A train operation situation prediction method based on embodied intelligence

Abstract

The application discloses a train operation situation prediction method based on embodied intelligence, which comprises the following steps: collecting multi-source heterogeneous operation data of a train operation process, and performing structured processing to serve as training data; constructing a train delay prediction model based on a gradient boosting tree; storing new operation data with interpretability into a data cache pool; performing incremental updating on the current train delay prediction model; collecting real-time train operation data, inputting the real-time train operation data into the train delay prediction model after incremental updating, and outputting predicted real-time train delay time. The application takes the train as an embodied carrier, captures the environment state in real time through a multi-source sensor, and updates the parameters of the prediction model based on incremental data generated by interaction to form a self-enhancing cycle.

Description

Technical Field

[0001] This invention relates to the field of intelligent rail transit systems, and specifically to a method for predicting train operation status based on embodied intelligence. Background Technology

[0002] Current train operation situation prediction models largely rely on offline-trained historical operation data and fixed parameter sets. Their core flaw lies in the lack of online sensing and dynamic adjustment capabilities for real-time disturbance signals. On one hand, the mismatch between the model update frequency and the disturbance propagation speed causes prediction results to lag behind the actual disturbance evolution process, creating a vicious cycle of "prediction failure → control inaccuracy → delay propagation." On the other hand, traditional batch training methods struggle to meet the real-time requirements of multi-source data fusion, causing prediction confidence to decay over time. Therefore, constructing a situation prediction model with online self-correction capabilities is an inevitable choice to overcome these bottlenecks. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention provides a train operation status prediction method based on embodied intelligence. The method uses the train as an embodied carrier, captures environmental status data of train operation in real time through multi-source sensors, and achieves accurate prediction of train time based on gradient boosting tree.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0005] A train operation status prediction method based on embodied intelligence is provided, which includes the following steps:

[0006] S1: Collect multi-source heterogeneous operational data during train operation, perform structured processing on the multi-source heterogeneous operational data, output multi-dimensional vector features, form a numerical input matrix, and use it as training data.

[0007] S2: Utilize numerical feature vectors from the training data Corresponding actual train delay time Construct a train delay prediction model based on gradient boosting tree;

[0008] S3: Collect new operational data during train operation, extract feature data from the new operational data, verify the interpretability of the new operational data from the perspectives of data validity, causal interpretability, model stability and predictive logic consistency, and store the interpretable new operational data into the data cache pool. When the amount of operational data in the data cache pool reaches the set threshold, execute step S4.

[0009] S4: Based on the newly added operational data in the data cache pool, input the newly added operational data into the current train delay prediction model, obtain the deviation direction and magnitude of the current train delay prediction model, and incrementally update the current train delay prediction model; S5: Collect real-time train operational data, input the real-time operational data into the incrementally updated train delay prediction model, and output the predicted real-time train delay time.

[0010] Further, step S1 includes:

[0011] S11: Collect multi-source heterogeneous operational data during train operation to obtain... N Group Operation Dataset

[0012] , i The number of data entries for each set of operational data. This is a field dataset for operational data, including planned arrival times. Original meteorological text and station name ; This represents the actual train delay time corresponding to the field dataset.

[0013] S12: Define the set of time parsing operator strings Text time Mapped to a scalar;

[0014] ;

[0015] in, For the set of real numbers, For hourly parsing functions, This is a minute parsing function. A scalar extraction function for the planned arrival time; Parse and map the string of arrival time, satisfying the following conditions:

[0016] ;

[0017] in, Arrival time string No. Bit-value mapping This is a string parsing function. k The number of digits in the arrival time string. The set of natural numbers;

[0018] S13: Define the original meteorological text Decomposition mapping ;

[0019] ;

[0020] in, The decomposition mapping function for the original meteorological text. These are the wind speed extraction function and the wind direction extraction function, respectively. These are the original meteorological text space and the wind direction label space, respectively.

[0021] Wind speed extraction function The method for feature extraction from the original meteorological text is as follows:

[0022] ;

[0023] in, The length of the original meteorological text. For numeric character indicator functions, The string position weight index of the original meteorological text. A mapping from characters to numbers. Original meteorological text The One character;

[0024] Wind direction extraction function The method for feature extraction from the original meteorological text is as follows:

[0025] ;

[0026] in, For a preset set of wind directions, This is a substring matching function for the original meteorological text. Wind direction meteorological text string The One character;

[0027] S14: Regarding station names Construct one-hot encoding mapping , Station assembly, station assembly The size is ;

[0028] ;

[0029] in, For station index mapping functions, K For the size of the station collection, The number of bits in the station encoding vector. This is a one-hot encoding mapping function;

[0030] S15: Extract the wind speed feature scalar Wind direction characteristic scalar Scalar of planned arrival time As categorical variable data, construct a high-dimensional sparse tensor , M The total number of categories for the categorical variable;

[0031] S16: Generate high-dimensional numerical feature vectors ;

[0032] ;

[0033] in, This represents a vector concatenation operation. For the normalization function of continuous feature data in operational data, For normalized values ​​of continuous feature data, For continuous feature data, D Numerical feature vectors Dimensions of intrinsic feature data Station name One-hot encoding;

[0034] ;

[0035] in, Continuous feature data The mean and standard deviation;

[0036] S17: Examine the numerical feature vector The system checks whether the feature data is within the valid range, deletes abnormal data that is not within the valid range, obtains multi-dimensional vector features, and forms a numerical input matrix.

[0037] Further, step S2 includes:

[0038] S21: Construction N Numerical feature vectors corresponding to group operation data set ;

[0039] S22: Construction N The set of actual train delay times corresponding to the group of operational data ;

[0040] S23: Construct an ensemble model of gradient boosting trees;

[0041] ;

[0042] in, This is a prediction function for train delay time, used to output the predicted train delay time; m The numbering of the decision tree, MFor the number of decision trees, The shrinkage rate of the decision tree. For the first m The prediction function of a decision tree. The input is a numerical feature vector. Let be the topology of the decision tree, and , For split feature index, Let be the splitting threshold of the split point, and let the leaf nodes of the topology be... , Let be the index number of the split point in the topology, indicating the first _____. The first decision tree Sub-feature splitting operation;

[0043] S24: Initialize the topology of the decision tree, and set the... Numerical feature vectors in Input a decision tree, output the predicted train delay time, and calculate the predicted train delay time. Relative to actual train delay time residual vector ;

[0044] ;

[0045] S25: Utilizing residual vectors The decision tree is partitioned into feature spaces, and a greedy strategy is used to search for the optimal split point in the topology to calculate the numerical feature vector. The Feature data Candidate split points Split gain ;

[0046] ;

[0047] in, , It is the sample index set corresponding to the training data. Left child node sample index set, It is the right child node sample index set. represents the L2 regularization coefficient for the weights of the leaf nodes in the model. This is a complexity penalty term;

[0048] S26: Select the option that maximizes the split gain Feature data and candidate split point combination Execute node splitting;

[0049] S27: Utilizing residual vectors and regularization coefficient Weights are assigned to the leaf nodes of the decision tree, and the resulting leaf node regions are then processed. Calculate the optimal weights of the leaf nodes. ;

[0050] ;

[0051] S28: Utilizing leaf node regions Optimal weight The gradient boosting tree ensemble model is updated and iteratively optimized to generate an extended gradient boosting tree ensemble model, thus obtaining the train delay prediction model.

[0052] ;

[0053] in, This is an indicator function.

[0054] Furthermore, step S2 also includes:

[0055] S29: After each iteration of optimization in the gradient boosting tree ensemble model, output a predicted train delay time and calculate the error between the predicted train delay time and the actual train delay time. ;

[0056] S210: Define the error threshold When continuous P All conditions are met during the round of iterative optimization. If the gradient boosting tree ensemble model is successfully optimized, it indicates that the gradient boosting tree ensemble model has been optimized. The gradient boosting tree ensemble model at this point is then directly output as the train delay prediction model.

[0057] Furthermore, step S2 also includes:

[0058] S211: Define the selection of each decision tree The dimensional feature data is split, and replacement is used in each round of iterative optimization. Each feature data and the corresponding actual train delay time are used as samples;

[0059] S212: Limit the maximum depth of a single decision tree and the minimum number of samples in a leaf node. Use a regularization term embedded in the objective function to jointly optimize the gradient boosting tree ensemble model. The overall loss function of the gradient boosting tree ensemble model during the optimization process is... for:

[0060] ;

[0061] in, Indicates the first The loss function for each sample. for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm; for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm square, Indicates the first The number of leaf nodes in a decision tree.

[0062] Further, step S3 includes:

[0063] S31: Collect new operational data during train operation and extract the corresponding feature data. and verify feature data If the continuous feature compliance, discrete feature legality and missing value tolerance meet the requirements, proceed to step S32; otherwise, delete it as abnormal data.

[0064] S32: Utilizing Feature Data Construct a subset of feature data corresponding to new operational data Using feature data subsets Training a temporary train delay prediction model And use SHAP values ​​to decompose feature data Contribution to prediction results :

[0065] ;

[0066] in, For feature data subset Excluding feature data Any subset of features after that, Representing a subset of feature data Size, Indicates using only subsets Feature-based train delay prediction model The output;

[0067] S33: Verify Contribution Does the dual contribution constraint apply?

[0068] , ;

[0069] in, For absolute contribution threshold, The relative contribution threshold;

[0070] If at least one feature data exists If the dual contribution constraint is satisfied, the new operational data is determined to have interpretable causal value, and step S34 is executed; otherwise, the new operational data is deleted as abnormal data.

[0071] S34: L2 norm normalized results defining feature importance as the SHAP value:

[0072] ;

[0073] in, For reference dataset size, These are the current train delay prediction model and the temporary train delay prediction model, respectively. Corresponding features The normalized importance strength on the reference dataset, To be respectively the first train delay prediction model under the current train delay prediction model Features in each sample With features SHAP value, Temporary train delay prediction models Next Features in each sample With features The SHAP value;

[0074] Calculate importance volatility :

[0075] ;

[0076] If any feature data exists satisfy , If the SHAP value fluctuation is within the allowable range, the model is determined to have experienced a significant logical drift, and the new operational data needs to be deleted as outlier data. Otherwise, the model stability requirements are met, and step S35 is executed.

[0077] Step S35: From history N Group Operation Dataset Randomly select reference sample set And the differences between models are quantified using the path integral method, specifically including:

[0078] Step S351: Perform feature space interpolation:

[0079] For the reference sample set Each sample in Construct from the reference vector to sample Linear path:

[0080] ;

[0081] in, For feature vectors Expected value

[0082] Step S352: Gradient field difference calculation:

[0083] Temporary train delay prediction model along a linear path Integrating the predictions with those of the current train delay prediction model yields the difference in prediction gradients. :

[0084] ;

[0085] in, For the sample Input a temporary train delay prediction model The output predicted value, Input the predicted value from the current train delay prediction model into the sample. It is the gradient operator. It is a path parameter;

[0086] Step S353: Perform prediction logic consistency verification:

[0087] If satisfied If the new operational data meets the consistency requirements of the prediction logic, it is determined that the new operational data is stored in the data cache pool; otherwise, the new operational data is deleted as abnormal data. Tolerance coefficient, This represents the gradient difference distribution during the updating of the historical train delay prediction model.

[0088] Furthermore, in step S31, the feature data is verified. The method for ensuring compliance with continuous features is as follows:

[0089] Validation feature data Is it within a reasonable range of historical distribution? , This is the confidence interval width coefficient. If it is, then the compliance with continuous feature requirements is satisfied; otherwise, the compliance with continuous feature requirements is not satisfied. For the first time in history j The mean of the characteristic data, For the first time in history j The standard deviation of the characteristic data;

[0090] , ;

[0091] in, It is a normalization factor. This refers to the size of the historical dataset.

[0092] Validation feature data The method for validating discrete features is as follows:

[0093] Validation feature data Does it belong to the pre-defined legal set? ,Right now ,in, For feature data Allowed discrete values, It is the total number of valid values;

[0094] If so, then determine the feature data. If the discrete features are valid, then the feature data is deemed invalid. The discrete feature validity requirement is not met.

[0095] Validation feature data The missing value tolerance method is as follows:

[0096] Define the threshold for the proportion of missing samples. With feature-level missing ratio threshold Highly missing samples and anomalous feature data are removed through dual constraints:

[0097] If feature data satisfy:

[0098] ;

[0099] Then determine the feature data If the missing value tolerance requirement is met, otherwise, the feature data is deemed unsuitable. The missing value tolerance requirement is not met; among them, It is a missing value indicator function. This is the total number of samples corresponding to the new operational data. Further, step S4 includes:

[0100] S41: Extract the feature data corresponding to the newly added operational data and input it into the current train delay prediction model, then output the predicted train delay time. and utilize actual train delay time Calculate the residuals corresponding to the newly added operational data. ;

[0101] ;

[0102] in, The output of the decision leaf nodes;

[0103] S42: Utilizing residuals Train a new decision tree and incrementally update the current train delay prediction model through gradient boosting.

[0104] ;

[0105] in, For the newly added L Decision tree, This is the prediction function of the incrementally updated train delay prediction model. This is the prediction function for the current train delay prediction model.

[0106] The beneficial effects of this invention are as follows:

[0107] 1. This invention discloses three methods for the structured processing of raw train operation data: time feature quantification, composite meteorological feature analysis, and one-hot coding of categorical variables. These methods construct a multi-dimensional input matrix that integrates spatiotemporal, environmental, and operational features, significantly improving data representation capabilities.

[0108] 2. The initial model is trained using a gradient boosting tree to predict train delay times. Simultaneously, to support subsequent incremental update validation, the feature contribution matrix is ​​initialized using the SHAP value method to achieve data interpretability analysis and obtain the model logic.

[0109] 3. This invention proposes a four-layer verification framework of data validity, causal interpretability, model stability, and prediction consistency. By using the threshold constraint of SHAP contribution value, it ensures that only new data with clear causal explanations triggers model updates, and stores the filtered data in a data buffer pool.

[0110] 4. Once the data in the data buffer pool reaches a certain amount, new data features are gradually absorbed by adding new decision trees. During the training process, a dynamically updated validation set is continuously used to monitor the evolution direction of the model, ensuring that incremental updates are trained in a more accurate direction.

[0111] This invention constructs an embodied intelligence-emphasized intelligent agent that continuously interacts with the physical environment through a real-time perception-action closed loop, dynamically optimizing its own cognitive model in the process. The train, as the embodied carrier, captures the environmental state in real time through multi-source sensors, while the predictive model updates parameters online based on incremental data generated by the interaction, forming a self-reinforcing loop. Attached Figure Description

[0112] Figure 1 This is a flowchart of a train operation status prediction method based on embodied intelligence.

[0113] Figure 2This is a block diagram illustrating the principle of training and incremental updating of the train delay prediction model. Detailed Implementation

[0114] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0115] like Figure 1 As shown, a train operation status prediction method based on embodied intelligence includes the following steps:

[0116] S1: Collect multi-source heterogeneous operational data during train operation, perform structured processing on the multi-source heterogeneous operational data, output multi-dimensional vector features, form a numerical input matrix, and use it as training data.

[0117] This involves the construction and dynamic updating of a train delay prediction model, one of the core implementation steps of which is the structured processing of multi-source heterogeneous operational data. The raw data encompasses multimodal features such as textual time stamps, unstructured meteorological descriptions, discrete station codes, and continuous sensor readings. It is necessary to transform this data into a standardized expression in a high-dimensional numerical tensor space by defining rigorous analytical operators and mapping functions, providing processable mathematical input for subsequent machine learning models. Specifically, the transformation from raw data to the prediction model is achieved through time feature quantization functions, meteorological decoupling functions, and category embedding spaces, ensuring the completeness of the data representation.

[0118] Step S1 specifically includes:

[0119] S11: Collect multi-source heterogeneous operational data during train operation to obtain... N Group Operation Dataset

[0120] , i The number of data entries for each set of operational data. This is a field dataset for operational data, including planned arrival times. Original meteorological text and station name ; This represents the actual train delay time corresponding to the field dataset.

[0121] S12: Define the set of time parsing operator strings Text time Mapped to a scalar;

[0122] ;

[0123] in, For the set of real numbers, For hourly parsing functions, This is a minute parsing function. A scalar extraction function for the planned arrival time; Parse and map the string of arrival time, satisfying the following conditions:

[0124] ;

[0125] in, Arrival time string No. Bit-value mapping This is a string parsing function. k The number of digits in the arrival time string. The set of natural numbers;

[0126] S13: Define the original meteorological text Decomposition mapping ;

[0127] ;

[0128] in, The decomposition mapping function for the original meteorological text. These are the wind speed extraction function and the wind direction extraction function, respectively. These are the original meteorological text space and the wind direction label space, respectively.

[0129] Wind speed extraction function The method for feature extraction from the original meteorological text is as follows:

[0130] ;

[0131] in, The length of the original meteorological text. For numeric character indicator functions, The string position weight index of the original meteorological text. A mapping from characters to numbers. Original meteorological text The One character;

[0132] Wind direction extraction function The method for feature extraction from the original meteorological text is as follows:

[0133] ;

[0134] in, For a preset set of wind directions, This is a substring matching function for the original meteorological text. Wind direction meteorological text string The One character;

[0135] S14: Regarding station names Construct one-hot encoding mapping , Station assembly, station assembly The size is ;

[0136]

[0137] in, For station index mapping functions, K For the size of the station collection, The number of bits in the station encoding vector. This is a one-hot encoding mapping function;

[0138] S15: Extract the wind speed feature scalar Wind direction characteristic scalar Scalar of planned arrival time As categorical variable data, construct a high-dimensional sparse tensor , M The total number of categories for the categorical variable;

[0139] S16: Generate high-dimensional numerical feature vectors ;

[0140] ;

[0141] in, This represents a vector concatenation operation. For the normalization function of continuous feature data in operational data, For normalized values ​​of continuous feature data, For continuous feature data, D Numerical feature vectors Dimensions of intrinsic feature data Station name One-hot encoding;

[0142] ;

[0143] in, Continuous feature data The mean and standard deviation;

[0144] S17: Examine the numerical feature vector The system checks whether the feature data is within the valid range, deletes abnormal data that is not within the valid range, obtains multi-dimensional vector features, and forms a numerical input matrix.

[0145] For example, the scalar of the arrival time of this inspection plan. To determine if the data is within a valid range, check whether the continuous feature data is within a physically reasonable range.

[0146] ;

[0147] in, For operational datasets that are within the legal range, For continuous feature data Within the legal scope.

[0148] like Figure 2 As shown, S2: Utilizing numerical feature vectors from the training data Corresponding actual train delay time Construct a train delay prediction model based on gradient boosting trees; Step S2 specifically includes the following steps:

[0149] S21: Construction N Numerical feature vectors corresponding to group operation data set ;

[0150] S22: Construction N The set of actual train delay times corresponding to the group of operational data ;

[0151] S23: Construct an ensemble model of gradient boosting trees;

[0152] ;

[0153] in, This is a prediction function for train delay time, used to output the predicted train delay time; m The numbering of the decision tree, M For the number of decision trees, The shrinkage rate of the decision tree. , For the first m The prediction function of a decision tree. The input is a numerical feature vector. Let be the topology of the decision tree, and , For the splitting feature index of the splitting point, The splitting threshold is given, and the leaf nodes of the topology are... , Let be the index number of the split point in the topology, indicating the first _____. The first decision tree Sub-feature splitting operation;

[0154] S24: Initialize the topology of the decision tree, and set the... Numerical feature vectors in Input a decision tree, output the predicted train delay time, and calculate the predicted train delay time. Relative to actual train delay time residual vector ;

[0155] ;

[0156] S25: Utilizing residual vectors The decision tree is partitioned into feature spaces, and a greedy strategy is used to search for the optimal split point in the topology to calculate the numerical feature vector. The Feature data Candidate split points Split gain ;

[0157] ;

[0158] in, , It is the sample index set corresponding to the training data. Left child node sample index set, It is the right child node sample index set. represents the L2 regularization coefficient for the weights of the leaf nodes in the model. This is a complexity penalty term;

[0159] S26: Select the option that maximizes the split gain Feature data and candidate split point combination Execute node splitting;

[0160] S27: Utilizing residual vectors and regularization coefficient Weights are assigned to the leaf nodes of the decision tree, and the generated leaf node regions are processed. Calculate the optimal weights of the leaf nodes. ;

[0161] ;

[0162] S28: Utilizing leaf node regions Optimal weight The gradient boosting tree ensemble model is updated and iteratively optimized to generate an extended gradient boosting tree ensemble model, thus obtaining the train delay prediction model.

[0163] ;

[0164] in, This is an indicator function.

[0165] Step S2 also includes:

[0166] S29: After each iteration of optimization in the gradient boosting tree ensemble model, output a predicted train delay time and calculate the error between the predicted train delay time and the actual train delay time. ;

[0167] S210: Define the error threshold When continuous P All conditions are met during the round of iterative optimization. This indicates that the gradient boosting tree ensemble model optimization is complete, and the current gradient boosting tree ensemble model is directly output as the train delay prediction model. Step S2 also includes:

[0168] S211: Define the selection of each decision tree The dimensional feature data is split, and replacement is used in each round of iterative optimization. Each feature data and the corresponding actual train delay time are used as samples;

[0169] S212: Limit the maximum depth of a single decision tree and the minimum number of samples in a leaf node. Use a regularization term embedded in the objective function to jointly optimize the gradient boosting tree ensemble model. The overall loss function of the gradient boosting tree ensemble model during the optimization process is... for:

[0170] ;

[0171] in, Indicates the first The loss function for each sample. for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm; for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm square, Indicates the first The number of leaf nodes in a decision tree.

[0172] S3: Collect new operational data during train operation, extract feature data from the new operational data, verify the interpretability of the new operational data from the perspectives of data validity, causal interpretability, model stability, and predictive logic consistency, and store the interpretable new operational data into the data cache pool. When the amount of operational data in the data cache pool reaches the set threshold, execute step S4.

[0173] Step S3 specifically includes the following steps:

[0174] S31: Collect new operational data during train operation and extract the corresponding feature data. and verify feature data If the continuous feature compliance, discrete feature legality and missing value tolerance meet the requirements, proceed to step S32; otherwise, delete it as abnormal data.

[0175] Validation feature data The method for ensuring compliance with continuous features is as follows:

[0176] Evaluation and verification feature data Is it within a reasonable range of historical distribution? , This is the confidence interval width coefficient. If it is, then the compliance with continuous feature requirements is satisfied; otherwise, the compliance with continuous feature requirements is not satisfied. For the first time in history j The mean of the characteristic data, For the first time in history j The standard deviation of the characteristic data;

[0177] , ;

[0178] in, It is a normalization factor. This refers to the size of the historical dataset;

[0179] Validation feature data The method for validating discrete features is as follows:

[0180] Validation feature data Does it belong to the pre-defined legal set? ,Right now ,in, For feature data Allowed discrete values, It is the total number of valid values;

[0181] If so, then determine the feature data. If the discrete features are valid, then the feature data is deemed invalid. The discrete feature validity requirement is not met.

[0182] Validation feature data The missing value tolerance method is as follows:

[0183] Define the threshold for the proportion of missing samples. With feature-level missing ratio threshold Highly missing samples and outlier data are removed through dual constraints:

[0184] If feature data satisfy:

[0185] ;

[0186] Then determine the feature data If the missing value tolerance requirement is met, otherwise, the feature data is deemed unsuitable. The missing value tolerance requirement is not met; among them, It is a missing value indicator function. This is the total number of samples corresponding to the new operational data.

[0187] S32: Utilizing Feature Data Construct a subset of feature data corresponding to new operational data Using feature data subsets Training a temporary train delay prediction model And use SHAP values ​​to decompose feature data Contribution to prediction results :

[0188] ;

[0189] in, For feature data subset Excluding feature data Any subset of features after that, Representing a subset of feature data Size, Indicates using only subsets Feature-based train delay prediction model The output;

[0190] S33: Verify Contribution Does the dual contribution constraint apply?

[0191] , ;

[0192] in, To establish an absolute contribution threshold, the key features must have a significant impact on the quantization of latency. The relative contribution threshold is used to ensure that its proportion in the overall feature importance is not less than [a certain percentage]. .

[0193] If at least one feature data exists If the dual contribution constraint is met, the new operational data is determined to have interpretable causal value, and step S34 is executed; otherwise, the new operational data is deleted as abnormal data.

[0194] S34: Model stability verification focuses on the logical consistency of the new and old models on key features to prevent abrupt changes in prediction logic caused by data distribution drift.

[0195] Define the feature importance index as the L2 norm normalized result of the SHAP value:

[0196] ;

[0197] in, For reference dataset size, These are the current train delay prediction model and the temporary train delay prediction model, respectively. Corresponding features The normalized importance strength on the reference dataset, To be respectively the first train delay prediction model under the current train delay prediction model Features in each sample With features SHAP value, Temporary train delay prediction models Next Features in each sample With features The SHAP value;

[0198] Calculate importance volatility :

[0199] ;

[0200] If any feature data exists satisfy , If the SHAP value fluctuation is within the allowable range, the model is determined to have experienced a significant logical drift, and the new operational data needs to be deleted as outlier data. Otherwise, the model stability requirements are met, and step S35 is executed.

[0201] Prediction logic consistency verification, as the final layer of verification, ensures the continuity of prediction behavior between the old and new models from the function space level.

[0202] Step S35: From history N Group Operation Dataset Randomly select reference sample set And the differences between models are quantified using the path integral method, specifically including:

[0203] Step S351: Perform feature space interpolation:

[0204] For the reference sample set Each sample in Construct from the reference vector to sample Linear path:

[0205] ;

[0206] in, For feature vectors Expected value;

[0207] Step S352: Gradient field difference calculation:

[0208] Temporary train delay prediction model along a linear path Integrating the predictions with those of the current train delay prediction model yields the difference in prediction gradients. :

[0209] ;

[0210] in, For the sample Input a temporary train delay prediction model The output predicted value, Input the predicted value from the current train delay prediction model into the sample. It is the gradient operator. It is a path parameter;

[0211] Step S353: Perform predictive logic consistency verification;

[0212] If satisfied If the new operational data meets the consistency requirements of the prediction logic, it is determined that the new operational data is stored in the data cache pool; otherwise, the new operational data is deleted as abnormal data. Tolerance coefficient, This represents the gradient difference distribution during the updating of the historical train delay prediction model. Consistency criterion: requires at least... The reference sample satisfies this requirement.

[0213] in Tolerance coefficient, This represents the gradient difference distribution during historical model updates. If it exceeds... If the sample trigger threshold is reached, it is determined that there is an unacceptable mutation in the prediction logic, and the data cannot be used as new data for incremental updates.

[0214] New data that has passed four layers of verification is stored in the data cache pool as new data with interpretable properties. When the amount of data in the data cache pool reaches a certain threshold, the incremental update operation of the model is executed.

[0215] S4: Based on the newly added operational data in the data cache pool, input the new operational data into the current train delay prediction model, obtain the direction and magnitude of the prediction deviation of the current train delay prediction model, and incrementally update the current train delay prediction model. Step S4 specifically includes the following steps:

[0216] S41: Extract the feature data corresponding to the newly added operational data and input it into the current train delay prediction model, then output the predicted train delay time. and utilize actual train delay time Calculate the residuals corresponding to the newly added operational data. ;

[0217] ;

[0218] in, The output of the decision leaf nodes;

[0219] S42: Utilizing residuals Train a new decision tree and incrementally update the current train delay prediction model through gradient boosting.

[0220] ;

[0221] in, For the newly added L Decision tree, This is the prediction function of the incrementally updated train delay prediction model. Let be the prediction function of the current train delay prediction model, and use the learning rate. Control the incremental update magnitude.

[0222] S5: Collects real-time train operation data, inputs the real-time operation data into the incrementally updated train delay prediction model, and outputs the predicted real-time train delay time.

[0223] This invention constructs an embodied intelligence-emphasized intelligent agent that continuously interacts with the physical environment through a real-time perception-action closed loop, dynamically optimizing its own cognitive model in the process. The train, as the embodied carrier, captures the environmental state in real time through multi-source sensors, while the predictive model updates parameters online based on incremental data generated by the interaction, forming a self-reinforcing loop.

Claims

1. A train operation status prediction method based on embodied intelligence, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous operational data during train operation, perform structured processing on the multi-source heterogeneous operational data, output multi-dimensional vector features, form a numerical input matrix, and use it as training data. S2: Utilize numerical feature vectors from the training data Corresponding actual train delay time Construct a train delay prediction model based on gradient boosting tree; S3: Collect new operational data during train operation, extract feature data from the new operational data, verify the interpretability of the new operational data from the perspectives of data validity, causal interpretability, model stability and predictive logic consistency, and store the interpretable new operational data into the data cache pool. When the amount of operational data in the data cache pool reaches the set threshold, execute step S4. S4: Based on the newly added operational data in the data cache pool, input the newly added operational data into the current train delay prediction model, obtain the deviation direction and magnitude of the current train delay prediction model, and incrementally update the current train delay prediction model; S5: Collect real-time train operational data, input the real-time operational data into the incrementally updated train delay prediction model, and output the predicted real-time train delay time. Step S3 includes: S31: Collect new operational data during train operation and extract the corresponding feature data. and verify feature data If the continuous feature compliance, discrete feature legality and missing value tolerance meet the requirements, proceed to step S32; otherwise, delete it as abnormal data. S32: Utilizing Feature Data Construct a subset of feature data corresponding to new operational data Using feature data subsets Training a temporary train delay prediction model And use SHAP values ​​to decompose feature data Contribution to prediction results : ； in, For feature data subset Excluding feature data Any subset of features after that, Representing a subset of feature data Size, Indicates using only subsets Feature-based train delay prediction model The output; S33: Verify Contribution Does the dual contribution constraint apply? ， ； in, For absolute contribution threshold, The relative contribution threshold, This represents the total number of samples corresponding to the new operational data. If at least one feature data exists If the dual contribution constraint is satisfied, the new operational data is determined to have interpretable causal value, and step S34 is executed; otherwise, the new operational data is deleted as abnormal data. S34: L2 norm normalized results defining feature importance as the SHAP value: ； in, For reference dataset size, These are the current train delay prediction model and the temporary train delay prediction model, respectively. Corresponding features The normalized importance strength on the reference dataset, To be respectively the first train delay prediction model under the current train delay prediction model Features in each sample With features SHAP value, Temporary train delay prediction models Next Features in each sample With features The SHAP value; Calculate importance volatility : ； If any feature data exists satisfy , If the SHAP value fluctuation is within the allowable range, the model is determined to have experienced a significant logical drift, and the new operational data needs to be deleted as outlier data. Otherwise, the model stability requirements are met, and step S35 is executed. Step S35: From history N Group Operation Dataset Randomly select reference sample set And the differences between models are quantified using the path integral method. i The number of data entries for each set of operational data. This is a field dataset for operational data, including planned arrival times. Original meteorological text and station name , This represents the actual train delay time corresponding to the field dataset. Specifically, it includes: Step S351: Perform feature space interpolation: For the reference sample set Each sample in Construct from the reference vector to sample Linear path: ； in, For feature vectors Expected value; Step S352: Gradient field difference calculation: Temporary train delay prediction model along a linear path Integrating the predictions with those of the current train delay prediction model yields the difference in prediction gradients. : ； in, For the sample Input a temporary train delay prediction model The output predicted value, Input the predicted value from the current train delay prediction model into the sample. It is the gradient operator. It is a path parameter; Step S353: Perform prediction logic consistency verification: If satisfied If the new operational data meets the consistency requirements of the prediction logic, it is determined that the new operational data is stored in the data cache pool; otherwise, the new operational data is deleted as abnormal data. Tolerance coefficient, This represents the gradient difference distribution during the updating of the historical train delay prediction model.

2. The train operation status prediction method based on embodied intelligence according to claim 1, characterized in that, Step S1 includes: S11: Collect multi-source heterogeneous operational data during train operation to obtain... N Group Operation Dataset ； S12: Define the set of time parsing operator strings Text time Mapped to a scalar; ； in, For the set of real numbers, For hourly parsing functions, This is a minute parsing function. A scalar extraction function for the planned arrival time; Parse and map the string of arrival time, satisfying the following conditions: ； in, Arrival time string No. Bit-value mapping This is a string parsing function. k The number of digits in the arrival time string. The set of natural numbers; S13: Define the original meteorological text Decomposition mapping ; ； in, The decomposition mapping function for the original meteorological text. These are the wind speed extraction function and the wind direction extraction function, respectively. These are the original meteorological text space and the wind direction label space, respectively. Wind speed extraction function The method for feature extraction from the original meteorological text is as follows: ； in, The length of the original meteorological text. For numeric character indicator functions, The string position weight index of the original meteorological text. A mapping from characters to numbers. Original meteorological text The One character; Wind direction extraction function The method for feature extraction from the original meteorological text is as follows: ； in, For a preset set of wind directions, This is a substring matching function for the original meteorological text. Wind direction meteorological text string The One character; S14: Regarding station names Construct one-hot encoding mapping , Station assembly, station assembly The size is ; ； in, For station index mapping functions, K For the size of the station collection, The number of bits in the station encoding vector. This is a one-hot encoding mapping function; S15: Extract the wind speed feature scalar Wind direction characteristic scalar Scalar of planned arrival time As categorical variable data, construct a high-dimensional sparse tensor , M The total number of categories for the categorical variable; S16: Generate high-dimensional numerical feature vectors ; ； in, This represents a vector concatenation operation. For the normalization function of continuous feature data in operational data, For normalized values ​​of continuous feature data, For continuous feature data, D Numerical feature vectors Dimensions of intrinsic feature data Station name One-hot encoding; ； in, Continuous feature data The mean and standard deviation; S17: Examine the numerical feature vector The system checks whether the feature data is within the valid range, deletes abnormal data that is not within the valid range, obtains multi-dimensional vector features, and forms a numerical input matrix.

3. The train operation status prediction method based on embodied intelligence according to claim 2, characterized in that... Therefore, step S2 includes: S21: Construction N Numerical feature vectors corresponding to group operation data set ; S22: Construction N The set of actual train delay times corresponding to the group of operational data ; S23: Construct an ensemble model of gradient boosting trees; ； in, This is a prediction function for train delay time, used to output the predicted train delay time; m The numbering of the decision tree, M For the number of decision trees, The shrinkage rate of the decision tree. For the first m The prediction function of a decision tree. The input is a numerical feature vector. Let be the topology of the decision tree, and , For split feature index, Let be the splitting threshold of the split point, and let the leaf nodes of the topology be... , Let be the index number of the split point in the topology, indicating the first _____. The first decision tree Sub-feature splitting operation; S24: Initialize the topology of the decision tree, and set the... Numerical feature vectors in Input a decision tree, output the predicted train delay time, and calculate the predicted train delay time. Relative to actual train delay time residual vector ; ； S25: Utilizing residual vectors The decision tree is partitioned into feature spaces, and a greedy strategy is used to search for the optimal split point in the topology to calculate the numerical feature vector. The Feature data Candidate split points Split gain ; ； in, , It is the sample index set corresponding to the training data. Left child node sample index set, It is the right child node sample index set. represents the L2 regularization coefficient for the weights of the leaf nodes in the model. This is a complexity penalty term; S26: Select the option that maximizes the split gain Feature data and candidate split point combination Execute node splitting; S27: Utilizing residual vectors and regularization coefficient Weights are assigned to the leaf nodes of the decision tree, and the resulting leaf node regions are then processed. Calculate the optimal weights of the leaf nodes. ; ； S28: Utilizing leaf node regions Optimal weight The gradient boosting tree ensemble model is updated and iteratively optimized to generate an extended gradient boosting tree ensemble model, thus obtaining the train delay prediction model. ； in, This is an indicator function.

4. The train operation status prediction method based on embodied intelligence according to claim 3, characterized in that, Step S2 further includes: S29: After each iteration of optimization in the gradient boosting tree ensemble model, output a predicted train delay time and calculate the error between the predicted train delay time and the actual train delay time. ; S210: Define the error threshold When continuous P All conditions are met during the round of iterative optimization. If the gradient boosting tree ensemble model is successfully optimized, it indicates that the gradient boosting tree ensemble model has been optimized. The gradient boosting tree ensemble model at this point is then directly output as the train delay prediction model.

5. The train operation status prediction method based on embodied intelligence according to claim 3, characterized in that, Step S2 further includes: S211: Define the selection of each decision tree The dimensional feature data is split, and replacement is used in each round of iterative optimization. Each feature data and the corresponding actual train delay time are used as samples; S212: Limit the maximum depth of a single decision tree and the minimum number of samples in a leaf node. Use a regularization term embedded in the objective function to jointly optimize the gradient boosting tree ensemble model. The overall loss function of the gradient boosting tree ensemble model during the optimization process is... for: ； in, Indicates the first The loss function for each sample. for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm; for Regularization coefficient, For the first The weights of all leaf nodes in the decision tree Norm square, Indicates the first The number of leaf nodes in a decision tree.

6. The train operation status prediction method based on embodied intelligence according to claim 1, characterized in that, In step S31, the feature data is verified. The method for ensuring compliance with continuous features is as follows: Validation feature data Is it within a reasonable range of historical distribution? , This is the confidence interval width coefficient. If it is, then the compliance with continuous feature requirements is satisfied; otherwise, the compliance with continuous feature requirements is not satisfied. For the first time in history j The mean of the characteristic data, For the first time in history j The standard deviation of the characteristic data; ， ； in, It is a normalization factor. This refers to the size of the historical dataset; Validation feature data The method for validating discrete features is as follows: Validation feature data Does it belong to the pre-defined legal set? ,Right now ,in, For feature data Allowed discrete values, It is the total number of valid values; If so, then determine the feature data. If the discrete features are valid, then the feature data is deemed invalid. The discrete feature validity requirement is not met. Validation feature data The missing value tolerance method is as follows: Define the threshold for the proportion of missing samples. With feature-level missing ratio threshold Highly missing samples and anomalous feature data are removed through dual constraints: If feature data satisfy: ； Then determine the feature data If the missing value tolerance requirement is met, otherwise, the feature data is deemed unsuitable. The missing value tolerance requirement is not met; among them, It is a missing value indicator function.

7. The train operation status prediction method based on embodied intelligence according to claim 3, characterized in that, Step S4 includes: S41: Extract the feature data corresponding to the newly added operational data and input it into the current train delay prediction model, then output the predicted train delay time. and utilize actual train delay time Calculate the residuals corresponding to the newly added operational data. ; ； in, The output of the decision leaf nodes; S42: Utilizing residuals Train a new decision tree and incrementally update the current train delay prediction model through gradient boosting. ； in, For the newly added L Decision tree, This is the prediction function of the incrementally updated train delay prediction model. This is the prediction function for the current train delay prediction model.

Images