Train movement authorization anomaly detection method based on artificial intelligence

Abstract

The invention relates to a train movement authorization anomaly detection method based on artificial intelligence. The method comprises the steps that movement authorization data are acquired through a train control system; the movement authorization data are input into a train movement authorization anomaly detection model, a train movement authorization anomaly detection result is output, the train movement authorization anomaly detection model carries out preliminary screening on the movement authorization data, the screened data are decomposed into periodic components and seasonal components, and the periodic components and the seasonal components are used for detecting the movement authorization data; and performing autocorrelation feature analysis on the periodic components to obtain periodic features, and aggregating the screened data, the seasonal components and the periodic features. According to the method, the movement authorization data on the train are obtained, anomaly detection is carried out, the detection result is sent to the automatic train protection system (ATP), different responses are made for different anomalies, and the running safety of the train is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of train operation control technology, and in particular to an artificial intelligence-based method for detecting abnormal train movement authorization. Background Technology

[0002] As the core of modern urban public transportation, the safety and reliability of urban rail transit systems are directly related to the lifeline of urban operations. In Communication-Based Train Control (CBTC) systems, Movement Authorization (MA) is a crucial command determining whether a train can move and the distance it can travel. Abnormalities in MA often indicate potential risks of speeding, collisions, or even accidents. Current research on CBTC systems primarily focuses on accurately generating MA data. However, these efforts have not adequately considered scenarios where wireless communication is interfered with or tampered with, potentially leading to erroneous or missing MA data received by the train, thus causing operational risks.

[0003] The currently suitable model, MVTS-Transformer, is a Transformer encoder framework for learning multivariate time series representations. It was proposed by Zerveas et al. at KDD 2021. This model uses unsupervised pre-training and the Transformer's self-attention mechanism to extract deep representations of time series, thereby supporting various downstream tasks such as classification, regression, and missing value imputation.

[0004] While the MVTS-Transformer performs exceptionally well in multivariate time series modeling, it suffers from significant computational inefficiencies, making it particularly unsuitable for resource-constrained real-time scenarios. Its core bottleneck lies in the self-attention mechanism of the standard Transformer, which has a computational complexity of O(L²). For high-frequency sampling time series such as train operation authorizations (MA), the large data volume leads to a significant increase in memory consumption and inference latency. Furthermore, MVTS employs a complete Transformer encoder architecture with a large number of parameters and deep computational paths, making it difficult to meet the real-time requirements of onboard equipment even when running on GPUs. In addition, MVTS is not optimized for anomaly detection tasks, treating all input data equally and lacking a pre-filtering mechanism for explicit anomalies, further exacerbating the computational burden. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based method for detecting train movement authorization anomalies. By acquiring movement authorization data on the train, anomaly detection is performed, and the detection results are sent to the Automatic Train Protection (ATP) system to take different responses to different anomalies, thereby ensuring the safe operation of the train.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] An AI-based method for detecting train movement authorization anomalies includes:

[0008] Obtain movement authorization data through the train control system;

[0009] The movement authorization data is input into the train movement authorization anomaly detection model, and the train movement authorization anomaly detection result is output. The train movement authorization anomaly detection model performs preliminary screening of the movement authorization data, decomposes the screened data into periodic components and seasonal components, performs autocorrelation feature analysis on the periodic components to obtain periodic features, and aggregates the screened data, seasonal components, and periodic features.

[0010] Optionally, preliminary screening of the mobile authorization data includes:

[0011] One or more of the following methods were used for preliminary screening: rate of change, correlation coefficient, Gaussian filtering combination method, sliding window Z-score thresholding method, IQR outlier detection, and one-dimensional convolutional residual detection.

[0012] Optionally, a preliminary screening can be performed using a combination of rate of change, correlation coefficient, and Gaussian filtering, including:

[0013] The pre-detection head processes the movement authorization data using a sliding window, calculates the rate of change and correlation coefficient of the movement authorization data within the window, and performs Gaussian filtering on the movement authorization data within the window.

[0014] Based on the rate of change, correlation coefficient, and changes in the movement authorization data within the window before and after filtering, abnormal data is initially identified and filtered to obtain the filtered data.

[0015] Optionally, decomposing the filtered data into periodic and seasonal components includes: inputting the filtered data into an encoder and decomposing it using an adaptive sliding window, a classical filter, a lightweight 1D-CNN encoder-decoder, wavelet decomposition, or a learnable basis function expansion method to obtain periodic and seasonal components, wherein the adaptive sliding window is obtained by using windows of different lengths for low-round training.

[0016] Optionally, the periodic component is:

[0017] ;

[0018] The seasonal component is:

[0019] ;

[0020] In the formula, It is a periodic component. The data represents seasonal components, and MA represents the filtered data. For average pooling, for.

[0021] Optionally, autocorrelation feature analysis of periodic components may be performed, including:

[0022] The autocorrelation function of the periodic component is calculated using the Fast Fourier Transform and the Wiener-Hinchin theorem:

[0023] ;

[0024] ;

[0025] in, Representing time series and its Time delay similarity between lagged sequences It is the power spectral density in the frequency domain. Indicates Fourier transform, Indicates complex conjugation. This represents the inverse Fourier transform.

[0026] Optionally, the encoder includes an autocorrelation module, a sequence decomposition module, and a feedforward module. The autocorrelation module of the encoder performs autocorrelation feature processing on the filtered data, the sequence decomposition module decomposes the time series data into periodic and seasonal components, and the feedforward module performs nonlinear transformation and feature extraction on the learned features.

[0027] The beneficial effects of this invention are as follows:

[0028] 1. The Fast Fourier Transform (FFT) is used to calculate the autocorrelation function of the train operation authorization (MA) sequence, identify the dependencies between periodic subsequences, and realize anomaly detection with O(L log L) complexity. The computational efficiency is significantly better than the O(L²) dot product attention of the traditional Transformer.

[0029] 2. Introduce a rule-based explicit anomaly filtering module (such as MA drop, missing values, etc.) before the deep model. This can quickly filter out obvious abnormal samples without training, reduce subsequent computation, and improve real-time performance.

[0030] 3. Dynamically decompose the MA sequence into trend-period components, and automatically search for the optimal window length in the early stage of training to enhance the model's adaptability to data from different routes and time periods, thus avoiding manual parameter tuning.

[0031] 4. After the model detection is completed, reconstruction strategies (such as early braking, delayed braking, and interpolation completion) are designed for the four types of MA anomalies: abnormally high values, abnormally low values, noise interference, and missing data, to ensure the continuity and safety of the system response after detection. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart of an artificial intelligence-based train movement authorization anomaly detection method according to an embodiment of the present invention;

[0034] Figure 2 This is a structural diagram of the train movement authorization anomaly detection model according to an embodiment of the present invention;

[0035] Figure 3 This is a comparison diagram of the original mobility authorization data and the processed mobility authorization data in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] This embodiment proposes an artificial intelligence-based method for detecting train movement authorization anomalies, including:

[0039] Obtain movement authorization data through the train control system;

[0040] The train movement authorization data is input into the train movement authorization anomaly detection model, which outputs the train movement authorization anomaly detection results. The train movement authorization anomaly detection model performs preliminary screening of the movement authorization data, decomposes the movement authorization data into periodic components and seasonal components, performs autocorrelation feature analysis on the periodic components to obtain periodic features, and aggregates the screened data, seasonal components, and periodic features.

[0041] Specifically, such as Figure 1 As shown, the model first receives movement authorization data from the CBTC system, which is directly related to train operation safety. In actual operation, the train continuously receives movement authorization commands from the control center, and this data enters the detection system in time-series format. The pre-detection stage uses simple rules to perform initial screening of the data. Considering the clear acceleration limits and braking distance requirements for train operation, this stage checks for sudden changes or deviations from reasonable ranges in the data. For example, an abnormally large increase in the authorization distance, or other data that clearly does not conform to train operation patterns, will be directly marked as abnormal. Figure 3 The image shows a comparison between the original and processed movement authorization data. An autocorrelation algorithm is used to analyze the periodicity of the data. Trains run cyclically on fixed lines, and their movement authorization data often exhibits certain patterns. The algorithm calculates the similarity between current and historical data to identify potential anomalies. This method is more adaptable to the operational characteristics of different lines and time periods than directly setting fixed thresholds. The sequence decomposition module divides the data into trend and fluctuation components. The trend component reflects the overall operational status of trains throughout the day, such as the frequent departures during morning and evening rush hours; the fluctuation component reflects the specific operational status of individual trains. This decomposition method helps distinguish between systemic changes and genuine anomalies. Finally, the detection model sends the detection results to the ATP to achieve precise control of train operations.

[0042] The entire processing procedure fully considers the actual constraints of train operation. The computational complexity is controlled within the capabilities of the onboard equipment, ensuring that detection results are promptly fed back to the control system. When an anomaly is detected, the system takes different measures depending on the type of anomaly, such as requiring driver confirmation, reducing operating speed, or triggering emergency braking. This invention's proposed train movement authorization anomaly detection method identifies various anomalies in movement authorization data, addressing the problem of anomaly detection in movement authorization data within rail transit systems and providing crucial protection for train operation safety. This technology can reduce operational risks caused by movement authorization anomalies and improve the reliability and security of the CBTC system.

[0043] Further preliminary screening of mobile authorization data includes:

[0044] One or more of the following methods were used for preliminary screening: rate of change, correlation coefficient, Gaussian filtering combination method, sliding window Z-score thresholding method, IQR outlier detection, and one-dimensional convolutional residual detection.

[0045] Specifically, the sliding window Z-score thresholding method calculates the Z-score using the mean and standard deviation within the sliding window; values ​​exceeding the threshold are considered abnormal.

[0046] IQR (interquartile range) outlier detection: Based on the interquartile range within the window, upper and lower bounds are set, and anything exceeding the bounds is considered an anomaly.

[0047] One-dimensional convolutional residual detection: The residual is calculated using a learnable lightweight 1D-CNN. If the residual is greater than a threshold, an abnormal signal is triggered.

[0048] The above methods can be combined in series or parallel, or a trainable weighted fusion can be introduced to achieve faster filtering with greater robustness.

[0049] Furthermore, a preliminary screening is conducted using a combination of rate of change, correlation coefficient, and Gaussian filtering methods, including:

[0050] The pre-detection head uses a sliding window to process the movement authorization data, calculates the rate of change and correlation coefficient of the movement authorization data within the window, and performs Gaussian filtering on the movement authorization data within the window;

[0051] Based on the rate of change, correlation coefficient, and changes in the movement authorization data within the window before and after filtering, abnormal data is initially identified and filtered to obtain the filtered data.

[0052] Furthermore, decomposing the filtered data into periodic and seasonal components involves inputting the filtered data into an encoder and decomposing it using an adaptive sliding window, a classical filter, a lightweight 1D-CNN encoder-decoder, wavelet decomposition, or a learnable basis function expansion method to obtain periodic and seasonal components. The adaptive sliding window is obtained by using windows of different lengths for low-round training.

[0053] Specifically, classic filters such as the STL (Seasonal-Trend-Loess), Hodrick-Prescott, Baxter-King, or Christiano-Fitzgerald filters can all be decomposed offline or online.

[0054] Lightweight 1D-CNN encoder-decoder: Outputs two channels, trend and season, in one go through a convolution-transposed convolution structure, achieving end-to-end decomposition.

[0055] Wavelet decomposition: The discrete wavelet transform (DWT) is used to decompose the sequence into multi-scale coefficients, and then the trend and seasonal components are reconstructed.

[0056] Learnable basis function expansion: weighted summation using sine-cosine basis, polynomial basis, or neural network basis functions, with parameters obtained through end-to-end training.

[0057] Furthermore, the periodic components are:

[0058] ;

[0059] The seasonal components are:

[0060] ;

[0061] In the formula, It is a periodic component. The data represents seasonal components, and MA represents the filtered data. For average pooling, This is a fill operation.

[0062] Furthermore, autocorrelation characteristic analysis of periodic components includes:

[0063] The autocorrelation function of the periodic component is calculated using the Fast Fourier Transform and the Wiener-Hinchin theorem:

[0064] ;

[0065] ;

[0066] in, Representing time series and its Time delay similarity between lagged sequences It is the power spectral density in the frequency domain. Indicates Fourier transform, Indicates complex conjugation. This represents the inverse Fourier transform.

[0067] Specifically, this embodiment replaces the dot product self-attention mechanism in the standard transformer model with an autocorrelation mechanism based on the Fast Fourier Transform (FFT), reducing the complexity from O(L²) to O(L log L), significantly reducing computational overhead for long sequences. Secondly, the model introduces a lightweight rule-based pre-detection head, which can quickly filter out obvious abnormal samples before the deep learning module runs. Furthermore, TACformer employs a sequence decomposition structure to decouple trend and periodic components, further reducing redundant computation, and dynamically adjusts the computational granularity through an adaptive sliding window mechanism, improving efficiency and stability when data distribution changes.

[0068] Furthermore, the encoder includes: an autocorrelation module, a sequence decomposition module, and a feedforward module, wherein the autocorrelation module of the encoder performs autocorrelation feature processing on the filtered data, the sequence decomposition module decomposes the time series data into periodic components and seasonal components, and the feedforward module performs nonlinear transformation and feature extraction on the learned features.

[0069] In one embodiment, attention may be used to replace the autorelation module, for example:

[0070] Sparse Attention: Only calculates attention weights in fixed or dynamic sparse patterns, such as LogSparse, Strided, or Random Sparse patterns, reducing the complexity to O(L log L) or O(Ld).

[0071] Locality Sensitive Hash Attention (LSH Attention): Maps high-dimensional query-key vectors to low-dimensional buckets using hash functions, reducing the number of dot product calculations and maintaining O(L log L) complexity.

[0072] Low-rank attention: The Q and K matrices are projected onto a low-rank subspace before attention calculation is performed.

[0073] Kernel Attention: This method uses random feature mapping to transform dot product attention into a linear form, achieving an O(L) approximation.

[0074] All of the above attention methods can be directly embedded into a multi-head structure, replacing the original auto-correlation module, without changing the overall encoder-decoder framework.

[0075] The method of this embodiment will be further described below with reference to the accompanying drawings:

[0076] An AI-based method for detecting train movement authorization anomalies includes the following steps:

[0077] Collect the movement authorization data received by the train and input it into the train movement authorization anomaly detection model.

[0078] like Figure 2 As shown, the train movement authorization anomaly detection model comprises three main components: a pre-detection head, an encoder, and a decoder. The pre-detection head performs initial screening of MA data based on preset rules, quickly identifying some obvious anomalies, thereby reducing the computational burden of subsequent deep learning detection and improving the overall model's computational efficiency. The encoder encodes the input sequence and extracts latent variables, handles the seasonality of the input data, uses autocorrelation mechanisms to discover dependencies between different subsequences in the input sequence, and aggregates information at the subsequence level to generate a latent representation containing key feature information of the input sequence. The decoder receives the latent variables output by the encoder and the initial input data, performs decoding operations to generate the final prediction result, continuously accumulates and refines the prediction result using the trend component of the latent variables obtained from the encoder, and fully utilizes past seasonal information through autocorrelation mechanisms to further uncover hidden dependencies in the data, comprehensively considering the trend and seasonal factors of the time series to generate more reliable results. Specifically:

[0079] In the train movement authorization anomaly detection model, the pre-detection head performs pre-detection on MA data based on rules, filtering out obvious abnormal data.

[0080] Because train speed and acceleration are strictly controlled in actual rail transit systems, this rule-based method is efficient and accurate when dealing with abnormal data that deviates significantly from the normal range.

[0081] The pre-detection head uses a sliding window of length n to calculate the rate of change and correlation coefficient of the MA data within the window. The formula for calculating the correlation coefficient is:

[0082] ;

[0083] in For data labels within the sliding window, This represents the MA data within the sliding window. The pre-detection head applies a Gaussian filter to the data within the window, using the following formula:

[0084] ;

[0085] in Standard deviation, For the core radius, This is the offset relative to the center point.

[0086] By analyzing the rate of change, correlation coefficient, and changes in MA before and after filtering, the pre-detection head can identify anomalies in the MA data.

[0087] In this model, sequence decomposition breaks down the input MA data into periodic components. and seasonal ingredients The formula is as follows:

[0088] ;

[0089] ;

[0090] Periodic components represent the overall trend of the data, while seasonal components reflect the data fluctuation patterns.

[0091] Since the sequence decomposition module uses a moving average window, the length of the moving average window directly affects the decomposition effect. Therefore, an automatic mechanism for finding the optimal window length was designed. In the early stages of model training, the model uses multiple windows of different lengths for low-level training to find the best-performing length range, and then performs a fine-grained search within that range to find the optimal window length.

[0092] In this model, the autocorrelation module performs dependency discovery and information aggregation based on the periodicity of time series data. It measures the similarity between different time points by calculating the autocorrelation function of the time series. The mathematical definition of autocorrelation is as follows:

[0093] ;

[0094] in, This represents the value of MA at time point t. Indicates a time delay. Representing time series and its The time delay similarity between lagged sequences. In this embodiment, for input MA data of length L, the autocorrelation module uses Fast Fourier Transform (FFT) to calculate the autocorrelation function according to the Wiener-Hinchin theorem:

[0095] ;

[0096] ;

[0097] in, Indicates Fourier transform, Indicates complex conjugation. This represents the inverse Fourier transform. It is the power spectral density in the frequency domain.

[0098] The specific encoder includes an autocorrelation module, a sequence decomposition module, and a feedforward module. The sequence decomposition module progressively decomposes the time series data into periodic and seasonal components; the feedforward module performs further nonlinear transformations and feature extraction on the learned features.

[0099] The encoder's autocorrelation module processes the "move authorization" signal after its initial processing by the "pre-detection head." The encoder's first sequence decomposition module receives input from the output of the encoder's internal autocorrelation module, resulting in the first sub-signals. These signals are then passed to the encoder's feedforward layer for further feature enhancement. The encoder's second sequence decomposition module receives input from the decoder's internal feedforward layer and the output signals of the sequence decomposition, outputting the second decomposed sub-signals.

[0100] In the Encoder, the data is processed sequentially by the autocorrelation module to discover periodic dependencies and aggregate information; then, the time series is decomposed into periodic and seasonal components by the sequence decomposition module; finally, it undergoes nonlinear transformation by the Feed Forward module.

[0101] The processed data then enters the Decoder module, which includes an autocorrelation module, a sequence decomposition module, and a feedforward module to further refine features and gradually build prediction results. Finally, the model combines the outputs of the pre-detection head and the Decoder. The output includes the original time series and the model's detection results. Anomalies detected by the pre-detection head are directly judged as anomalies, while those not detected by the pre-detection head are further judged as anomalies based on the Decoder results, generating the final train movement authorization anomaly detection result.

[0102] The Decoder's first autocorrelation module takes periodic components as input and performs time-delay aggregation to discover periodic dependencies.

[0103] The first sequence decomposition module of the Decoder continues to decompose the previous module and extracts periodic components.

[0104] The second autocorrelation module of the Decoder performs time-delay aggregation on the results of the previous sequence decomposition and the encoder output to discover periodic dependencies.

[0105] The second sequence decomposition module of the Decoder continues to decompose the previous autocorrelation module to extract periodic components.

[0106] The Decoder's feedforward module receives the results of sequence decomposition, performs nonlinear transformations, and enhances the model's learning ability.

[0107] The second sequence decomposition module of the Decoder decomposes the previous module and extracts periodic components.

[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for detecting train movement authorization anomalies based on artificial intelligence, characterized in that, include: Obtain movement authorization data through the train control system; The movement authorization data is input into the train movement authorization anomaly detection model, and the train movement authorization anomaly detection result is output. The train movement authorization anomaly detection model performs preliminary screening of the movement authorization data, decomposes the screened data into periodic components and seasonal components, performs autocorrelation feature analysis on the periodic components to obtain periodic features, and aggregates the screened data, seasonal components, and periodic features.

2. The train movement authorization anomaly detection method based on artificial intelligence according to claim 1, characterized in that, The initial screening of the mobile authorization data includes: One or more of the following methods were used for preliminary screening: rate of change, correlation coefficient, Gaussian filtering combination method, sliding window Z-score thresholding method, IQR outlier detection, and one-dimensional convolutional residual detection.

3. The train movement authorization anomaly detection method based on artificial intelligence according to claim 2, characterized in that, The initial screening process, which combines the methods of rate of change, correlation coefficient, and Gaussian filtering, includes: The pre-detection head processes the movement authorization data using a sliding window, calculates the rate of change and correlation coefficient of the movement authorization data within the window, and performs Gaussian filtering on the movement authorization data within the window. Based on the rate of change, correlation coefficient, and changes in the movement authorization data within the window before and after filtering, abnormal data is initially identified and filtered to obtain the filtered data.

4. The train movement authorization anomaly detection method based on artificial intelligence according to claim 1, characterized in that, Decomposing the filtered data into periodic and seasonal components includes: inputting the filtered data into an encoder and decomposing it using an adaptive sliding window, a classical filter, a lightweight 1D-CNN encoder-decoder, wavelet decomposition, or a learnable basis function expansion method to obtain periodic and seasonal components. The adaptive sliding window is obtained by using windows of different lengths for low-round training.

5. The train movement authorization anomaly detection method based on artificial intelligence according to claim 4, characterized in that, The periodic component is: ； The seasonal component is: ； In the formula, It is a periodic component. The data represents seasonal components, and MA represents the filtered data. For average pooling, This is a fill operation.

6. The train movement authorization anomaly detection method based on artificial intelligence according to claim 4, characterized in that, Autocorrelation analysis of periodic components includes: The autocorrelation function of the periodic component is calculated using the Fast Fourier Transform and the Wiener-Hinchin theorem: ； ； in, Representing time series and its Time delay similarity between lagged sequences It is the power spectral density in the frequency domain. Indicates Fourier transform, Indicates complex conjugation. This represents the inverse Fourier transform.

7. The train movement authorization anomaly detection method based on artificial intelligence according to claim 4, characterized in that, The encoder includes an autocorrelation module, a sequence decomposition module, and a feedforward module. The autocorrelation module performs autocorrelation feature processing on the filtered data, the sequence decomposition module decomposes the time series data into periodic and seasonal components, and the feedforward module performs nonlinear transformation and feature extraction on the learned features.

Images