Railway dispatching security data aggregation processing method based on AI model training

Abstract

The invention discloses a railway dispatching safety data aggregation processing method based on AI model training, and the method comprises the following steps: S1, collecting data related to railway dispatching from a plurality of data sources, including train operation data, dispatching commands, track conditions, signal information and meteorological data, carrying out the cleaning and standardization processing of the collected data, and obtaining a standard data set; the method comprises the following steps: S1, selecting a data set, S2, integrating different types of data by using a multi-modal fusion method in deep learning, S3, extracting key features from the fused data set, performing time sequence feature extraction by using an auto-encoder method, and performing feature selection by using XGBoost, S4, performing AI model training by using the processed data set, and selecting a random forest for training; and S5, applying the trained AI model to real-time data. According to the invention, a solid technical support is provided for improving the safety and the intelligent level of a railway dispatching system.

Description

Technical Field

[0001] The present invention relates to the field of AI model technology, and in particular to a railway dispatching safety data aggregation processing method based on AI model training. Background Art

[0002] Railway dispatching is a key link in the railway transportation system and is directly related to the safety and efficiency of train operation. In the railway dispatching process, it is necessary to comprehensively consider multiple factors such as train operation data, dispatching orders, track conditions, signal information and meteorological data to ensure the safe and punctual operation of trains. However, the existing railway dispatching system mainly relies on manual experience to make decisions, and there are several core difficulties: 1. Numerous and scattered data sources: The data sources involved in railway dispatching are diverse, including train operation status, dispatching orders, track conditions, signal information, and meteorological data. These data are usually scattered in different systems, lacking unified standards and formats, making it difficult to conduct effective comprehensive analysis.

[0003] 2. Complex data processing: Due to the diversity and complexity of railway dispatching data, data needs to be cleaned, standardized and aggregated to ensure data quality and consistency. This process usually requires a lot of manual intervention and expertise, which is time-consuming, labor-intensive and prone to errors.

[0004] 3. Difficulty in feature extraction and selection: In the process of analyzing railway dispatching data, it is necessary to extract and select key features that can reflect dispatching safety from a large amount of high-dimensional data. Existing methods mostly rely on traditional statistical methods and expert experience, which makes it difficult to fully explore the potential patterns and trends in the data, resulting in limited prediction accuracy and model performance.

[0005] 4. High real-time requirements: Railway dispatching systems need to be able to quickly analyze and process real-time data, timely predict possible safety hazards and dispatching risks, and provide corresponding decision support and safety warnings. Existing systems are insufficient in terms of real-time and accuracy, and cannot meet the requirements of safe and efficient railway transportation.

[0006] Therefore, how to provide a railway dispatching safety data aggregation and processing method based on AI model training to achieve comprehensive processing of multi-source data, efficient feature extraction and selection, real-time prediction and decision support has become a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0007] One object of the present invention is to propose a railway dispatching safety data aggregation processing method based on AI model training.

[0008] A railway dispatching safety data aggregation processing method based on AI model training according to an embodiment of the present invention includes the following steps: S1. Collect railway dispatch-related data from multiple data sources, including train operation data, dispatching orders, track conditions, signal information and meteorological data, and clean and standardize the collected data; S2. Use the multimodal fusion method in deep learning to integrate different types of data to form a unified data set; S3, extract key features from the fused data set, use the autoencoder method to extract time series features, and use XGBoost for feature selection; S4. Use the processed data set to train the AI ​​model, select random forest for training, and optimize the model parameters through hyperparameter tuning technology; S5. Real-time prediction and decision support: Apply the trained AI model to real-time data to make real-time predictions of safety hazards and scheduling risks, and provide decision support and safety warnings.

[0009] Optionally, S1 includes the following steps: S11. Collecting data related to railway dispatching in real time from multiple data sources, including train operation data, dispatching orders, track conditions, signal information and meteorological data; S12. Represent the collected train operation data as a time series x t , where x t represents the train position, speed and acceleration at time t; S13, structurally process the dispatch command data and extract key information c in the dispatch command i , where c i Represents the content and timestamp of the i-th scheduling command; S14, spatially map the track condition data to generate a track condition matrix R = {r ij}, where r ij represents the status of the i-th track segment at the j-th time; S15. Collect signal information data and convert it into a signal state sequence s t , where s t Represents the signal state at time t; S16, obtain meteorological data, and perform spatial and temporal matching to form a meteorological condition matrix W = {w ij}, where w ij represents the meteorological conditions at the i-th location at the j-th time; S17. Perform data cleaning on all collected data, including processing missing values, outliers and noise data, using interpolation to process missing values, and using triple standard deviation method to detect and process outliers; S18. Standardize the cleaned data and use the z-score standardization method to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0010] Optionally, S2 includes the following steps: S21, the pre-processed train operation data x t , Scheduling command data c i , track condition data R = {r ij}、Signal information data t and meteorological data W = {w ij} Perform preliminary feature extraction to form a feature vector set f x ,f c ,f r ,f s ,f w , where x t represents the train position, speed and acceleration at time t, c i Represents the content and timestamp of the i-th scheduling command, r ij represents the status of the track segment i at time j, s t represents the signal state at time t, w ij represents the meteorological conditions at the i-th location at the j-th time, f x 、f c 、f r 、f s 、f w : The feature vectors representing train operation data, dispatching command data, track condition data, signal information data and meteorological data respectively; S22. Use a deep neural network to encode each type of feature vector separately to form an intermediate representation h x ,h c ,h r ,h s ,h w , where DNN represents the deep neural network function, h x =DNN(f x ) represents the intermediate representation of train operation data, h c =DNN(f c ) represents the intermediate representation of the scheduling command data, h r =DNN(f r ): represents the intermediate representation of track status data, h s =DNN(f s) represents the intermediate representation of signal information data, h w =DNN(f w ) represents an intermediate representation of meteorological data; S23. Use the self-attention mechanism to perform weighted fusion of different types of intermediate representations; S24. According to the calculated attention weights, the intermediate representations are weighted and summed to form the fused feature representation H: H = α x h x +α c h c +α r h r +α s h s +α w h w ; Among them, H represents the fused feature representation; S25, the fused feature representation H is processed through the fully connected layer to reduce the dimension and obtain a unified data set representation {z k}, where: z k =FC(H); Among them, z k Represents the feature representation after dimensionality reduction; S26. Representation of a unified dataset k} to form the final fusion data set {z ′ k}.

[0011] Optionally, S3 specifically includes the following steps: S31. The integrated railway dispatching safety data set z ′ k Perform preliminary processing and construct the feature matrix F = {f kj}, where f kj represents the jth feature of the kth sample, the feature is the track condition, z ′ k represents the normalized kth sample, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and F represents the feature matrix of all samples; S32, use the autoencoder method to extract time series features from the feature matrix F, and build an autoencoder model, including the encoder and decoder parts. The encoder takes the input feature f kj Mapped to the hidden layer representation h k , the decoder represents the hidden layer h k Reconstruct back to input features, the optimization objective is: Among them, θ represents the parameters of the autoencoder model, including the encoder parameters θ E and decoder parameters θ D , represents the reconstructed feature matrix, Where E(θ E ,F) represents the output of the encoder, D(θ D ,E(θ E , F) represents the output of the decoder, ∥ represents the norm of the reconstruction error, usually the L2 norm, λ represents the regularization parameter used to control the model complexity, L represents the number of layers of the autoencoder, W l represents the weight matrix of the lth layer, b l Represents the bias vector of the lth layer; S33. Extracting time series features {h k}: h k =E(θ E ,f k ); Among them, h k Indicates the historical operation mode of the train, the change of track conditions, and the change of signal status. E ,f k ) indicates that the encoder transforms the feature f of the kth sample k The process of mapping to the hidden layer; S34. Use the XGBoost algorithm for feature selection and build a set of decision trees. The objective function is: Among them, Θ represents the parameters of the XGBoost model, N represents the number of samples, that is, the number of samples of railway scheduling safety data, L represents the loss function, which is used to measure the error between the predicted value and the true value, and y i represents the true label of the i-th sample, such as the safety status of train scheduling, represents the predicted value of the i-th sample, T represents the number of decision trees Ω(f t ) represents the regularization term of the t-th tree: Among them, γ represents the regularization parameter, which controls the complexity of the tree, d represents the number of features, and w j represents the weight of feature j; S35. Calculate the importance score of each feature I j ; S36, according to the importance score I j Sort the features, select the top n features with the highest scores, and construct the final feature set F ′Factors reflecting the safety of railway dispatching, where n is the number of preset features; S37, the final feature set F ′ It serves as input data for subsequent model training to further improve the accuracy and reliability of railway scheduling safety predictions.

[0012] Optionally, the S35 specifically includes: S351, for each round of iteration t, using the current residual r i t Build a new decision tree and update the model's predictions: in, represents the predicted value of the i-th sample after the t+1th round of iteration, represents the prediction of the railway scheduling safety data by the model after the t+1th round of iteration, η: learning rate, used to control the update amplitude of the predicted value at each iteration, f t (x i ) represents the predicted value of the t-th tree for sample i, and represents the contribution of the t-th decision tree to the railway scheduling safety data; S352. In each round of iteration, the node splitting gain of the current tree is calculated. The gain formula is: Among them, G s represents the splitting gain of node s, and represents the information gain obtained by the current node when splitting. L represents the sample set of the left child node, represents the railway scheduling safety data sample belonging to the left child node after splitting, and I R represents the sample set of the right child node, represents the railway scheduling safety data samples belonging to the right child node after the split, I represents the sample set of the current node, represents all railway scheduling safety data samples before the split, g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. i represents the second-order gradient of the i-th sample, represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error. λ represents the regularization parameter, which is used to control the complexity of the model and avoid overfitting. S353. For each feature j, accumulate its split gain on all nodes and calculate the importance score I j : Among them, I j represents the importance score of the jth feature, represents the importance of feature j in all decision trees, T represents the number of decision trees, represents the total number of decision trees constructed, S trepresents the set of all nodes in the t-th tree, represents all nodes in the t-th decision tree, G js represents the splitting gain of the jth feature at node s, and represents the information gain obtained when feature j is split at node s; S354. For each node s, the gain is calculated according to the accumulated value of the first-order gradient and the second-order gradient. The calculation formula is: Among them, G js represents the splitting gain of the jth feature at node s, and represents the contribution of feature j when splitting at node s. s represents the sample set on node s, represents all railway scheduling safety data samples belonging to node s, and g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. i It represents the second-order gradient of the i-th sample, which represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error.

[0013] Optionally, the S4 specifically includes: S41, using the processed feature set F ′ Construct training data set {(x i ,y i )}, where F ′ represents the final feature set, which contains the features with the highest scores. The features represent the key information of railway scheduling safety data. i represents the label of the ith sample, indicating the safety status of railway scheduling, such as safety, potential risk or failure, x i The feature vector representing the i-th sample contains train operation data, dispatching commands, track conditions, signal information, and meteorological data; S42, select the random forest algorithm for model training, and construct multiple decision trees {T m}, where each tree is trained using a different subset of features and samples: {T m}={T1,T2,…,T M}; Among them, {T m} represents the set of decision trees in the random forest, T m : the mth decision tree, M: the number of decision trees, indicating the total number of trees contained in the random forest; S43. Optimize model parameters through hyperparameter tuning technology to determine the best hyperparameter combination {θ k}, including the number of trees M, the maximum depth D, and the minimum number of sample splits N min : {θk}={θ1,θ2,…,θ K}; Among them, {θ k} represents the hyperparameter set, θ k : The kth hyperparameter, K represents the number of hyperparameters, M represents the number of decision trees, which represents the total number of trees contained in the random forest, D represents the maximum depth of the decision tree, which represents the maximum number of layers of the tree, and N min Indicates the minimum number of sample splits, which indicates the minimum number of samples required for a node to split; S44. Use the grid search method to search for the best parameter combination in the preset parameter space, and the optimization goal is: Among them, θ represents the hyperparameter to be optimized, N: the number of samples, which represents the total number of samples in the training data set, Loss function, measuring the predicted value and the true value y i The difference between S45. Evaluate model performance through cross-validation; S46. Using the best hyperparameter combination Retrain the random forest model to get the final model in, The final random forest model represents the predicted value of the input feature vector x, T m (x): The predicted value of sample x by the mth decision tree. M represents the number of decision trees and the total number of trees in the random forest.

[0014] Optionally, the S45 specifically includes: S451, the training data set {(x i ,y i )} is divided into K non-overlapping subsets, denoted as {D1,D2,…,D K}, where the number of samples contained in each subset is Among them, {(x i ,y i )} represents the training data set, which contains the features and labels of railway dispatching safety data. K represents the number of cross-validation folds, and the training data set is divided into K subsets. k represents the kth subset, each subset contains a part of the railway dispatching safety data samples, N represents the total number of samples in the training data set, and represents the number of samples of the railway dispatching safety data; S452. Perform the following operations on each hyperparameter combination θ: For the kth subset D k , which is used as the validation set, and the remaining K-1 subsets ({D1,D2,…,D k-1 ,D k+1 ,…,D K}Merge into training set: represents the training set in the kth fold, which is composed of the K-1 subsets and contains the features and labels of railway scheduling safety data; represents the validation set in the k-th fold, which consists of the k-th subset and contains the features and labels of the railway scheduling safety data; In the training set Train the random forest model and get the model in, represents the predicted value of the input feature vector x by the random forest model trained on the training set in the kth fold, and represents the prediction of railway scheduling safety data, represents the mth decision tree in the kth fold, the predicted value of the input feature vector x, represents the prediction of railway scheduling safety data, M represents the number of decision trees, and represents the total number of trees contained in the random forest; Using a validation set Evaluate model performance and calculate loss function The average value of: Among them, L k represents the average loss in the kth fold, represents the model performance on the validation set, represents the number of samples in the validation set, Represents the model in the kth fold for the validation set sample x i The predicted value of Represents the loss function, measuring the predicted value and the true value y i The difference between is used to evaluate the prediction accuracy of railway dispatching safety data; S453. Calculate the average loss of all compromises as the performance indicator of the hyperparameter combination θ: Among them, L cv (θ) represents the cross-validation average loss of the hyperparameter combination θ, represents the model performance in all folds, K represents the number of cross-validation folds, L k represents the average loss in the kth fold; S454. Select the hyperparameter combination with the minimum cross-validation average loss as the best hyperparameter combination: in, represents the best hyperparameter combination after optimization, represents the kth optimized hyperparameter, L cv (θ) represents the cross-validation average loss for the hyperparameter combination θ.

[0015] Optionally, the S5 specifically includes: S51. Collect and preprocess real-time railway dispatching data, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and convert these data into feature vectors x real-time , Among them, x real-time A feature vector representing real-time data, including the latest train operation status, dispatching orders, track conditions, signal information, and meteorological data; S52, the real-time data feature vector x real-time Input to the trained AI model Generate real-time prediction results in, It indicates the prediction result of AI model on real-time data and the safety status of current railway dispatching. Represents a trained AI model, which is used to predict the input feature vector x; S53. Based on the prediction results Analyze potential safety hazards and scheduling risks, and determine whether there are potential risks based on preset thresholds: Among them, Risk represents the real-time risk level, which is classified according to the prediction results, τ high Represents a high risk threshold, exceeding this value indicates high risk, τ low represents a low risk threshold, below which the risk is low; S54. Provide corresponding decision support according to the risk level Risk, including scheduling adjustment suggestions and safety measures suggestions. The specific steps are: when the risk level is high, take emergency scheduling adjustment measures immediately; when the risk level is medium, issue an early warning and recommend the dispatcher to take preventive measures; when the risk level is low, continue monitoring without taking additional measures.

[0016] The beneficial effects of the present invention are: The railway dispatching safety data aggregation processing method based on AI model training proposed in this invention describes in detail the specific steps of realizing railway dispatching safety data aggregation processing by comprehensively utilizing advanced algorithms such as computer deep learning and random forest, and has the following significant advantages: 1. Strong comprehensive data processing capabilities: This method can effectively integrate multiple data sources such as train operation data, dispatching commands, track conditions, signal information and meteorological data through the aggregation and processing of multi-source data, and perform data cleaning, standardization and aggregation processing to ensure the comprehensiveness and consistency of the data, thereby improving the accuracy and reliability of data analysis.

[0017] 2. Efficient feature extraction and selection: In the feature engineering process, this method uses autoencoders to extract time series features and uses the XGBoost algorithm for feature selection to extract key features from large amounts of high-dimensional data, improving the accuracy and generalization of the model. The accuracy of feature selection enables the model to better capture important patterns and trends in the data.

[0018] 3. Advanced model training and optimization: This method uses the random forest algorithm to train the model and combines hyperparameter tuning technology to optimize the model parameters to ensure high performance and stability of the model. The application of cross-validation technology further improves the reliability of the model and provides a solid foundation for real-time prediction of railway scheduling safety.

[0019] 4. Real-time prediction and decision support: This method applies the trained AI model to real-time data, which can predict safety hazards and scheduling risks in real time, and provide corresponding decision support and safety warnings. By dynamically updating the model input data, the real-time and accuracy of the railway dispatching system is ensured, helping dispatchers to take timely measures to prevent potential accidents and improve the safety and efficiency of railway dispatching.

[0020] 5. High level of intelligence and automation: The present invention uses advanced AI technology and data processing methods to realize the intelligent and automated processing of railway dispatching safety data, significantly reducing the dependence on manual experience, reducing the safety hazards and errors caused by manual operation, and improving the overall intelligence level and operation efficiency of the railway dispatching system.

[0021] In summary, the present invention provides a solid technical support for improving the safety and intelligence level of the railway dispatching system. It has the advantages of strong comprehensive data processing, accurate feature extraction, efficient model performance, reliable real-time prediction and high intelligence level, and can significantly improve the safety and efficiency of railway dispatching. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying 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 of the present invention. In the accompanying drawings: Figure 1 This is an overall flow chart of a railway dispatching safety data aggregation processing method based on AI model training proposed in the present invention. The flow chart describes in detail the entire process from data collection and preprocessing, multimodal data fusion, high-dimensional feature extraction and selection, model training and optimization to real-time prediction and decision support; Figure 2 This is a detailed schematic diagram of the data collection and preprocessing module, showing the steps of collecting, cleaning and standardizing multi-source data such as train operation data, dispatching commands, track conditions, signal information and meteorological data. DETAILED DESCRIPTION

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0024] refer to Figure 1 , a railway dispatching safety data aggregation processing method based on AI model training, comprising the following steps: S1, collecting railway dispatching related data from multiple data sources, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and cleaning and standardizing the collected data; In this implementation, S1 includes the following steps: S11. Collecting data related to railway dispatching in real time from multiple data sources, including train operation data, dispatching orders, track conditions, signal information and meteorological data; S12. Represent the collected train operation data as a time series x t , where x t represents the train position, speed and acceleration at time t; S13, structurally process the dispatch command data and extract key information c in the dispatch command i , where c i Represents the content and timestamp of the i-th scheduling command; S14, spatially map the track condition data to generate a track condition matrix R = {r ij}, where r ij represents the status of the i-th track segment at the j-th time; S15. Collect signal information data and convert it into a signal state sequence s t , where s t Represents the signal state at time t; S16, obtain meteorological data, and perform spatial and temporal matching to form a meteorological condition matrix W = {w ij}, where w ij represents the meteorological conditions at the i-th location at the j-th time; S17. Perform data cleaning on all collected data, including processing missing values, outliers and noise data, using interpolation to process missing values, and using triple standard deviation method to detect and process outliers; S18. Standardize the cleaned data and use the z-score standardization method to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0025] S2. Integrate different types of data using a multimodal fusion method in deep learning to form a unified data set; Optionally, S2 includes the following steps: S21, the pre-processed train operation data x t , Scheduling command data c i , track condition data R = {r ij}、Signal information data t and meteorological data W = {w ij} Perform preliminary feature extraction to form a feature vector set f x ,f c ,f r ,f s ,f w , where x t represents the train position, speed and acceleration at time t, c i Represents the content and timestamp of the i-th scheduling command, r ij represents the status of the track segment i at time j, s t represents the signal state at time t, w ij represents the meteorological conditions at the i-th location at the j-th time, f x 、f c 、f r 、f s 、f w : The feature vectors representing train operation data, dispatching command data, track condition data, signal information data and meteorological data respectively; S22. Use a deep neural network to encode each type of feature vector separately to form an intermediate representation h x ,h c ,h r ,h s ,h w , where DNN represents the deep neural network function, h x =DNN(f x ) represents the intermediate representation of train operation data, hc =DNN(f c ) represents the intermediate representation of the scheduling command data, h r =DNN(f r ): represents the intermediate representation of track status data, h s =DNN(f s ) represents the intermediate representation of signal information data, h w =DNN(f w ) represents an intermediate representation of meteorological data; S23. Use the self-attention mechanism to perform weighted fusion of different types of intermediate representations; S24. According to the calculated attention weights, the intermediate representations are weighted and summed to form the fused feature representation H: H = α x h x +α c h c +α r h r +α s h s +α w h w ; Among them, H represents the fused feature representation; S25, the fused feature representation H is processed through the fully connected layer to reduce the dimension and obtain a unified data set representation {z k}, where: z k =FC(H); Among them, z k Represents the feature representation after dimensionality reduction; S26. Representation of a unified dataset k} to form the final fusion data set {z ′ k}.

[0026] S3, extract key features from the fused data set, use the autoencoder method to extract time series features, and use XGBoost for feature selection; In this implementation, S3 specifically includes the following steps: S31. The integrated railway dispatching safety data set z ′ k Perform preliminary processing and construct the feature matrix F = {f kj}, where f kj represents the jth feature of the kth sample, the feature is the track condition, z ′ k represents the normalized kth sample, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and F represents the feature matrix of all samples; S32, use the autoencoder method to extract time series features from the feature matrix F, and build an autoencoder model, including the encoder and decoder parts. The encoder takes the input feature f kj Mapped to the hidden layer representation h k , the decoder represents the hidden layer h k Reconstruct back to input features, the optimization objective is: Among them, θ represents the parameters of the autoencoder model, including the encoder parameters θ E and decoder parameters θ D , represents the reconstructed feature matrix, Where E(θ E ,F) represents the output of the encoder, D(θ D ,E(θ E , F) represents the output of the decoder, ∥ represents the norm of the reconstruction error, usually the L2 norm, γ represents the regularization parameter used to control the complexity of the model, L represents the number of layers of the autoencoder, and W l represents the weight matrix of the lth layer, b l Represents the bias vector of the lth layer; S33. Extracting time series features {h k}: h k =E(θ E ,f k ); Among them, h k Indicates the historical operation mode of the train, the change of track conditions, and the change of signal status. E ,f k ) indicates that the encoder transforms the feature f of the kth sample k The process of mapping to the hidden layer; S34. Use the XGBoost algorithm for feature selection and build a set of decision trees. The objective function is: Among them, Θ represents the parameters of the XGBoost model, N represents the number of samples, that is, the number of samples of railway scheduling safety data, L represents the loss function, which is used to measure the error between the predicted value and the true value, and y i represents the true label of the i-th sample, such as the safety status of train scheduling, represents the predicted value of the i-th sample, T represents the number of decision trees Ω(f t ) represents the regularization term of the t-th tree: Among them, γ represents the regularization parameter, which controls the complexity of the tree, d represents the number of features, and w j represents the weight of feature j; S35. Calculate the importance score of each feature I j ; S36, according to the importance score I j Sort the features, select the top n features with the highest scores, and construct the final feature set F ′ Factors reflecting the safety of railway dispatching, where n is the number of preset features; S37, the final feature set F ′ It serves as input data for subsequent model training to further improve the accuracy and reliability of railway scheduling safety predictions.

[0027] S35. Calculate the importance score of each feature I j ; In this implementation, S35 specifically includes: S351, for each iteration t, using the current residual Build a new decision tree and update the model's predictions: in, represents the predicted value of the i-th sample after the t+1th round of iteration, represents the prediction of the railway scheduling safety data by the model after the t+1th round of iteration, η: learning rate, used to control the update amplitude of the predicted value at each iteration, f t (x i ) represents the predicted value of the t-th tree for sample i, and represents the contribution of the t-th decision tree to the railway scheduling safety data; S352. In each round of iteration, the node splitting gain of the current tree is calculated. The gain formula is: Among them, G s represents the splitting gain of node s, and represents the information gain obtained by the current node when splitting. L represents the sample set of the left child node, represents the railway scheduling safety data sample belonging to the left child node after splitting, and I R represents the sample set of the right child node, represents the railway scheduling safety data samples belonging to the right child node after the split, I represents the sample set of the current node, represents all railway scheduling safety data samples before the split, g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. irepresents the second-order gradient of the i-th sample, represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error. λ represents the regularization parameter, which is used to control the complexity of the model and avoid overfitting. S353. For each feature j, accumulate its split gain on all nodes and calculate the importance score I j : Among them, I j represents the importance score of the jth feature, represents the importance of feature j in all decision trees, T represents the number of decision trees, represents the total number of decision trees constructed, S t represents the set of all nodes in the t-th tree, represents all nodes in the t-th decision tree, G js represents the splitting gain of the jth feature at node s, and represents the information gain obtained when feature j is split at node s; S354. For each node s, the gain is calculated according to the accumulated value of the first-order gradient and the second-order gradient. The calculation formula is: Among them, G js represents the splitting gain of the jth feature at node s, and represents the contribution of feature j when splitting at node s. s represents the sample set on node s, represents all railway scheduling safety data samples belonging to node s, and g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. i It represents the second-order gradient of the i-th sample, which represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error.

[0028] S4. Use the processed data set to train the AI ​​model, select random forest for training, and optimize the model parameters through hyperparameter tuning technology; In this implementation, S4 specifically includes: S41, using the processed feature set F ′ Construct training data set {(x i ,y i )}, where F ′ represents the final feature set, which contains the features with the highest scores. The features represent the key information of railway scheduling safety data. i represents the label of the ith sample, indicating the safety status of railway scheduling, such as safety, potential risk or failure, x i The feature vector representing the i-th sample contains train operation data, dispatching commands, track conditions, signal information, and meteorological data; S42, select the random forest algorithm for model training, and construct multiple decision trees {T m}, where each tree is trained using a different subset of features and samples: {T m}={T1,T2,…,T M}; Among them, {T m} represents the set of decision trees in the random forest, T m : the mth decision tree, M: the number of decision trees, indicating the total number of trees contained in the random forest; S43. Optimize model parameters through hyperparameter tuning technology to determine the best hyperparameter combination {θ k}, including the number of trees M, the maximum depth D, and the minimum number of sample splits N min : {θ k}={θ1,θ2,…,θ K}; Among them, {θ k} represents the hyperparameter set, θ k : The kth hyperparameter, K represents the number of hyperparameters, M represents the number of decision trees, which represents the total number of trees contained in the random forest, D represents the maximum depth of the decision tree, which represents the maximum number of layers of the tree, and N min Indicates the minimum number of sample splits, which indicates the minimum number of samples required for a node to split; S44. Use the grid search method to search for the best parameter combination in the preset parameter space, and the optimization goal is: Among them, θ represents the hyperparameter to be optimized, N: the number of samples, which represents the total number of samples in the training data set, Loss function, measuring the predicted value and the true value y i The difference between S45. Evaluate model performance through cross-validation; S46. Using the best hyperparameter combination Retrain the random forest model to get the final model in, The final random forest model represents the predicted value of the input feature vector x, T m (x): The predicted value of sample x by the mth decision tree. M represents the number of decision trees and the total number of trees in the random forest.

[0029] S45. Evaluate model performance through cross-validation; In this implementation manner, S45 specifically includes: S451, the training data set {(x i ,y i )} is divided into K non-overlapping subsets, denoted as {D1,D2,…,D K}, where the number of samples contained in each subset is Among them, {(x i ,y i )} represents the training data set, which contains the features and labels of railway dispatching safety data. K represents the number of cross-validation folds, and the training data set is divided into K subsets. k represents the kth subset, each subset contains a part of the railway dispatching safety data samples, N represents the total number of samples in the training data set, and represents the number of samples of the railway dispatching safety data; S452. Perform the following operations on each hyperparameter combination θ: For the kth subset D k , which is used as the validation set, and the remaining K-1 subsets ({D1,D2,…,D k-1 ,D k+1 ,…,D K}Merge into training set: represents the training set in the kth fold, which is composed of the K-1 subsets and contains the features and labels of railway scheduling safety data; represents the validation set in the k-th fold, which consists of the k-th subset and contains the features and labels of the railway scheduling safety data; In the training set Train the random forest model and get the model in, represents the predicted value of the input feature vector x by the random forest model trained on the training set in the kth fold, and represents the prediction of railway scheduling safety data, represents the mth decision tree in the kth fold, the predicted value of the input feature vector x, represents the prediction of railway scheduling safety data, M represents the number of decision trees, and represents the total number of trees contained in the random forest; Using a validation set Evaluate model performance and calculate loss function The average value of: Among them, L k represents the average loss in the kth fold, represents the model performance on the validation set, represents the number of samples in the validation set, Represents the model in the kth fold for the validation set sample x i The predicted value of Represents the loss function, measuring the predicted value and the true value y i The difference between is used to evaluate the prediction accuracy of railway dispatching safety data; S453. Calculate the average loss of all compromises as the performance indicator of the hyperparameter combination θ: Among them, L cv (θ) represents the cross-validation average loss of the hyperparameter combination θ, represents the model performance in all folds, K represents the number of cross-validation folds, L k represents the average loss in the kth fold; S454. Select the hyperparameter combination with the minimum cross-validation average loss as the best hyperparameter combination: in, represents the best hyperparameter combination after optimization, represents the kth optimized hyperparameter, L cv (θ) represents the cross-validation average loss for the hyperparameter combination θ.

[0030] S5. Real-time prediction and decision support: Apply the trained AI model to real-time data to make real-time predictions of safety hazards and scheduling risks, and provide decision support and safety warnings.

[0031] In this implementation, S5 specifically includes: S51, collect and pre-process real-time railway dispatching data, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and convert these data into feature vectors x real-time , Among them, x real-time A feature vector representing real-time data, including the latest train status, dispatching orders, track conditions, signal information, and meteorological data; S52, the real-time data feature vector x real-time Input to the trained AI model Generate real-time prediction results in, It indicates the prediction result of AI model on real-time data and the safety status of current railway dispatching. Represents a trained AI model, which is used to predict the input feature vector x; S53. Based on the prediction results Analyze potential safety hazards and scheduling risks, and determine whether there are potential risks based on preset thresholds: Among them, Risk represents the real-time risk level, which is classified according to the prediction results, τ high Represents a high risk threshold, exceeding this value indicates high risk, τ low represents a low risk threshold, below which the risk is low; S54. Provide corresponding decision support according to the risk level Risk, including scheduling adjustment suggestions and safety measures suggestions. The specific steps are: when the risk level is high, take emergency scheduling adjustment measures immediately; when the risk level is medium, issue an early warning and recommend the dispatcher to take preventive measures; when the risk level is low, continue monitoring without taking additional measures.

[0032] Embodiment 1: In the actual railway dispatching system, it is crucial to ensure the safety of train operation and the effectiveness of dispatching. Take the Beijing-Shanghai high-speed railway trunk line in June 2023 as an example. The trunk line needs to dispatch 300 trains every day. Dispatchers need to monitor the train operation status, track conditions, signal information, meteorological data and other data in real time to ensure the safe operation of trains. However, the existing system has significant deficiencies in data processing, feature extraction, model prediction and decision support. By applying the railway dispatching safety data aggregation processing method based on AI model training of the present invention, we can significantly improve the safety and efficiency of railway dispatching.

[0033] In this embodiment, it is first necessary to collect and preprocess data from multiple data sources, including train operation data, dispatching commands, track conditions, signal information, and meteorological data. These data are acquired in real time through various sensors and monitoring systems, and collected through the data acquisition module. For the data on June 10, 2023, we use data cleaning and standardization processing techniques to remove noise and outliers, and convert the data into a unified standard format to ensure the quality and consistency of the data. For example, in the data collection of that day, the train operation data involves train speed and position data recorded once a minute, totaling 1440 records; the dispatching command data involves all dispatching instructions for the day, totaling 250; the track condition data includes track detection data once a day and every hour, totaling 24 records; signal information and meteorological data are recorded once a minute, totaling 1440 records.

[0034] After data cleaning and standardization, we use the multimodal fusion method in deep learning to effectively integrate different types of data to form a unified data set. Specifically, we use an autoencoder to extract features from the data. The encoder part maps the high-dimensional input data to a low-dimensional hidden layer representation, and then reconstructs the input data through the decoder part, and optimizes the model parameters by minimizing the reconstruction error. In this process, the autoencoder can automatically learn and extract key features that can reflect the safety of scheduling, thereby improving the accuracy of data analysis.

[0035] Next, we use the XGBoost algorithm for feature selection. The XGBoost algorithm constructs a gradient boosting decision tree, evaluates the importance of each feature, and selects the most representative features. For example, by analyzing the data on June 10, 2023, the XGBoost algorithm selects key features such as the rate of change of train speed, the rate of change of track conditions, the change of signal status, and the change of meteorological conditions. These features have high split gains in multiple decision tree splits and can significantly affect the prediction performance of the model.

[0036] After feature extraction and selection, we used the random forest algorithm to train the model. Through hyperparameter tuning technology, we determined the best hyperparameter combination, including 100 decision trees, 10 maximum depth, and 5 minimum sample splits. In practical applications, we verified the performance of the model through cross-validation technology to ensure that the model has high prediction accuracy and stability. Specific data shows that the prediction accuracy of the training set and the validation set reached 95% and 93% respectively, and the model has strong generalization ability.

[0037] By applying the trained AI model to real-time data, we can achieve real-time prediction of dispatch safety hazards and risks. For example, during a dispatch process on June 11, 2023, the model analyzed the current train operation status, track conditions and meteorological conditions in real time, predicted that a certain train had potential track risks, and issued a warning signal when the risk level reached the high risk threshold. The dispatcher took timely dispatch adjustment measures based on the warning information provided by the system to avoid possible accidents.

[0038] Table 1 Data collection Table 2 Feature extraction and selection results Table 3 Model training and optimization parameters Parameter name Parameter Symbols Parameter Value Number of decision trees M 100 Maximum Depth D 10 Minimum number of sample splits N_min 5 Training set accuracy 95％ Validation set accuracy 93％ Table 4: Real-time prediction and decision support examples Table 5: System performance comparison before and after application index Before application After application Improve the situation Accident rate 0.8％ 0.4％ 50% reduction Scheduling efficiency 70％ 91％ 30% improvement Workload high middle Significantly reduce Data processing accuracy 80％ 95％ 15% improvement Combined with the above table, in order to verify the beneficial effects of the present invention, we compared the dispatching safety and efficiency before and after the adoption of the present invention. Before the adoption of the present invention, dispatchers needed to rely on experience for data analysis and decision-making, the accuracy and real-time performance of data processing were low, and safety hazards were not identified in a timely manner. After the adoption of the present invention, the system can automatically process and analyze a large amount of complex data, provide accurate predictions and decision support in real time, and greatly improve the safety and efficiency of dispatching. Specific data show that after the adoption of the present invention, the accident rate of the dispatching system dropped from 0.8% to 0.4%, the dispatching efficiency increased by 30%, the workload of the dispatcher was significantly reduced, and the safety and reliability were greatly improved.

[0039] In summary, the present invention realizes intelligent and automatic processing of railway dispatching safety data by effectively integrating multi-source data, adopting advanced feature extraction and selection algorithms, and combining random forest models for training and optimization. After applying the present invention, the safety and efficiency of the railway dispatching system are significantly improved, providing a strong guarantee for the safe operation of railway transportation.

[0040] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A railway dispatching safety data aggregation processing method based on AI model training, characterized in that: The steps include: S1. Collect railway dispatch-related data from multiple data sources, including train operation data, dispatching orders, track conditions, signal information and meteorological data, and clean and standardize the collected data; S2. Use the multimodal fusion method in deep learning to integrate different types of data to form a unified data set; S3, extract key features from the fused data set, use the autoencoder method to extract time series features, and use XGBoost for feature selection; S4. Use the processed data set to train the AI ​​model, select random forest for training, and optimize the model parameters through hyperparameter tuning technology; S5. Real-time prediction and decision support: Apply the trained AI model to real-time data to make real-time predictions of safety hazards and scheduling risks, and provide decision support and safety warnings.

2. According to the method for railway dispatching safety data aggregation processing based on AI model training according to claim 1, it is characterized in that: The S1 comprises the following steps: S11. Collecting data related to railway dispatching in real time from multiple data sources, including train operation data, dispatching orders, track conditions, signal information and meteorological data; S12. Represent the collected train operation data as a time series x t , where x t represents the train position, speed and acceleration at time t; S13, structurally process the dispatch command data and extract key information c in the dispatch command i , where c i Represents the content and timestamp of the i-th scheduling command; S14, spatially map the track condition data to generate a track condition matrix R = {r ij }, where r ij represents the status of the i-th track segment at the j-th time; S15. Collect signal information data and convert it into a signal state sequence s t , where s t Represents the signal state at time t; S16, obtain meteorological data, and perform spatial and temporal matching to form a meteorological condition matrix W = {w ij }, where w ij represents the meteorological conditions at the i-th location at the j-th time; S17. Perform data cleaning on all collected data, including processing missing values, outliers and noise data, using interpolation to process missing values, and using triple standard deviation method to detect and process outliers; S18. Standardize the cleaned data and use the z-score standardization method to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1.

3. According to the railway dispatching safety data aggregation processing method based on AI model training according to claim 1, it is characterized in that: The S2 comprises the following steps: S21, the pre-processed train operation data x t , dispatch command data c i , track condition data R = {r ij }、Signal information data t and meteorological data W = {w ij } Perform preliminary feature extraction to form a feature vector set f x ,f c ,f r ,f s ,f w , where x t represents the train position, speed and acceleration at time t, c i Represents the content and timestamp of the i-th scheduling command, r ij represents the status of the i-th track segment at the j-th time, s t represents the signal state at time t, w ij represents the meteorological conditions at the i-th location at the j-th time, f x 、f c 、f r 、f s 、f w : The feature vectors representing train operation data, dispatching command data, track condition data, signal information data and meteorological data respectively; S22. Use a deep neural network to encode each type of feature vector separately to form an intermediate representation h x ,h c ,h r ,h s ,h w , where DNN represents the deep neural network function, h x =DNN(f x ) represents the intermediate representation of train operation data, h c =DNN(f c ) represents the intermediate representation of the scheduling command data, h r =DNN(f r ): represents the intermediate representation of track status data, h s =DNN(f s ) represents the intermediate representation of signal information data, h w =DNN(f w ) represents an intermediate representation of meteorological data; S23. Use the self-attention mechanism to perform weighted fusion of different types of intermediate representations; S24. According to the calculated attention weights, the intermediate representations are weighted and summed to form the fused feature representation H: H=a x h x +a c h c +a r h r +a s h s +a w h w ; Among them, H represents the fused feature representation; S25, the fused feature representation H is processed through the fully connected layer to reduce the dimension and obtain a unified data set representation {z k }, where: z k =FC(H); Among them, z k Represents the feature representation after dimensionality reduction; S26. Representation of a unified dataset k } to form the final fusion data set {z ′ k }.

4. According to the railway dispatching safety data aggregation processing method based on AI model training according to claim 1, it is characterized in that: The S3 specifically includes the following steps: S31. The integrated railway dispatching safety data set z ′ k Perform preliminary processing and construct the feature matrix F = {f kj }, where f kj represents the jth feature of the kth sample, the feature is the track condition, z ′ k represents the normalized kth sample, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and F represents the feature matrix of all samples; S32, use the autoencoder method to extract time series features from the feature matrix F, and build an autoencoder model, including the encoder and decoder parts. The encoder takes the input feature f kj Mapped to the hidden layer representation h k , the decoder represents the hidden layer h k Reconstruct back to input features, the optimization objective is: Among them, θ represents the parameters of the autoencoder model, including the encoder parameters θ E and decoder parameters θ D , represents the reconstructed feature matrix, Where E(θ E ,F) represents the output of the encoder, D(θ D ,E(θ E , F) represents the output of the decoder, ∥ represents the norm of the reconstruction error, usually the L2 norm, λ represents the regularization parameter used to control the model complexity, L represents the number of layers of the autoencoder, and W l represents the weight matrix of the lth layer, b l Represents the bias vector of the lth layer; S33. Extracting time series features {h k }: h k =E(θ E ,f k ); Among them, h k Indicates the historical operation mode of the train, the change of track conditions, and the change of signal status. E ,f k ) indicates that the encoder transforms the feature f of the kth sample k The process of mapping to the hidden layer; S34. Use the XGBoost algorithm for feature selection and build a set of decision trees. The objective function is: Among them, Θ represents the parameters of the XGBoost model, N represents the number of samples, that is, the number of samples of railway scheduling safety data, L represents the loss function, which is used to measure the error between the predicted value and the true value, and y i represents the true label of the i-th sample, such as the safety status of train scheduling, represents the predicted value of the i-th sample, T represents the number of decision trees Ω(f t ) represents the regularization term of the t-th tree: Among them, γ represents the regularization parameter, which controls the complexity of the tree, d represents the number of features, and w j represents the weight of feature j; S35. Calculate the importance score of each feature I j ; S36, according to the importance score I j Sort the features, select the top n features with the highest scores, and construct the final feature set F ′ Factors reflecting the safety of railway dispatching, where n is the number of preset features; S37, the final feature set F ′ It serves as input data for subsequent model training to further improve the accuracy and reliability of railway scheduling safety predictions.

5. According to the method of railway dispatching safety data aggregation processing based on AI model training in claim 1, it is characterized in that: The S35 specifically includes: S351, for each iteration t, using the current residual Build a new decision tree and update the model's predictions: in, represents the predicted value of the i-th sample after the t+1th round of iteration, represents the prediction of the railway scheduling safety data by the model after the t+1th round of iteration, η: learning rate, used to control the update amplitude of the predicted value at each iteration, f t (x i ) represents the predicted value of the t-th tree for sample i, and represents the contribution of the t-th decision tree to the railway scheduling safety data; S352. In each round of iteration, the node splitting gain of the current tree is calculated. The gain formula is: Among them, G s represents the splitting gain of node s, and represents the information gain obtained by the current node when splitting. L represents the sample set of the left child node, represents the railway scheduling safety data sample belonging to the left child node after splitting, and I R represents the sample set of the right child node, represents the railway scheduling safety data samples belonging to the right child node after the split, I represents the sample set of the current node, represents all railway scheduling safety data samples before the split, g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. i represents the second-order gradient of the i-th sample, represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error. λ represents the regularization parameter, which is used to control the complexity of the model and avoid overfitting. S353. For each feature j, accumulate its split gain on all nodes and calculate the importance score I j : Among them, I j represents the importance score of the jth feature, represents the importance of feature j in all decision trees, T represents the number of decision trees, represents the total number of decision trees constructed, S t represents the set of all nodes in the t-th tree, represents all nodes in the t-th decision tree, G js represents the splitting gain of the jth feature at node s, and represents the information gain obtained when feature j is split at node s; S354. For each node s, the gain is calculated according to the accumulated value of the first-order gradient and the second-order gradient. The calculation formula is: Among them, G js represents the splitting gain of the jth feature at node s, and represents the contribution of feature j when splitting at node s. s represents the sample set on node s, represents all railway scheduling safety data samples belonging to node s, and g i represents the first-order gradient of the i-th sample, which represents the derivative of the loss function to the predicted value and is used to measure the prediction error. i It represents the second-order gradient of the i-th sample, which represents the second-order derivative of the loss function with respect to the predicted value, and is used to measure the rate of change of the prediction error.

6. According to the railway dispatching safety data aggregation processing method based on AI model training according to claim 1, it is characterized in that: The S4 specifically includes: S41, using the processed feature set F ′ Construct training data set {(x i ,y i )}, where F ′ represents the final feature set, which contains the features with the highest scores. The features represent the key information of railway scheduling safety data. i represents the label of the ith sample, indicating the safety status of railway scheduling, such as safety, potential risk or failure, x i The feature vector representing the i-th sample contains train operation data, dispatching commands, track conditions, signal information, and meteorological data; S42, select the random forest algorithm for model training, and construct multiple decision trees {T m }, where each tree is trained using a different subset of features and samples: {T m }={T1,T2,…,T M }; Among them, {T m } represents the set of decision trees in the random forest, T m : the mth decision tree, M: the number of decision trees, indicating the total number of trees contained in the random forest; S43. Optimize model parameters through hyperparameter tuning technology to determine the best hyperparameter combination {θ k }, including the number of trees M, the maximum depth D, and the minimum number of sample splits N min : {i} k }={θ1,θ2,…,θ K }; Among them, {θ k } represents the hyperparameter set, θ k : The kth hyperparameter, K represents the number of hyperparameters, M represents the number of decision trees, which represents the total number of trees contained in the random forest, D represents the maximum depth of the decision tree, which represents the maximum number of layers of the tree, and N min Indicates the minimum number of sample splits, which indicates the minimum number of samples required for a node to split; S44. Use the grid search method to search for the best parameter combination in the preset parameter space, and the optimization goal is: Among them, θ represents the hyperparameter to be optimized, N: the number of samples, which represents the total number of samples in the training data set, Loss function, measuring the predicted value and the true value y i The difference between S45. Evaluate model performance through cross-validation; S46. Using the best hyperparameter combination Retrain the random forest model to get the final model in, The final random forest model represents the predicted value of the input feature vector x, T m (x): The predicted value of sample x by the mth decision tree. M represents the number of decision trees and the total number of trees in the random forest.

7. According to the method of railway dispatching safety data aggregation processing based on AI model training in claim 1, it is characterized in that: The S45 specifically includes: S451, the training data set {(x i ,y i )} is divided into K non-overlapping subsets, denoted as {D1,D2,…,D K }, where the number of samples contained in each subset is Among them, {(x i ,y i )} represents the training data set, which contains the features and labels of railway dispatching safety data. K represents the number of cross-validation folds, and the training data set is divided into K subsets. k represents the kth subset, each subset contains a part of the railway dispatching safety data samples, N represents the total number of samples in the training data set, and represents the number of samples of the railway dispatching safety data; S452. Perform the following operations on each hyperparameter combination θ: For the kth subset D k , which is used as the validation set, and the remaining K-1 subsets ({D1,D2,…,D k-1 ,D k+1 ,…,D K }Merge into training set: represents the training set in the kth fold, which is composed of the K-1 subsets and contains the features and labels of railway scheduling safety data; represents the validation set in the k-th fold, which consists of the k-th subset and contains the features and labels of the railway scheduling safety data; In the training set Train the random forest model and get the model in, represents the predicted value of the input feature vector x by the random forest model trained on the training set in the kth fold, represents the prediction of railway scheduling safety data, represents the mth decision tree in the kth fold, the predicted value of the input feature vector x, represents the prediction of railway scheduling safety data, M represents the number of decision trees, and represents the total number of trees contained in the random forest; Using a validation set Evaluate model performance and calculate loss function The average value of: Among them, L k represents the average loss in the kth fold, represents the model performance on the validation set, represents the number of samples in the validation set, Represents the model in the kth fold for the validation set sample x i The predicted value of Represents the loss function, measuring the predicted value and the true value y i The difference between is used to evaluate the prediction accuracy of railway dispatching safety data; S453. Calculate the average loss of all compromises as the performance indicator of the hyperparameter combination θ: Among them, L cv (θ) represents the cross-validation average loss of the hyperparameter combination θ, represents the model performance in all folds, K represents the number of cross-validation folds, L k represents the average loss in the kth fold; S454. Select the hyperparameter combination with the minimum cross-validation average loss as the best hyperparameter combination: in, represents the best hyperparameter combination after optimization, represents the kth optimized hyperparameter, L cv (θ) represents the cross-validation average loss for the hyperparameter combination θ.

8. According to the method of railway dispatching safety data aggregation processing based on AI model training in claim 1, it is characterized in that: The S5 specifically includes: S51, collect and pre-process real-time railway dispatching data, including train operation data, dispatching commands, track conditions, signal information and meteorological data, and convert these data into feature vectors x real-time , Among them, x real-time A feature vector representing real-time data, including the latest train status, dispatching orders, track conditions, signal information, and meteorological data; S52, the real-time data feature vector x real-time Input to the trained AI model Generate real-time prediction results in, It indicates the prediction result of the AI ​​model on real-time data and the safety status of the current railway dispatching. Represents a trained AI model, which is used to predict the input feature vector x; S53. Based on the prediction results Analyze potential safety hazards and scheduling risks, and determine whether there are potential risks based on preset thresholds: Among them, Risk represents the real-time risk level, which is classified according to the prediction results, τ high Represents a high risk threshold, exceeding this value indicates high risk, τ low represents a low risk threshold, below which the risk is low; S54. Provide corresponding decision support according to the risk level Risk, including scheduling adjustment suggestions and safety measures suggestions. The specific steps are: when the risk level is high, take emergency scheduling adjustment measures immediately; when the risk level is medium, issue an early warning and recommend the dispatcher to take preventive measures; when the risk level is low, continue monitoring without taking additional measures.

Images