random forest algorithm is an integrated learning algorithm that makes predictions by constructing multiple decision trees, which has the advantages of improving prediction accuracy and reducing overfitting. The following steps are used to achieve data prediction and analysis:

[0063] The data set was constructed using historical test data and annotated mud-bubbling states, in which the training set and test set accounted for 70% and 30% of the total data set, respectively.

[0064] Randomly select samples: Use the replacement sampling method to randomly select multiple subsets from the training data set to ensure sample diversity.

[0065] Randomly select features: At each node split, some features are randomly selected for splitting, thereby increasing the generalization ability of the model.

[0066] Ensemble learning: weight or vote on the prediction results of multiple trees and output the final prediction result.

[0067] Each decision tree in the random forest is constructed independently. The specific steps are as follows:

[0068] Random selection of samples: By using the replacement sampling method, a training set subset Ssub of size N is generated from the training data set to ensure that the sample data of each decision tree has a certain degree of difference.

[0069] Random selection of features: For each node split, m features are randomly selected, and the number of m is at least 5 to ensure the generalization ability of the model. The number of features of each node is usually the square root of the total number of selected features, and the optimal splitting feature is selected from the feature set used by the node.

[0070] Determination of the splitting standard: Gini impurity is used as the standard for splitting nodes, where the Gini impurity formula is:

[0071]

[0072] Sort by Gini impurity from small to large, and select the feature subset with the smallest Gini impurity. This value represents the quantification of classification accuracy.

[0073] The tree termination conditions are: the tree reaches the maximum depth; the number of node samples is less than the preset minimum number of samples; further splitting fails to improve the information gain.

[0074] After multiple decision trees are trained, predictions are made using ensemble learning methods, including:

[0075] The prediction results of all trees are voted by majority, the category with the most occurrences is selected as the final prediction result, and the number of votes is output as the possibility distribution to provide data support for subsequent maintenance strategies and fault prevention.

[0076]

[0077] To improve the performance of the model and optimize the model, the following key hyperparameters need to be optimized in random forest training:

[0078] Number of trees: determines the number of trees in the forest. More trees generally lead to higher accuracy, but the complexity increases and the model calculation takes longer.

[0079] Maximum tree depth: Controls the complexity of the tree to avoid overfitting or underfitting caused by too large or too small depth.

[0080] Maximum number of features per node: The number of optional features when a node is split, usually set to the square root of the total number of features.

[0081] Minimum number of sample splits: Limit the minimum number of samples for node splitting to control the depth of the tree.

[0082] Minimum number of leaf node samples: Ensure that each leaf node contains at least the number of samples to reduce the risk of model overfitting.

[0083] The present invention optimizes the above model by the following steps:

[0084] Cross-validation: By dividing the data set and performing multiple rounds of training and validation, the generalization ability of the model is evaluated.

[0085] Grid Search: Systematically search for combinations of hyperparameters to find the optimal ones.

[0086] Feature importance analysis and selection: Evaluate the contribution of features to model prediction and select key features to further improve model performance.

[0087] The model performance is specifically evaluated through the accuracy, precision, recall, and F1 score of the classification problem to verify its generalization ability and robustness.

[0088] After the model is established, the real-time discrimination system of the present invention obtains real-time data through the automatic data collection module, transmits the collected data to the cloud data processing platform using the data transmission module, and analyzes based on the pre-trained model to generate discrimination results and adaptively optimize by combining the discrimination results and the actual situation. When the cloud feedback of the actual situation does not match the discrimination result, the adaptive optimization of the model is achieved by filling the data set and re-optimizing the model according to the new data set, selecting hyperparameters. Through the adaptive function, this method can cope with the discrimination of mud bubbling under different environmental conditions, and can achieve real-time monitoring and prediction under the action of freeze-thaw cycles, rainfall and dynamic loads.

[0089] This method can also be used to identify other types of damage under complex geological conditions, such as intelligent identification of subgrade softening and roadbed settlement.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent method for distinguishing mud and slurry on railway ballasted track, characterized in that: The following steps are involved: 1) Data collection: Use the automated data collection equipment installed in the track bed and subgrade to monitor the data in real time, and transmit the collected data to the cloud processing platform in real time through the wireless communication module; 2) Data preprocessing: The cloud processing platform preprocesses the collected raw data and performs feature extraction; 3) Model training: Use historical test data and annotated mud-swelling states to build a data set to train a random forest classification model, and then optimize the model; 4) Real-time discrimination: The real-time collected monitoring data is input into the trained random forest model, the mud-bubbling state and its probability distribution are output, and the model is optimized through a real-time feedback mechanism.

2. The intelligent railway ballasted track mud bubbling identification method according to claim 1 is characterized in that: The data collection steps include the following: By arranging sensors at key positions of the ballast bed and the subgrade, including moisture content sensors, pore water pressure sensors, soil pressure gauges, temperature sensors and smart particles, the status information of the railway ballasted track is collected; the sampling frequency of the sensors is set to 0.01 Hz to 10 Hz; the collected information includes moisture content, pore water pressure, soil pressure, and the movement state and vibration response of ballast particles.

3. The intelligent railway ballasted track mud bursting and mud emitting identification method according to claim 1 is characterized in that: The data preprocessing step first performs outlier detection, missing value filling and normalization on the original data, and then uses an algorithm to extract features, including introducing the liquid and plastic limits of the soil, and further calculating the volume moisture content of the soil to obtain the features, which are specifically manifested as follows: Among them, W t is the input feature, W is the monitoring data, W L is the soil liquid limit, W P is the plastic limit of the soil; the gradient of the pore water pressure information is calculated and the gradient is normalized as a feature, which is specifically expressed as follows: Among them, Δu is the rate of change of pore water pressure, u(t) and u(t+Δt) represent the pore water pressure at time t and t+Δt respectively, and Δt is the sampling interval. The effective stress is further obtained from the pore water pressure and soil pressure as a feature. The specific process is as follows: Among them, Gu is the normalized pore water pressure gradient, Δu is the change of pore water pressure in the height direction, and Δz is the corresponding height change; the original value of the monitored temperature is used as the feature; the sliding window method is used to process the original data collected by the intelligent particles to obtain the vibration acceleration characteristics of the ballast.

4. The intelligent railway ballasted track mud bubbling identification method according to claim 1 is characterized in that: The data set constructed in the model training, wherein the training set and the test set account for 70% and 30% of the total data respectively; In the model training, features are screened, multiple decision trees are constructed through the random forest algorithm to cover multi-dimensional data for centralized learning, importance scores are given to features involved in decision-making in training, and key features are screened out for discriminant model training; The grid search method is used to optimize the hyperparameters of the random forest model, including the number of decision trees, the maximum tree depth, and the minimum number of split samples; and the classification accuracy, recall rate, and F1 value are calculated through cross-validation.

5. The intelligent railway ballasted track mud bubbling identification method according to claim 1 is characterized in that: In the real-time discrimination step, the model outputs two states of mud bubbling: normal and mud bubbling, and is accompanied by a state possibility distribution.

6. The intelligent railway ballasted track mud bubbling identification method according to claim 1 is characterized in that: The method can compare the real state based on the real-time output result and perform feedback adjustment in combination with the monitored data.

Images