Abnormal monitoring and early warning method for motion control state of road train

Abstract

The invention discloses an abnormal monitoring and early warning method for a road train motion control state, and relates to the technical field of vehicle operation monitoring. According to the method, multi-mode state data of a head train and a vehicle-mounted train of a road train are collected through a multi-source sensor, and preprocessing such as standardization, missing value filling and abnormal value elimination is carried out; the category data and the numeric data are fused through learnable embedded codes, and a unified high-dimensional feature vector is constructed; according to the method, a historical fusion feature sequence is obtained, a Transform model is used for modeling the historical fusion feature sequence, a key state value at a future moment is predicted, and finally, real-time judgment and early warning of an abnormal state are achieved by comparing whether the deviation between a predicted value and a real-time observation value exceeds a threshold value or not. According to the method, the time sequence modeling and feature extraction capabilities of the Transform model are fully utilized, the problems that multi-source heterogeneous data fusion is difficult and the complex time sequence dependency relationship modeling capability is insufficient in the road train state monitoring are effectively solved, and high-precision and high-real-time anomaly monitoring and early warning are realized.

Description

Technical Field

[0001] This invention relates to the field of highway train operation monitoring technology, and in particular to a method for abnormal monitoring and early warning of the motion control status of highway trains. Background Technology

[0002] Highway trains consist of a lead car and multiple accompanying cars, each containing numerous state variables, resulting in a vast number of interconnected and complex state variables. The operating environment of highway trains is complex and variable, with state variable data exhibiting complex temporal characteristics and non-stationarity, characterized by dramatic dynamic changes and long-term dependencies. Existing technologies struggle to achieve joint prediction when dealing with these complex state variables, failing to fully consider the interactions between them, and thus resulting in insufficient monitoring accuracy. The state variable data sources for highway trains are diverse, with complex and heterogeneous data types, posing challenges for data fusion and feature extraction in existing technologies, hindering the provision of a high-quality data foundation for subsequent monitoring and prediction. Existing monitoring methods suffer from delays in data acquisition and analysis, failing to meet real-time requirements and hindering timely detection and early warning of abnormal states. In normal state modeling, existing technologies rely on simplified assumptions and historical data, making it difficult to adapt to the actual conditions of highway trains under various operating conditions, leading to discrepancies between predicted and actual results.

[0003] Application number CN202411342354.4 uses time window filtering technology to optimize initial vehicle measurement data and combines neighborhood mutual information and automatic encoders for feature fusion to monitor vehicle energy consumption. Application number CN202411918077.7 collects vibration signals through sensors, transforms them in the time and frequency domains, and uses specific algorithms to analyze spectrograms and spectrum widths to achieve gearbox anomaly monitoring and alarm. However, research on monitoring the overall motion control state of highway trains has not yet been carried out. This invention analyzes data under normal operating conditions of highway trains, predicts data for the next moment, and compares real-time data with predicted data for anomaly monitoring, thereby achieving accurate and real-time monitoring of the motion control state of highway trains and providing reliable protection for the safe operation of highway trains. Summary of the Invention

[0004] In view of this, this invention addresses the shortcomings of existing methods for monitoring the motion control status of highway trains in terms of data processing, real-time monitoring, and modeling. It proposes a Transformer-based method for monitoring and early warning of abnormal motion control status in highway trains. This method constructs a data prediction model under normal operating conditions to predict the status data at future times and compares it with real-time collected data to achieve accurate identification and early warning of abnormal states. This method fully utilizes the advantages of the Transformer model in temporal modeling and feature extraction, possessing high accuracy, strong real-time performance, and good environmental adaptability. It effectively overcomes the limitations of existing technologies in joint prediction of state variables, dynamic feature capture, and multi-source data fusion.

[0005] This invention provides a method for monitoring and early warning of abnormalities in the motion control state of highway trains, comprising the following steps: S1 deploys multiple types of sensors on key components of the lead car and each accompanying car of the highway train to collect key state variable data in real time, including speed, acceleration, braking pressure and steering angle. S2, preprocess the collected key state variable data to obtain processed numerical data and categorical data; S3, the processed numerical data obtained in step S2 is used as the numerical feature vector; the processed categorical data is input into the embedding layer after independently building a vocabulary and performing label encoding, and mapped into a low-dimensional dense vector to obtain the categorical data embedding vector; the numerical feature vector and the categorical data embedding vector are concatenated in the feature dimension to obtain the fused feature vector at each time step. S4. A time-series prediction model is constructed using a Transformer encoder-decoder architecture. The continuous fused feature vector sequence within the historical time window is used as the input of the time-series prediction model, and positional encoding information is introduced. The time-series prediction model is used to predict numerical data at multiple future times and outputs a sequence of predicted numerical feature values ​​for future times. S5. Using the sliding window method, time series training samples are constructed based on the fused feature vector sequence obtained in step S3, and the time series prediction model described in step S4 is trained with the goal of predicting future numerical data. S6, using the trained time-series prediction model to monitor highway trains in real time: execute steps S2 to S3 on the real-time collected data to generate the current fused feature vector and update the historical sequence buffer queue; input the updated historical sequence buffer queue into the time-series prediction model to obtain the numerical feature prediction values ​​for future moments; compare the actual observation value at the next moment with the corresponding prediction value and calculate the absolute deviation; when the absolute deviation of any key state variable exceeds its preset threshold, it is determined to be an abnormal state and an early warning is triggered.

[0006] Optionally, in step S2, the preprocessing includes: standardizing the numerical data, imputing missing values ​​in the categorical data, and identifying and replacing outliers in the numerical data based on statistical methods.

[0007] Optionally, in step S2, the identification and replacement of outliers in the numerical data based on statistical methods specifically includes: calculating the local mean and standard deviation of the data within a sliding time window; identifying data points whose absolute deviation from the local mean exceeds three times the standard deviation as outliers; and replacing the outliers with the median within the sliding time window.

[0008] Optionally, in step S4, the introduction of position encoding information specifically involves: using sine-cosine position encoding, adding the position information to the fused feature vector, and then inputting it into the model.

[0009] Optionally, in step S5, the construction of time series training samples specifically involves: setting the time window length, prediction step size, and sliding step size, and extracting samples from the fused feature vector sequence in chronological order; the input of each sample is a historical fused feature vector sequence, and its corresponding true output is numerical data from multiple subsequent consecutive time points.

[0010] Optionally, in step S5, the loss function for training the time series prediction model is the mean squared error.

[0011] Optionally, in step S6, when the early warning mechanism is triggered, the timestamp of the anomaly, the vehicle identification involved, the specific state variables and their values ​​for the deviation exceeding the limit, the corresponding historical data segments, and the comparison information between the predicted value and the actual value are recorded simultaneously.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects: 1. The method of this invention effectively extracts the dynamic correlation features between state variables through the self-attention mechanism in step S4, significantly improving the model's ability to model the system's evolution trend, thereby enhancing the accuracy and reliability of anomaly monitoring and strengthening the safety of highway train operation. It solves the problems in existing technologies regarding the difficulty in modeling complex time-series data relationships and the difficulty in capturing long-term dependent features in highway train motion control state monitoring.

[0013] 2. The method of this invention combines a multimodal data processing framework of z-score standardization, text feature extraction, and embedding encoding. This framework enables effective and unified representation of different types of data, improves data usability and analysis accuracy, provides high-quality data support for subsequent time series prediction and anomaly detection, and enhances the overall monitoring and early warning capabilities of the system. It solves the problems of data fusion difficulties and insufficient feature extraction in traditional monitoring methods when processing multi-source heterogeneous data.

[0014] 3. The method of this invention employs an adaptive learning rate dynamic adjustment mechanism. This mechanism intelligently adjusts the learning rate based on the loss changes and convergence status during model training, effectively improving the model's convergence speed, suppressing overfitting, and significantly enhancing the accuracy of fault monitoring and the robustness of the early warning system. It solves the problems of low model training efficiency and unstable prediction performance in existing highway train fault monitoring systems caused by the diversity and complexity of data. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the Transformer model architecture of the present invention; Figure 2 A schematic diagram of the collected highway train data; Figure 3 This relates to the construction of the time window in the method of the present invention; Figure 4 This is a schematic diagram of forward filling in the method of the present invention. Detailed Implementation

[0016] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0017] A specific embodiment of the present invention, such as Figures 1-4 This paper discloses a method for monitoring and early warning of abnormal motion control status of highway trains based on Transformer. This method possesses high accuracy, strong real-time performance, and good environmental adaptability, and is suitable for status monitoring and early warning of highway trains under complex operating conditions. The method includes the following steps: Step 1: Deploy multiple types of sensors on the lead car and key components of each car in the highway train to collect real-time data on key state variables, including speed, acceleration, braking pressure, and steering angle. Specifically, the key components include: The steering system of the tractor (head vehicle): A steering angle sensor is installed at the steering knuckle or steering tie rod to monitor the steering angle of the front wheels in real time; Braking system of each vehicle (including lead vehicle and all accompanying vehicles): Pressure sensors are installed in the brake chamber or brake line to collect the braking pressure signal of each axle and reflect the braking force distribution and response status. Near the chassis or center of gravity of each vehicle: Install inertial measurement units (IMUs) or acceleration sensors to synchronously acquire longitudinal, lateral, and vertical acceleration data; Wheel hubs or transmission systems of each vehicle: Install wheel speed sensors or read vehicle speed signals via CAN bus to calculate and output vehicle speed and the speed difference between each axle, thereby assisting in determining whether sideslip, folding or instability has occurred; For vehicles equipped with electronic control systems (such as smart trailers), auxiliary status parameters such as gear position, ABS activation signal, and vehicle attitude angle can be obtained through the on-board controller (ECU) as supplementary feature inputs.

[0018] The sensors transmit data to the central data processing unit in real time via the vehicle CAN bus, Ethernet, or wireless communication module, ensuring that the status information of the lead vehicle and each accompanying vehicle is strictly synchronized in time, providing a highly consistent data foundation for the subsequent construction of global motion control status characteristics.

[0019] Step 2: Preprocess the collected key state variable data to obtain processed numerical and categorical data.

[0020] Specifically, preprocessing includes data standardization, missing value imputation, and outlier removal, ensuring the accuracy and integrity of the data; Key state variable data includes two types of data: (1) Numerical data: including continuous physical quantities or quantifiable indicators, such as vehicle speed (km / h), longitudinal acceleration, lateral acceleration (m / s²), (lead vehicle only) steering angle (°), yaw rate (° / s), etc. These data are all real-valued time-series signals, suitable for statistical analysis and deep learning model input.

[0021] (2) Categorical data (i.e., text data): This includes discrete state variables such as vehicle type identifiers (e.g., “lead vehicle”, “trailer 1”, “trailer 2”), braking system operating modes (e.g., “conventional braking”, “emergency braking”, “retarder activated”), and ABS status (“activated” / “inactive”). These variables exist raw in string or integer form and need to be encoded.

[0022] First, the numerical data (continuous physical quantities) in the collected key state variable data are standardized using the z-score standardization method, with the expression as follows:

[0023] Where X is the original numerical data in the key state variable data, μ is the mean of the training set, σ is the standard deviation, and X' is the standardized data.

[0024] This invention removes the influence of dimensions through standardization, making the data fall within the same numerical range, which facilitates subsequent unified processing and analysis.

[0025] Then, missing values ​​are filled in for the collected key state variable data.

[0026] Regarding missing value handling: For missing values ​​in categorical data (such as ABS status not being reported at a certain time), a forward fill (ffill) strategy is used, that is, the current missing value is replaced with the categorical value from the previous valid time step. The expression is:

[0027] in, Indicates the current value (missing). This represents the non-missing values ​​from the previous time step. This operation is performed before mapping to integer numbers to ensure the integrity of the input sequence.

[0028] Finally, outlier removal was performed on the collected key state variable data: Numerical boundaries for highway train data are set based on the physical operating characteristics of highway trains, directly filtering out data exceeding reasonable ranges. For example, the maximum design speed of highway trains typically does not exceed 120 km / h; therefore, if a speed observation at a certain moment is greater than 120 km / h or less than 0 km / h (negative values ​​have no physical meaning), it is considered an anomaly. These anomaly data are processed using boundary value replacement to preserve the continuity of the time series. For example, speed observations exceeding 120 km / h are replaced with 120 km / h, and speeds below 0 km / h are replaced with 0 km / h.

[0029] For data that does not exceed the numerical boundaries, data points that deviate from the normal distribution are identified using statistical methods and replaced with robust estimates. The specific steps are as follows: The first step is to select a time window length T (10 seconds) and calculate the corresponding number of data points N = T * f based on the sampling frequency f. For example, f = 50Hz and N = 500. For the standardized data sequence, calculate the moving average within each time window. With sliding standard deviation , as a local statistical feature.

[0030] The second step is to set the time window. Each data point within Calculate its relationship with the moving average. absolute value of deviation =∣ |; If >3 If the condition is met, the data point is considered an outlier. This rule is based on the statistical principle of "3... Principle: Under the assumption of normal distribution, 99.7% of the data lies within [μ]. 3 μ+3 Values ​​outside the range are considered low-probability events and will be further verified or corrected.

[0031] The third step is to use the median of the current time window for data points that are identified as outliers. Replace the original value. The median is more robust to extreme values ​​(unaffected by single-point outliers) and better represents the "normal" state within a window than the mean. For example, if data points within a certain time window... The median for one data point was 0.5, while the median for the other data point was 4.2 (exceeding 3). If ), then replace it with 0.5 to eliminate the interference of local anomalies on subsequent analysis.

[0032] The fourth step involves advancing the sliding window in fixed steps (usually one sampling point), repeating the statistical calculations, anomaly detection, and repair operations described above until the entire data sequence has been processed. This process ensures that anomaly detection and repair are continuous and dynamically adaptable over time, effectively addressing data drift issues under non-stationary conditions.

[0033] Step 3: Encode the preprocessed numerical and categorical data into a vector form that can be processed by the Transformer encoder-decoder temporal prediction model, and concatenate them along the feature dimension to obtain a fused feature vector. Construct a multidimensional state feature space.

[0034] The fusion process of this invention ensures that the Transformer encoder-decoder temporal prediction model can simultaneously capture the dynamic changes of continuous physical quantities and the semantic information of discrete state variables, providing high-quality input for subsequent Transformer temporal prediction models.

[0035] This invention employs learnable embedding coding to uniformly encode and fuse feature vectors for both numerical and categorical data. The processing flow is as follows: Step 31: Process numerical data; The values ​​of the preprocessed numerical data from step 2 at time step t are combined into a numerical feature vector:

[0036] in, Indicates vehicle speed; These represent longitudinal and lateral accelerations, respectively. Indicates the steering angle; The total dimension representing the numerical feature; It is represented as a vector; Represents the set of real numbers (this is the entire text); Indicates the number of categories in a categorical variable; This represents the embedding dimension for each category; Step 32, process categorical data; The first step is to create a separate vocabulary for each categorical data type. For example, the vocabulary for "vehicle type" would be {"head_truck": 0, "trailer_1": 1, "trailer_2": 2}; The second step is to convert categorical data (such as vehicle type "trailer_1", braking mode "emergency braking", ABS status "activated", etc.) into corresponding integer IDs (e.g., "trailer_1" → 1) through label encoding.

[0037] Step 33: Design an independent learnable embedding layer for each categorical data type. Input the integer ID into the learnable embedding layer and map it to a low-dimensional dense vector (i.e., the embedding vector), expressed as:

[0038] Where, e∈ This is the embedded low-dimensional dense vector. For the embedded dimension; For embedding matrix, ∈ Embedding matrix The parameters are automatically learned and optimized during the Transformer model training process; id is the integer ID after label encoding.

[0039] For example, if the ID corresponding to the ABS state "on" is 1, the embedding dimension =8, then e is an 8-dimensional vector representing the semantic features of the "on" state.

[0040] Furthermore, the learnable embedding layer is part of the Transformer encoder-decoder temporal prediction model, and its parameters are automatically optimized during model training.

[0041] Step 34: Linearly concatenate the low-dimensional dense vectors of all categorical data in a preset order (e.g., ABS status → braking mode → vehicle type) to form a categorical data embedding vector. ,in, The number of categorical data variables (e.g., ABS status, braking mode, vehicle type, a total of 3 variables, therefore) =3); For the embedded dimension.

[0042] Step 35, convert the numerical data vector and categorical data embedding vectors The features are concatenated along the feature dimension to form a fused feature vector. , =[ ], where d is the total feature dimension, d= + × The fused feature vector at all times All of them are located in the same d-dimensional real space, which is the multi-dimensional state feature space constructed in this invention. It is used to uniformly represent the complete operating state of the highway train at any time and to serve as the input basis for the subsequent Transformer time series prediction model.

[0043] The fused feature vector at each time step These form temporal samples in a multidimensional state feature space, which are then used as input for the subsequent Transformer model.

[0044] Furthermore, before concatenating along the feature dimensions, the numerical data is standardized using z-scores and then transformed into dimensionless standard real numbers (i.e., numerical data vectors). ); whereby the standard real number converted to dimensionless does not go through the embedding layer, but is directly used as a component of the feature vector, and is concatenated with the embedding vector of the categorical data in the feature dimension to form a complete fused feature vector; the fused feature vector is organized into a time sequence and used as the input of the Transformer encoder-decoder time sequence prediction model (i.e. the model described in step 4) for future state prediction and anomaly detection.

[0045] Step 4: Organize the fused feature vectors obtained in Step S3 into a time-series sequence, and select the current time and L consecutive time steps before it (i.e., the historical window length is L) to form the historical input window; based on the fused feature vectors from Step S3, and using the Transformer encoder-decoder model, process the historical fused feature sequence of the highway train. Model the vehicle and generate predictions of continuous physical quantities (such as vehicle speed, acceleration, and steering angle) for multiple future steps.

[0046] Specifically, the Transformer encoder-decoder model employs a 4-layer encoder + 4-layer decoder architecture, with 8 attention heads per layer and a model dimension d=32. This architecture is particularly suitable for long-term time-series prediction and dynamic trajectory generation tasks, effectively capturing the complex nonlinear mapping relationship between historical states and future evolution.

[0047] First, the input time window length is Historical fusion feature sequence :

[0048] in, This represents the state fusion feature vector at the start of the historical window, i.e., tracing back from the current time t. The time point corresponding to each step; Let d be the fused feature vector at time t, with dimension d = + × , Indicates the number of numerical state variables; Represents the number of categories of categorical state variables; This represents the vector dimension of each categorical variable after it has been mapped through the embedding layer. Secondly, standard sine-cosine positional encoding is used as the input time window length. Historical fusion feature sequence Each time step is assigned a unique location identifier, thus preserving the temporal sequence of historical states. The location identifier PE is added to the input historical fusion feature sequence. This forms the augmented input sequence with temporal position information that is ultimately input to the Transformer encoder. The expression is:

[0049] in, {0,1,..., } represents the relative position index of the time step within the input window. { } represents the dimension index in the fused feature vector; This indicates that at position pos, the first i Dimensional positional encoding value; and The wavelength scaling factor, through exponential decay, controls the different frequencies of sine / cosine waves corresponding to different dimensions, enabling the model to capture various time-scale dependencies ranging from short to long periods. Dimension d = + × , Indicates the number of numerical state variables; Represents the number of categories of categorical state variables; This represents the vector dimension of each categorical variable after it has been mapped through the embedding layer.

[0050] The final input is:

[0051] in, This represents the augmented input sequence with temporal location information, which serves as the final input to the Transformer model; then, The input is fed into the Transformer encoder-decoder architecture to obtain a sequence of numerical feature vectors for future time steps. The specific steps are as follows: Encoder enhances input sequences with temporal position information Perform multi-layer self-attention computation to extract state representations that include the global context; The decoder, conditioned on this state representation and combined with the positional encoding of the target sequence, progressively generates a predicted sequence composed of numerical feature vectors for the next horizon step through a cross-attention mechanism. .

[0052] For example: The model output is a predicted sequence composed of numerical feature vectors at the next horizontal time, which is used for motion state prediction and anomaly detection.

[0053] Step 5: Construct time series training samples and train the Transformer time series prediction model.

[0054] Specifically, in obtaining the fused feature vector Based on this, the present invention constructs the input-output sample pairs required for supervised learning using a sliding window method, and trains the Transformer time series prediction model.

[0055] Specifically, the time window length is set to... The prediction step size is horizontal, and the sliding step size is stride. For the preprocessed fused feature sequence... The samples were extracted in chronological order as follows:

[0056]

[0057] in, For the first The input of each sample includes historical fusion features; This indicates the first [number]th [item] in the input window. The historical state of each sample at the last time step; For the first input window The true future numerical features corresponding to each sample; A numerical feature vector representing the last future moment within the prediction range; All samples are combined to form the training set. M represents the total number of samples. This training set is input into the Transformer encoder-decoder model described in step 4, and end-to-end training is performed using mean squared error as the loss function until the model converges or meets the early stopping condition. After training, a time-series prediction model that can be used for real-time prediction is obtained.

[0058] Step 6: Deploy the trained time series prediction model on the vehicle-mounted edge computing unit or cloud monitoring platform to monitor and provide early warning of the operation status of highway trains in real time.

[0059] First, for the newly collected key state variable data at the current moment, repeat steps 2 (numerical feature standardization and variance filtering) and 3 (categorical feature embedding and fusion) to generate the current fused feature vector. The key state variable data is the core input data set, which includes both numerical and categorical data.

[0060] Secondly, maintain a length of The buffer queue is continuously updated with the latest historical fusion feature sequence to obtain the updated input sequence, which is then used as input to the time series prediction model.

[0061] Then, the updated input sequence is fed into the trained Transformer time-series prediction model to obtain the numerical feature predictions for the next horizon step. , This represents the model's prediction vector for the numerical state variables (such as vehicle speed, acceleration, etc.) at the horizontal step from the current time t (i.e., absolute time point t+horizon). The elements of the prediction vector are the future offset steps from the current time t. k The j-th numerical data Predicted value.

[0062] Finally, calculate the number of future offset steps at the current time t. k The actual observation value of the j-th numerical data The number of future offset steps corresponding to the current time t k The j-th numerical data absolute deviation of predicted values :

[0063] in, This indicates the total number of numerical state variables, such as the number of dimensions of continuous physical quantities like vehicle speed, longitudinal acceleration, lateral acceleration, vertical acceleration, steering angle, and yaw rate; t represents the current time (i.e., the end of the historical window), which is also the starting reference time point for prediction. k This represents the number of future offset steps at the current time t. k ∈{1,2,…,horizon}).

[0064] Pre-set dynamic or static thresholds for each key state variable data. (This can be determined through historical normal operating condition statistics or expert experience). If any key characteristic is satisfied:

[0065] If so, the system is determined to be in an abnormal state.

[0066] Once an anomaly is detected, an early warning mechanism is immediately triggered, and the following information is recorded: timestamp of the anomaly occurrence; vehicle ID and trailer configuration; specific feature name and value of the deviation exceeding the limit; corresponding historical window original data segment; and comparison curve of predicted value and actual value.

[0067] This invention employs the Transformer architecture to construct a time-series prediction model, solving the data fusion and feature extraction problems faced by traditional methods when processing multi-source heterogeneous data. Through real-time data comparison and anomaly monitoring mechanisms, it promptly detects and issues early warnings of abnormal situations, significantly improving the accuracy of monitoring and the reliability of early warnings.

[0068] Figure 2 The data generated during the operation of the highway train is presented in a table and diagram format. Figure 3 The time window is constructed as shown in the sliding window diagram. Each box represents a set of samples, and each set of samples corresponds to a predicted value through calculation. Figure 4 The method of the present invention demonstrates that, during data preprocessing, if there is a missing data point, it is filled with data from the next time step.

[0069] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for abnormal monitoring and early warning of the motion control status of highway trains, characterized in that, Includes the following steps: S1 deploys multiple types of sensors on key components of the lead car and each accompanying car of the highway train to collect key state variable data in real time, including speed, acceleration, braking pressure and steering angle. S2, preprocess the collected key state variable data to obtain processed numerical data and categorical data; S3, the processed numerical data obtained in step S2 is used as the numerical feature vector; the processed categorical data is input into the embedding layer after independently building a vocabulary and performing label encoding, and mapped into a low-dimensional dense vector to obtain the categorical data embedding vector; the numerical feature vector and the categorical data embedding vector are concatenated in the feature dimension to obtain the fused feature vector at each time step. S4. A time-series prediction model is constructed using a Transformer encoder-decoder architecture. The continuous fused feature vector sequence within the historical time window is used as the input of the time-series prediction model, and positional encoding information is introduced. The time-series prediction model is used to predict numerical data at multiple future times and outputs a sequence of predicted numerical feature values ​​for future times. S5. Using the sliding window method, time series training samples are constructed based on the fused feature vector sequence obtained in step S3, and the time series prediction model described in step S4 is trained with the goal of predicting future numerical data. S6, Use the trained time series prediction model to monitor highway trains in real time: Perform steps S2 to S3 on the real-time collected data to generate the current fused feature vector and update the historical sequence buffer queue; Input the updated historical sequence buffer queue into the time series prediction model to obtain the numerical feature prediction values ​​for future times; The actual observed value at the next moment is compared with the corresponding predicted value to calculate the absolute deviation; when the absolute deviation of any key state variable exceeds its preset threshold, it is judged as an abnormal state and an early warning is triggered.

2. The method according to claim 1, characterized in that, In step S2, the preprocessing includes: standardizing the numerical data, imputing missing values ​​in the categorical data, and identifying and replacing outliers in the numerical data based on statistical methods.

3. The method according to claim 1 or 2, characterized in that, In step S2, the identification and replacement of outliers in the numerical data based on statistical methods specifically includes: calculating the local mean and standard deviation of the data within a sliding time window; identifying data points whose absolute deviation from the local mean exceeds three times the standard deviation as outliers; and replacing the outliers with the median within the sliding time window.

4. The method according to claim 1, characterized in that, In step S4, the introduction of position encoding information specifically involves using sine-cosine position encoding to add the position information to the fused feature vector and then inputting it into the model.

5. The method according to claim 1, characterized in that, In step S5, the construction of time series training samples specifically involves: setting the time window length, prediction step size, and sliding step size, and extracting samples from the fused feature vector sequence in chronological order; the input of each sample is a historical fused feature vector sequence, and its corresponding true output is numerical data from multiple subsequent consecutive time points.

6. The method according to claim 1, characterized in that, In step S5, the loss function for training the time series prediction model is the mean squared error.

7. The method according to claim 1, characterized in that, In step S6, when the early warning mechanism is triggered, the timestamp of the anomaly, the vehicle identification involved, the specific state variables and their values ​​for the deviation exceeding the limit, the corresponding historical data segments, and the comparison information between the predicted value and the actual value are recorded simultaneously.

Images