Train wheel slip real-time detection method and device based on time sequence prediction model

Abstract

The invention belongs to the technical field of rail transit train operation safety detection, discloses a train wheel slip real-time detection method based on a time sequence prediction model, and aims to improve slip detection precision and guarantee driving safety. The method comprises the steps of collecting multi-dimensional data such as wheel rotating speed, displacement and acceleration in a vehicle-mounted driving log, dividing a data set after cleaning, sliding window segmentation and normalization processing, building a combined model fusing an LSTM neural network, an ARIMA statistical model and a self-adaptive weight synthesis module, completing training and parameter calibration through an early stop method, ADF inspection and an AIC criterion, and obtaining the vehicle-mounted driving log. According to the scheme, the prediction speed and the real speed are compared based on a dynamic threshold value to judge the slip state, errors are predicted and corrected through self-adaptive iteration, the device comprises a data extraction module, a preprocessing module, a model storage module, a speed prediction module, a slip judgment module and a result output module, hardware does not need to be newly added, the model is lightweight, end-side deployment can be achieved, high-precision and real-time early warning is achieved, and the method is suitable for large-scale popularization and application. The method adapts to a complex line environment and reduces operation cost.

Description

Technical Field

[0001] This invention belongs to the field of rail transit train operation safety detection technology, specifically relating to a real-time detection method and device for train wheel slippage based on a time series prediction model. Background Technology

[0002] The core requirement of rail transit operation is to ensure the smoothness and safety of train operation, which is directly related to the safety of passengers. As a key guarantee for train operation safety, the Automatic Train Protection (ATP) system's stable operation highly depends on accurate speed and positioning information. The ATP system not only makes decisions on subsequent train operations based on this information, but also verifies the compliance of the control structure. Therefore, speed measurement and positioning functions are the core foundation of safe rail transit operation.

[0003] Currently, urban rail lines generally use a combination of wheel speed sensors and transponders for speed measurement and positioning. This method calculates train speed and position by counting the number of wheel rotations, but it has significant technical limitations: wheel slippage can cause the measured speed and distance to be lower than the actual values, and this error accumulates with repeated slippage, leading to a decrease in positioning accuracy. In severe cases, this can cause safety accidents such as overshooting signals or rear-end collisions. Therefore, timely and accurate detection and alarming of wheel slippage are crucial means to improve train positioning accuracy and avoid safety risks.

[0004] Existing train slippage and idler detection technologies have explored multiple directions: some solutions rely on traditional measurement methods such as wheel and axle sensors, acceleration sensors, Doppler radar, or inertial navigation, and achieve detection by calculating speed difference, acceleration difference, slip rate, or speed difference ratio, combined with fixed threshold judgment; other solutions introduce intelligent algorithms such as neural networks and LSTM-AutoEncoder, and identify slippage state through feature extraction, model training, or difficulty in reconstructing abnormal data.

[0005] However, existing technologies still have significant shortcomings: on the one hand, solutions relying on traditional measurement methods are easily affected by wheel conditions, track environment and sensor errors, and a unified threshold cannot be adapted to the complex environment of different lines, resulting in insufficient detection accuracy and accumulation of positioning errors; on the other hand, although some solutions introduce intelligent algorithms, the application of deep learning, intelligent feature extraction and adaptive algorithms is still relatively limited, and a mature high-precision detection system has not yet been formed.

[0006] In summary, existing technologies cannot achieve high-precision detection of wheel slippage sections without increasing hardware costs. How to break through this bottleneck through technological innovation has become a key technical problem that urgently needs to be solved in the current rail transit field. Summary of the Invention

[0007] In view of the problems mentioned above in the background art, the purpose of this invention is to provide a method and device for real-time detection of train wheel slippage based on a time series prediction model.

[0008] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0009] A real-time detection method for train wheel slippage based on a time series prediction model includes the following steps:

[0010] S1. Collect multi-dimensional sensor data during train operation and construct a labeled dataset containing normal operation data and slippage event data;

[0011] S2. Construct an LSTM-ARIMA combined prediction model that integrates deep learning and statistical analysis. The model integrates an LSTM neural network module, an ARIMA statistical model module, and an adaptive weight synthesis module.

[0012] S3. Based on the labeled dataset, complete the training optimization of the LSTM model and the parameter calibration of the ARIMA model;

[0013] S4. Output the predicted speed of the train for the next operating cycle through the combined prediction model;

[0014] S5. Compare the predicted speed with the real-time collected wheel speed, determine whether the train is in a slipping state based on the deviation characteristics, and after entering the slipping state, use an adaptive iterative prediction method to estimate the real speed, and then determine whether to exit the slipping state according to preset conditions.

[0015] A real-time detection device for train wheel slippage based on a time series prediction model is used in accordance with the real-time detection method for train wheel slippage based on a time series prediction model as described in claims 1 to 9. The device is characterized by comprising a data extraction module, a data preprocessing module, a model storage module, a speed prediction module, a slippage judgment module, and a result output module.

[0016] The data extraction module is connected to the vehicle driving log system to extract sensor data such as speed, displacement, and acceleration related to skidding.

[0017] The data preprocessing module cleans, segments, and normalizes the raw data to ensure that the data quality is consistent with the model input.

[0018] The model storage module stores the trained LSTM-ARIMA combined model for subsequent prediction and judgment.

[0019] The speed prediction module uses LSTM and ARIMA models to predict train speed respectively, and fuses the prediction results through an adaptive weighting module.

[0020] The slippage detection module compares the predicted speed with the actual speed to determine whether slippage has occurred, and performs iterative prediction and exit judgment.

[0021] The result output module will send the detection results and prediction data back to the vehicle system or upload them to the cloud platform for subsequent analysis and alarm.

[0022] Further specifying, the multi-dimensional sensing data in step S1 includes wheel speed, train displacement, and vehicle acceleration, and the data is collected from the onboard operation logs of actual subway operation projects. After data collection, obvious abnormal records with missing or incorrect transponder data are first cleared, followed by time-series alignment processing. Then, a sliding window sampling strategy is used to segment the data sequence, archiving normal operation data and slippage event data separately. The window length of the sliding window is preset to 5-20 data points, and the time step is preset to 1-5 operating cycles. Both can be adaptively adjusted according to the characteristics of the train operation line and the data sampling frequency. The data originates from actual operation scenarios, and the preprocessing process ensures data quality. The sliding window parameters have both preset range and adaptive adjustment capabilities, adapting to different lines and sampling frequencies, and improving the rationality of data segmentation.

[0023] Further specifying, step S1 also includes normalizing the standardized dataset, using an adaptive extreme value calibration algorithm to update the data extreme value range in real time, and achieving dimensional uniformity of sensor data under different line and climate conditions. The normalization formula is:

[0024]

[0025] Where x is the data to be processed, x min x is the minimum value among all data. max x' represents the maximum value among all data points, and x' represents the normalized data. The data sources are aligned with actual operational scenarios, and the preprocessing process ensures data quality. Adaptive extreme value calibration and normalization processing achieve unified dimensions of sensor data across multiple scenarios, eliminate interference from differences in lines and climate on model input, and enhance model adaptability.

[0026] Further specifying, in step S1, the standardized labeled dataset is divided using stratified sampling: normal operation data is divided into training set, validation set, and test set at a ratio of 70%, 15%, and 15% respectively; all slippage event data are assigned to the test set, and neither the validation set nor the test set participates in the model training process, ensuring the objectivity and generalization ability of the evaluation results. Stratified sampling of the dataset, with slippage data separately assigned to the test set and the validation set not participating in training, ensures the objectivity and fairness of the model evaluation and comprehensively verifies the slippage scene detection performance.

[0027] Further specifying, the LSTM neural network module in step S2 consists of a moving average layer, at least two cascaded LSTM layers, and a linear output layer. The specific propagation process is as follows: the original input sequence is smoothed using a one-dimensional convolutional process in the time dimension through a fixed-size sliding window, restoring the smoothed sequence to its original time-order format as input to the LSTM network; the first LSTM layer receives data step-by-step, updating the "memory unit" and "hidden state" based on the current input and the state of the previous time step, and outputs the sequence to the subsequent LSTM layer; the subsequent LSTM layer receives the output of the preceding layer and extracts high-order time-series features, its output dimension can be the same as the preceding layer or adaptively adjusted at a ratio of 1:0.5 to 2; the final hidden state of the target layer is input to the linear mapping layer, converting it into a velocity prediction result in the target output space. The multi-layer structure of the LSTM module combined with moving average preprocessing can both reduce noise and smooth the data, and extract high-order time-series features; the adaptive adjustment of the output dimension adapts to different data complexities, improving the accuracy of velocity prediction.

[0028] Further specifying, in step S2, the adaptive weight synthesis module introduces an online learning mechanism based on prediction performance, and the weight update cycle is consistent with the train operation cycle; the weight ratio of the two models is dynamically adjusted according to the historical difference between the predicted values ​​and the actual values ​​of the LSTM model and the ARIMA model, with the model with the smaller prediction error being assigned a higher weight, and finally fused to obtain the prediction speed of the next operation cycle. The weights are dynamically adjusted based on the prediction error, and the advantages of the LSTM and ARIMA models are integrated in real time to make up for the limitations of a single model, making the prediction results more accurate and robust.

[0029] Further specifying, the training process of the LSTM model in step S3 is as follows: input the data of the first N periods to predict the data of the (N+1)th period; use the mean squared error loss function to evaluate the prediction bias, the formula is:

[0030]

[0031] Among them, y i These are the values ​​after normalizing the training data. For the model, use data y i The reconstructed value after reconstruction;

[0032] After each training round, the error value is predicted and calculated using the validation set, and the learning rate is dynamically adjusted based on the error. A gradient descent optimization strategy is used to stabilize the error between the training set and the validation set. Training stops when the predicted value on the validation set meets the conditions set by the early stopping method. Regularization constraints are introduced during training to suppress overfitting and ensure the model's generalization ability. By integrating MSE loss, dynamic learning rate, early stopping method, and regularization, the model achieves efficient convergence, effectively suppresses overfitting, and ensures the model's generalization ability and stability in different scenarios.

[0033] Further specifying, the parameter calibration process of the ARIMA model in step S3 includes:

[0034] The ADF stationarity test is used to determine whether the original time series data is stationary. If it is not stationary, it is successively differencing until the test is passed, and the difference order is determined. The relevant formula is:

[0035]

[0036] Where: Δy t y is the difference value at the current time point. t-1 The original value at the previous time point, t is the time trend term, and δ i These are the lag difference coefficients, used to control higher-order autocorrelation, ε t This is the error term;

[0037] The optimal parameter combination (p, d, q) is selected from the candidate space using a parametric grid search combined with the AIC criterion. The relevant formula is as follows:

[0038] AIC = 2k - 2ln(L)

[0039] The optimal model is subjected to residual testing to ensure that the model fit meets the preset requirements. The standardized parameter calibration process (ADF test + grid search + AIC criterion + residual test) accurately determines the optimal ARIMA parameters, thereby improving the model's ability to capture linear time series features and its fitting effect.

[0040] Further specifying, the slippage determination in step S5 adopts a dynamic threshold mechanism, and the threshold value is adaptively adjusted according to the train's operating line environment and current operating status. When the deviation between the predicted speed and the actual speed exceeds the dynamic threshold, the train is determined to have entered a slippage state. After entering the slippage state, the speed estimation value of the slippage area is corrected by an adaptive iterative prediction algorithm. When the deviation returns to the threshold range and meets the preset stability conditions, the train is determined to have exited the slippage state. The dynamic threshold adapts to complex operating environments, and the adaptive iterative prediction corrects the speed in the slippage area, achieving accurate identification of the slippage state and reliable exit judgment, reducing the risk of misjudgment and missed judgment.

[0041] The beneficial effects of using the present invention are as follows:

[0042] This invention is based on real sensor data collected in actual subway operation projects, including speed, displacement, acceleration, etc., and introduces deep learning methods and mathematical statistics to fuse and analyze multi-source data, thereby significantly improving the detection accuracy of wheel slippage.

[0043] This invention optimizes the accuracy of the train positioning system solely through algorithms without adding hardware, effectively saving costs. Compared to existing methods, it reduces reliance on expensive sensors and satellite navigation equipment.

[0044] This invention enables real-time detection and early warning, quickly identifies and alerts to slippage conditions, and further enhances the safety assurance capability of train operation.

[0045] The model constructed by this invention is highly lightweight, consumes few resources, and can be directly deployed on edge devices, demonstrating good application feasibility.

[0046] This invention can also provide a more accurate reference information source for multi-sensor and satellite navigation fusion positioning research, thereby improving the overall safety and reliability of train operation. Attached Figure Description

[0047] The present invention can be further illustrated by the non-limiting embodiments given in the accompanying drawings;

[0048] Figure 1 This is a flowchart of a real-time detection method for train wheel slippage based on a time series prediction model, as described in an embodiment of the present invention.

[0049] Figure 2 This is a framework diagram of the real-time detection model for train wheel slippage based on LSTM-ARIMA built in this invention.

[0050] Figure 3 This is the logarithmic loss curve obtained when training the LSTM prediction model in this invention.

[0051] Figure 4 This invention demonstrates the real-time recognition effect of partial slippage. Figure 1 ;

[0052] Figure 5 This invention demonstrates the real-time recognition effect of partial slippage. Figure 2 ;

[0053] Figure 6 This invention demonstrates the real-time recognition effect of partial slippage. Figure 3 ;

[0054] Figure 7 This is a structural diagram of a train slippage detection device based on deep learning in an embodiment of the present invention. Detailed Implementation

[0055] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0056] like Figures 1 to 7 As shown, the present invention provides a real-time detection method for train wheel slippage based on a time series prediction model, comprising the following steps:

[0057] S1. Collect multi-dimensional sensor data during train operation and construct a labeled dataset containing normal operation data and slippage event data;

[0058] S2. Build an LSTM-ARIMA combined prediction model that integrates deep learning and statistical analysis. The model integrates an LSTM neural network module, an ARIMA statistical model module, and an adaptive weight synthesis module.

[0059] S3. Based on the labeled dataset, complete the training optimization of the LSTM model and the parameter calibration of the ARIMA model;

[0060] S4. Output the predicted speed of the train for the next running cycle through the combined prediction model;

[0061] S5. Compare the predicted speed with the real-time collected wheel speed, determine whether the train is in a slipping state based on the deviation characteristics, and use an adaptive iterative prediction method to estimate the real speed after entering the slipping state, and then determine whether to exit the slipping state according to the preset conditions.

[0062] A real-time detection device for train wheel slippage based on a time series prediction model is used, which includes a data extraction module, a data preprocessing module, a model storage module, a speed prediction module, a slippage judgment module, and a result output module.

[0063] The data extraction module is connected to the vehicle's driving log system to extract sensor data such as speed, displacement, and acceleration related to slippage.

[0064] The data preprocessing module cleans, segments, and normalizes the raw data to ensure that the data quality is consistent with the model input.

[0065] The model storage module stores the trained LSTM-ARIMA combined model for subsequent prediction and judgment;

[0066] The speed prediction module uses LSTM and ARIMA models to predict train speed respectively, and then fuses the prediction results through an adaptive weighting module.

[0067] The slippage detection module compares the predicted speed with the actual speed to determine whether slippage has occurred, and performs iterative prediction and exit judgment.

[0068] The output module sends the detection results and prediction data back to the vehicle system or uploads them to the cloud platform for subsequent analysis and alarm purposes.

[0069] In this embodiment, the onboard driving logs of each project are first collected. Speed ​​sensor information related to wheel slippage is extracted from the driving logs of actual subway operation projects, covering key parameters such as speed, displacement, and acceleration. The raw data is preprocessed to remove obvious abnormal records such as missing transponder data or format errors to ensure data quality. A sliding window sampling strategy is used to segment the data sequence according to a preset window length (Window_size) and time step (Time_step), archiving normal operation data and slippage event data separately. Then, all data involved in training and detection are normalized using the following formula:

[0070]

[0071] Where x is the data to be processed, x min x is the minimum value among all data. max x' represents the maximum value among all data points, and x' represents the data after normalization.

[0072] Finally, the normalized normal data was divided into training, validation, and test sets at a ratio of 70%, 15%, and 15%, respectively. All slippage data was merged into the test set. Neither the validation nor the test set was used for model training, thus constructing a labeled dataset. Subsequently, an LSTM-ARIMA combined prediction model integrating deep learning and statistical analysis was built. This model includes an LSTM neural network module, an ARIMA statistical model module, and an adaptive weight synthesis module. The LSTM neural network module consists of a moving average layer, an LSTM layer, and a linear output layer. The ARIMA statistical model module is built based on time series statistical patterns. The adaptive weight synthesis module is used to dynamically fuse the prediction results of the two models. Model training and parameter calibration were completed based on the labeled dataset. The LSTM model was trained using the training set, and the data from the first N periods was used to predict the data from the (N+1)th period. The mean squared error loss function was used to evaluate the prediction bias, as shown in the formula:

[0073]

[0074] Among them, y i These are the values ​​after normalizing the training data. For the model, use data y i The reconstructed value after reconstruction;

[0075] After each training epoch, the validation set is used to predict and calculate the error value. The learning rate is dynamically adjusted, and a gradient descent optimization strategy is used to stabilize the error. Training stops when the early stopping condition is met. The validation set is not used for parameter updates to suppress overfitting. Simultaneously, the ARIMA model's parameters are calibrated. The ADF stationarity test is used to determine whether the original time series data is stationary. If it is not stationary, it is successively differencing until the test is passed, determining the differencing order. The relevant formula is:

[0076]

[0077] Where: Δy t y is the difference value at the current time point. t-1 The original value at the previous time point, t is the time trend term, and δ i These are the lag difference coefficients, used to control higher-order autocorrelation, ε t This is the error term;

[0078] The optimal parameter combination (p, d, q) is selected from the candidate space using a parametric grid search combined with the AIC criterion. The relevant formula is as follows:

[0079] AIC = 2k - 2ln(L);

[0080] The train speed for the next cycle is predicted using LSTM and ARIMA models based on the data from the previous N cycles. The adaptive weight synthesis module introduces an online learning mechanism based on prediction performance, dynamically adjusting the weights according to the difference between the predicted and actual values ​​from the two modules in the past, and then fusing them to obtain the final predicted speed data. The predicted speed data is compared with the real-time collected wheel speeds, and the difference between the two is calculated. When the difference reaches a set threshold, the train enters a slippage state. Once in the slippage state, an adaptive iterative prediction method is used to estimate the actual speed in the slippage area. The predicted data is then compared with the actual data, and the train exits the slippage state when the set conditions are met.

[0081] Specifically, the propagation process of the LSTM neural network module includes several key steps: the moving average layer smooths the original input sequence in the time dimension, performs one-dimensional convolution through a fixed sliding window, and restores the smoothed sequence to its original time order format as the input to the LSTM layer; the first LSTM layer receives the input sequence data step by step, updates the state information according to the current input and the "memory unit" and "hidden state" of the previous time step, and outputs the processed sequence to the subsequent LSTM layer; the subsequent LSTM layer receives the output sequence of the preceding layer, extracts time series features at a higher level using the same calculation method, and its output dimension can be the same as that of the preceding layer or adjusted as needed, finally obtaining the hidden state of the last time step of the target layer; this hidden state is input into the linear mapping layer, which converts the high-dimensional feature vector into the result of the target output space for subsequent loss calculation or inverse normalization processing. In determining the parameters of the ARIMA model, the stationarity test (ADF) is used to determine whether the original time series data is stationary. If a unit root exists, the series is non-stationary, requiring successive differencing until the differencing order d is determined. Then, a parameter grid search is used to traverse candidate combinations of autoregressive order p and moving average order q. The AIC criterion is used to seek a balance between goodness of fit and model complexity, selecting the optimal parameter combination (p, d, q) with the least information loss. The adaptive weight synthesis module's online learning mechanism continuously calculates the prediction errors of the two models within historical periods and dynamically adjusts the weight ratios, assigning higher weights to models with smaller prediction errors. Weighted fusion achieves complementary advantages. During slippage determination, the adaptive iterative prediction method uses the latest real data and historical prediction data as input to continuously update the speed estimate of the slippage area, avoiding error accumulation and ensuring the accuracy of state judgment.

[0082] A detection device for implementing the above methods is also provided. This device includes a data extraction module, a data preprocessing module, a model storage module, a speed prediction module, a slippage judgment module, and a result output module. The data extraction module connects to the vehicle's driving log system, extracting real-time sensor data related to slippage, such as speed, displacement, and acceleration, ensuring the integrity and real-time nature of data acquisition. The data preprocessing module receives the raw data and sequentially performs abnormal data cleaning, sliding window segmentation, and normalization processing, strictly adhering to a preset normalization formula to output standardized data that meets the model input requirements, ensuring data quality. The model storage module adopts a lightweight storage design, specifically storing the trained LSTM-ARIMA combined prediction model, including all parameters and structural information of the LSTM neural network module, ARIMA statistical model module, and adaptive weight synthesis module, adapting to the resource constraints of the edge device and providing model support for subsequent predictions. The speed prediction module operates in parallel. Two models are invoked to predict the speed of preprocessed time-series data. The results are then fused by an adaptive weight synthesis module according to the logic of "dynamic weighting based on prediction error" to output a high-precision predicted speed. The slippage detection module receives the predicted speed and real-time actual speed data, calculates the difference and compares it with a preset threshold to identify the slippage state. Once the slippage state is entered, an adaptive iterative prediction algorithm is triggered to correct the speed estimate. At the same time, the change in the difference is monitored, and the slippage state is exited when the exit condition is met. The result output module transmits the detection results, speed prediction data and iterative correction data bidirectionally back to the vehicle control system or uploads them to the cloud monitoring platform through the communication interface. This supports real-time safety warnings and control decisions of the vehicle system, and also provides data support for historical data backtracking and operation optimization analysis in the cloud.

[0083] Based on real sensor data collected from actual subway operation projects, this study innovatively integrates deep learning and mathematical statistics methods to analyze multi-source data, significantly improving the detection accuracy of wheel slippage and effectively solving the problem of insufficient detection accuracy in traditional methods. It innovatively employs algorithms to optimize the positioning system performance, eliminating the need for additional hardware and reducing reliance on expensive sensors and satellite navigation equipment. This effectively saves on rail transit operation and renovation costs while ensuring detection effectiveness. The constructed model is highly lightweight and consumes few resources, allowing direct deployment in end-side devices to achieve real-time detection and rapid early warning of slippage, responding promptly to slippage conditions to avoid safety risks and further enhancing the safety assurance capabilities of train operation. Through innovative mechanisms such as adaptive weight adjustment and dynamic thresholds, it adapts to different line environments and train operating states, exhibiting stronger robustness and avoiding the adaptation limitations of uniform thresholds. Furthermore, it provides a higher-precision reference information source for multi-sensor and satellite navigation fusion positioning research, comprehensively improving the safety and reliability of train operation, and possessing good application feasibility and promotional value.

[0084] A preferred method for real-time detection of train wheel slippage based on a time series prediction model is characterized by the following: In step S1, the multi-dimensional sensing data includes wheel speed, train displacement, and vehicle acceleration, and the data is collected from the onboard operation logs of actual subway operation projects; after data collection, obvious abnormal records with missing or incorrect transponder data are first cleared, and then time-series alignment is performed. Subsequently, a sliding window sampling strategy is used to segment the data sequence, and normal operation data and slippage event data are archived separately; the window length of the sliding window is preset to 5-20 data points, and the time step is preset to 1-5 operating cycles, and both can be adaptively adjusted according to the characteristics of the train operation line and the data sampling frequency.

[0085] In this implementation case, the multi-dimensional sensor data were all collected from the onboard train operation logs of actual subway operation projects, specifically including three core parameters: wheel speed, train displacement, and vehicle acceleration. After data collection, obvious abnormal records such as missing transponder data and incorrect formatting were cleared first to avoid invalid data interfering with subsequent processing. Then, time-series alignment was performed to ensure that sensor data of different dimensions were accurately matched on the time axis, eliminating the impact of time deviation. Next, a sliding window sampling strategy was used to segment the continuous time-series data, and normal operation data and slippage event data were archived and stored separately. The window length of the sliding window was preset to 5-20 data points, and the time step was preset to 1-5 operating cycles. In actual application, it can be adaptively adjusted according to the characteristics of the train line such as gradient and curvature, as well as the data sampling frequency (10-100Hz). For example, on line sections with large gradients and complex track environments, the window length can be adjusted to 5-10 data points to improve the segmentation targeting. On line sections with flat tracks and small data fluctuations, the window length can be extended to 15-20 data points to ensure data integrity.

[0086] To address the differences in route characteristics and data sampling frequencies, the adaptive adjustment of sliding window parameters can be achieved through a combination of preset rules and real-time data feedback: by continuously analyzing the fluctuation frequency of recent data and the probability of slippage events, the window length and time step are dynamically optimized. When the data sampling frequency increases, the time step is appropriately shortened to retain more refined temporal features; when the route has high smoothness and strong data stability, the window length is extended to improve data smoothness; when a region with frequent slippage events is detected, the window length and time step are reduced to accurately capture the characteristic changes at the moment of slippage. Simultaneously, multiple sets of route scenario parameter templates can be preset. For different scenarios such as plain routes, mountain routes, and densely populated urban routes, the appropriate initial sliding window parameters can be quickly called up, and then fine-tuned using real-time data to improve parameter adjustment efficiency.

[0087] The data sources directly align with actual operational scenarios. Abnormal data removal and time-series alignment ensure data quality from the outset, preventing invalid data from affecting the accuracy of model training and detection results. The sliding window not only has clearly defined preset parameter ranges to ensure operability but also possesses adaptive adjustment capabilities, flexibly adapting to differences in different line environments and data sampling frequencies. This allows data segmentation to better reflect actual train operation, providing more accurate and targeted data source support for subsequent model training and effectively improving the adaptability and reliability of the entire detection solution in complex scenarios.

[0088] The preferred step S1 also includes normalizing the standardized dataset and using an adaptive extreme value calibration algorithm to update the extreme value range of the data in real time, so as to achieve dimensional uniformity of sensor data under different line and climate conditions. The normalization formula is:

[0089]

[0090] Where x is the data to be processed, x min x is the minimum value among all data. max x is the maximum value among all data, and x' is the data after normalization.

[0091] In this implementation case, the adaptive extreme value calibration algorithm can be set with an update cycle consistent with the train operation cycle to ensure that the extreme value range can respond to data changes in real time. Simultaneously, an abnormal extreme value filtering mechanism is introduced. When a data point exceeds three times or more of the historical extreme value, it is temporarily excluded from the extreme value update range. Only after two to three similar data points appear consecutively is the extreme value range recalculated and updated, preventing a single extreme data point from interfering with the overall accuracy of extreme value judgment. Furthermore, extreme value calibration thresholds can be set separately for different types of sensor data (wheel speed, train displacement, vehicle acceleration). Based on the differences in the physical characteristics of various data types, the sensitivity of extreme value updates is optimized, further improving the accuracy of normalization processing.

[0092] By normalizing the data, the inconsistency in the dimensions of sensor data caused by differences in line slope, climate humidity, and operating speed is effectively eliminated, allowing the model to receive standardized input data and significantly improving the stability and convergence speed of model training. The introduction of the adaptive extreme value calibration algorithm avoids the shortcomings of the traditional fixed extreme value normalization method in terms of insufficient adaptability when the data distribution changes, ensuring that the normalization result always fits the real-time data distribution characteristics, further enhancing the model's generalization ability under different lines and climate conditions, and laying a solid data foundation for subsequent high-precision speed prediction and slippage detection.

[0093] In the preferred step S1, the standardized labeled dataset is divided using stratified sampling: normal operation data is divided into training set, validation set and test set at a ratio of 70%, 15% and 15% respectively; all slippage event data are assigned to the test set, and neither the validation set nor the test set participates in the model training process, ensuring the objectivity and generalization ability of the evaluation results.

[0094] In this implementation case, a stratified sampling method is used to scientifically divide the standardized labeled dataset. Normal operation data is strictly allocated to the model training set, validation set, and test set according to a ratio of 70%, 15%, and 15%, respectively. The training set is used for learning and optimizing the model's core parameters; the validation set is used to monitor overfitting during training and to provide a basis for dynamically adjusting the learning rate and triggering early stopping; and the test set is used to comprehensively evaluate the model's overall detection performance. All slippage event data, including minor slippage with a slip ratio of 0.05-0.1, moderate slippage with a slip ratio of 0.1-0.2, severe slippage with a slip ratio of 0.2-0.3, as well as short-term slippage lasting 0.2-1s and long-term slippage lasting 1-5s, are all included in the test set. This ensures that the test set can comprehensively cover all types of slippage situations that may occur in actual operation. The validation set and test set do not participate in parameter updates during the entire model training process; they are only used as a basis for evaluating model performance to avoid interference from training data with evaluation results.

[0095] When partitioning the dataset, it can be further subdivided and categorized according to the specific type of slippage event. Slippage data with different slip rates, durations, and occurrence scenarios are labeled and assigned to the test set separately, ensuring that the test set has comprehensive and representative coverage of slippage scenarios. Simultaneously, during stratified sampling, the probability of occurrence of normal operation and slippage states in actual subway operation data is statistically analyzed to ensure that the distribution characteristics of normal operation data in the training and validation sets are consistent with the distribution of actual operation data. For example, different data distributions during peak hours, off-peak hours, and nighttime hours are simulated, improving the relevance and practicality of model training. Furthermore, a small amount of real operation data not included in the partitioning can be reserved as an independent validation set for additional blind testing after model training, further validating the model's generalization ability.

[0096] The stratified sampling method ensures a reasonable distribution of data in the training, validation, and test sets, avoiding bias in model training due to uneven data partitioning. This allows the model to fully learn the feature differences between normal operation and slippage states within a balanced data distribution. Slippage data is separately included in the test set and covers a variety of typical scenarios, enabling comprehensive and objective verification of the model's detection performance under various slippage conditions. This avoids model bias caused by an excessively high or low proportion of slippage samples in the training data. The validation set does not participate in parameter updates, effectively suppressing overfitting and making the model evaluation results more realistic and valuable, providing a reliable performance basis for subsequent model optimization and practical applications.

[0097] In preferred step S2, the LSTM neural network module consists of a moving average layer, at least two cascaded LSTM layers, and a linear output layer. The specific propagation process is as follows: the original input sequence is smoothed by a one-dimensional convolutional process in the time dimension through a sliding window of fixed size, and the smoothed sequence is restored to the original time sequence format as the input of the LSTM network; the first LSTM layer receives data step by step in time, updates the "memory unit" and "hidden state" according to the current input and the state of the previous time step, and outputs the sequence to the subsequent LSTM layer; the subsequent LSTM layer receives the output of the preceding layer and extracts high-order time series features, and its output dimension can be the same as that of the preceding layer or adaptively adjusted according to a ratio of 1:0.5 to 2; the final hidden state of the target layer is input into the linear mapping layer and converted into the velocity prediction result of the target output space.

[0098] In this implementation, the LSTM neural network module consists of a moving average layer, at least two cascaded LSTM layers, and a linear output layer. These components work together to complete data preprocessing, feature extraction, and prediction output. The specific propagation process is as follows: The moving average layer performs one-dimensional convolutional smoothing on the original input sequence through a fixed-size sliding window, filtering out high-frequency noise and random interference in the data. The smoothed sequence is then restored to its original time-order format and used as input data for subsequent LSTM layers. The first LSTM layer receives the input sequence data step-by-step. Based on the current input data and the "memory unit" and "hidden state" of the previous time step, it updates the state information through gating mechanisms such as the input gate, forget gate, and output gate. Effective features are initially extracted and processed, and a new sequence is output to the subsequent LSTM layer. The subsequent LSTM layer receives the output sequence of the preceding layer and uses the same gating calculation method as the first layer to further extract the nonlinear features and long-term dependencies of the time series at a higher level. Its output dimension can be kept the same as that of the preceding layer, or it can be adaptively adjusted at a ratio of 1:0.5 to 2 to adapt to data features of different complexities. Finally, the final hidden state of the target layer (usually the last LSTM layer) at the last time step is input into the linear mapping layer to convert the high-dimensional feature vector into the velocity prediction result of the target output space. This result can be directly used for subsequent loss calculation, or restored to the original data dimension through inverse normalization.

[0099] The number of LSTM layers can be flexibly adjusted according to the complexity of the data features. For situations with large data fluctuations and complex temporal features (such as mountain routes or frequent start-stop scenarios), the number of LSTM layers can be increased to 3-4 layers to improve the ability to extract higher-order features. The number of LSTM units in each layer can be adapted to the dimension of the input features. For example, when the dimension of the input features is 10, the number of LSTM units can be set to 20-40 to ensure that the model has sufficient feature processing capabilities. The window size of the moving average layer can be dynamically adjusted according to the data sampling frequency. When the sampling frequency is 10-30Hz, the window size is set to 3-5 time steps; when the sampling frequency is 30-100Hz, the window size is set to 5-10 time steps to achieve a balance between noise reduction and feature preservation. In addition, a batch normalization layer can be added between the LSTM layer and the linear output layer to further improve the stability and convergence speed of model training.

[0100] The one-dimensional convolutional smoothing of the moving average layer effectively filters high-frequency noise in the original data, improves the stability of the input data, and provides a cleaner data foundation for subsequent feature extraction. At least two cascaded LSTM layers can extract shallow and high-order features of time-series data layer by layer, accurately capturing the long-term dependencies of wheel speed changes and adapting to the temporal characteristics of speed changes during train operation. The adaptive adjustment of the output dimension allows the module to flexibly adapt to datasets of different complexities, efficiently processing both simple plain line data and complex mountain line data. The linear output layer realizes the accurate mapping of high-dimensional features to speed prediction values. The structural design of the entire LSTM neural network module takes into account the core requirements of noise reduction, feature extraction, and prediction output, significantly improving the accuracy and stability of speed prediction.

[0101] In the preferred step S2, the adaptive weight synthesis module introduces an online learning mechanism based on prediction performance, and the weight update cycle is consistent with the train operation cycle. The weight ratio of the two models is dynamically adjusted according to the difference between the predicted values ​​and the actual values ​​of the LSTM model and the ARIMA model in history. The model with the smaller prediction error is assigned a higher weight, and finally the predicted speed of the next operation cycle is obtained by fusion.

[0102] In this implementation case, the adaptive weight synthesis module introduces an online learning mechanism based on prediction performance. The weight update cycle is consistent with the train operation cycle, ensuring that the weight adjustment can respond in real time to changes in the model's prediction performance. The module continuously records the predicted values ​​and corresponding actual speed values ​​of the LSTM neural network module and the ARIMA statistical model module in the historical operation cycle, calculates the prediction error of the two (including absolute error, relative error, and sum of squared errors), and dynamically adjusts the weight ratio of the two models based on the prediction error. The model with the smaller prediction error is assigned a higher weight ratio. Finally, the prediction results of the two models are merged by weighted summation to obtain the final predicted speed for the next operation cycle, ensuring that the prediction result can give full play to the advantages of the two models.

[0103] When adjusting weights, an error decay factor can be introduced, and historical prediction errors can be calculated using an exponential moving average. This allows the module to focus more on the prediction performance of recent models (e.g., the last 5-10 running cycles), improving the sensitivity and timeliness of weight adjustments and preventing outdated error data from affecting current weight allocation. Simultaneously, upper and lower limits for weight percentages can be set, limiting the weight percentage of a single model to between 0.2 and 0.8. This prevents a single model from having an excessively high or low weight percentage, ensuring that the two models can continue to work synergistically and avoiding a significant drop in prediction accuracy due to a model temporarily failing in a specific scenario. Furthermore, initial weight ratios can be preset based on different route scenarios (e.g., plains routes, mountain routes, routes in rain or snow), and then fine-tuned based on real-time prediction errors, further improving the efficiency and accuracy of weight adjustments.

[0104] The online learning mechanism based on prediction performance provides a clear quantitative basis for weight adjustment, avoiding the problem of insufficient adaptability of fixed-weight fusion methods in different scenarios. The weight update cycle is consistent with the train operation cycle, ensuring that the prediction results can respond to changes in model performance in real time and achieve dynamic complementarity of the advantages of the two models. When the LSTM neural network module predicts more accurately in complex nonlinear scenarios (such as frequent starts and stops, uneven tracks), its weight ratio is automatically increased. When the ARIMA statistical model module performs better in stable linear scenarios (such as high-speed uniform operation, plain lines), its weight ratio is automatically increased, which significantly improves the prediction accuracy and robustness of the combined model and effectively makes up for the adaptability limitations of a single model in different scenarios.

[0105] The training process of the LSTM model in the preferred step S3 is as follows: Predict the data for the (N+1)th period by inputting the data from the first N periods; evaluate the prediction bias using the mean squared error loss function, as shown in the formula:

[0106]

[0107] Among them, y i These are the values ​​after normalizing the training data. For the model, use data y i The reconstructed value after reconstruction;

[0108] After each training round, the error value is predicted and calculated using the validation set, and the learning rate is dynamically adjusted based on the error. A gradient descent optimization strategy is used to stabilize the error between the training set and the validation set. Training is stopped when the predicted value of the validation set meets the conditions set by the early stopping method. Regularization constraints are introduced during training to suppress overfitting and ensure the model's generalization ability.

[0109] In this implementation case, the value of N can be dynamically adjusted according to the scale and complexity of the training data. When the data volume is large and the features are complex (such as including multiple line scenarios and a large number of slippage samples), N is appropriately increased (30-50) to capture richer temporal features; when the data volume is small and the features are simple, N is decreased (10-20) to improve training efficiency. The learning rate adjustment adopts an adaptive decay strategy, with the decay coefficient set to 0.5-0.8, which avoids training stagnation caused by the learning rate decreasing too quickly, and can quickly converge to the optimal solution; regularization constraints. The strength of regularization can be dynamically adjusted based on the degree of overfitting on the validation set. For example, when the difference between the validation set error and the training set error exceeds 30%, the regularization strength is increased (e.g., the Dropout rate is increased from 0.1 to 0.3, and the L2 regularization penalty coefficient is increased from 1e-4 to 1e-3), and vice versa. This seeks a balance between suppressing overfitting and preserving the model's fitting ability. In addition, batch training can be used, and the batch size can be adjusted according to the computing resources of the edge device (e.g., 32, 64, 128) to further improve training efficiency.

[0110] The mean squared error loss function accurately measures prediction bias, providing a clear optimization direction for model training and ensuring that the model focuses on reducing overall prediction error. The combination of dynamic learning rate and gradient descent optimization strategy achieves efficient model convergence, avoiding the slow convergence or oscillation problems caused by fixed learning rate. The dual effect of early stopping and regularization constraints effectively suppresses overfitting, ensuring that the model performs well not only on the training set but also on unseen validation and test sets. The design of the entire training process takes into account both convergence speed and generalization ability, enabling the LSTM model to stably learn the temporal features in the data, providing solid support for high-precision speed prediction.

[0111] The parameter calibration process of the ARIMA model in preferred step S3 includes:

[0112] The ADF stationarity test is used to determine whether the original time series data is stationary. If it is not stationary, it is successively differencing until the test is passed, and the difference order is determined. The relevant formula is:

[0113]

[0114] Where: Δy t y is the difference value at the current time point. t-1 The original value at the previous time point, t is the time trend term, and δ i These are the lag difference coefficients, used to control higher-order autocorrelation, ε t This is the error term;

[0115] The optimal parameter combination (p, d, q) is selected from the candidate space using a parametric grid search combined with the AIC criterion. The relevant formula is as follows:

[0116] AIC = 2k - 2ln(L)

[0117] Perform residual testing on the optimal model to ensure that the model fit meets the preset requirements.

[0118] In this implementation case, the candidate parameter range can be narrowed down during parameter grid search based on the temporal characteristics of the data. For example, for data with small fluctuations and stable trends, the candidate range for p and q can be narrowed down to 1-5, reducing the number of searches and improving search efficiency. For data with large fluctuations and complex trends, the complete candidate range of 1-8 is retained to ensure that no optimal parameters are missed. During the ADF test, the form of the test model (whether it includes a trend term and a constant term) can be flexibly adjusted based on whether the time trend of the data is significant. For example, when the data has a clear linear trend, a test model that includes both a trend term and a constant term can be selected to improve the accuracy of the stationarity judgment. The residual test can introduce various statistical indicators, such as the Shapiro-Wilke test (for normality verification) and the Ljung-Box test (for verification of no autocorrelation). By comprehensively judging whether the residual sequence meets the model assumptions through multiple indicators, the reliability of the test results can be ensured. In addition, the optimal parameter combination can be fine-tuned (e.g., p±1, q±1) to re-verify the model performance and ensure that the selected parameter combination is globally optimal.

[0119] The standardized parameter calibration process provides a clear logic and basis for determining the parameters of the ARIMA model, avoiding parameter selection bias caused by subjective experience and improving the scientific nature and reproducibility of parameter determination. The ADF stationarity test ensures that the original data meets the stationarity requirements of the ARIMA model, laying the foundation for model fitting and avoiding prediction bias caused by non-stationary data. The parameter grid search combined with the AIC criterion comprehensively traverses candidate parameters while taking into account model complexity, avoiding overfitting or underfitting problems. The residual test further verifies the model's fitting effect, ensuring that the model can accurately capture the linear dependence of time series data, improving the prediction accuracy and stability of the ARIMA model, complementing the advantages of the LSTM neural network module, and providing support for high-precision prediction of the combined model.

[0120] In the preferred step S5, the slippage determination adopts a dynamic threshold mechanism. The threshold value is adaptively adjusted according to the train's operating line environment and current operating status. When the deviation between the predicted speed and the actual speed exceeds the dynamic threshold, the train is determined to have entered a slippage state. After entering the slippage state, the speed estimation value of the slippage area is corrected by an adaptive iterative prediction algorithm. When the deviation returns to the threshold range and meets the preset stability conditions, the train is determined to have exited the slippage state.

[0121] In this implementation case, a dynamic threshold mechanism is used for slippage detection. The dynamic threshold is adaptively adjusted based on the current track environment (such as gradient, curvature, and rail friction coefficient) and current operating status (such as speed and acceleration). By integrating multi-dimensional influencing factors, the threshold calculation logic is constructed to ensure that the threshold can meet the slippage detection requirements in different scenarios. The predicted speed output by the combined prediction model is compared with the real-time collected wheel speed, and the absolute deviation value between the two is calculated. When the deviation value exceeds the dynamic threshold and continues for at least two consecutive operating cycles, the train is determined to have entered a slippage state. After entering the slippage state, an adaptive iterative prediction algorithm is started. In each operating cycle, the latest collected real data and historical prediction data are used as input to update the speed prediction result and continuously correct the estimated value of the real speed in the slippage area to avoid error accumulation. When the deviation value returns to the range of the dynamic threshold and runs stably for more than three cycles, the train is determined to have exited the slippage state and the normal detection logic is restored.

[0122] The calculation of dynamic thresholds can incorporate machine learning models (such as decision trees and random forests). By training a threshold prediction model with historical data, the model can directly output the optimal threshold based on real-time input line environment and operating status parameters, improving the accuracy and response speed of threshold adjustment. The adaptive iterative prediction algorithm can set the iteration step size and dynamically adjust the weight of iteration updates based on the degree of slippage. The more severe the slippage (the higher the proportion of deviations exceeding the threshold), the smaller the iteration step size and the more refined the correction, ensuring accurate estimation of the true speed even in severe slippage scenarios. The exit condition can include judgment on the rate of speed change; when the rate of speed change stabilizes within ±0.5 m / s... 2 When the train is within the specified range, it can be further confirmed that it has exited the slipping state, avoiding misjudgment due to the return of instantaneous deviation. In addition, the threshold reference range for different scenarios can be preset. For example, when the rail surface friction coefficient is low in rainy or snowy weather, the dynamic threshold can be appropriately increased to avoid misjudgment of slight slippage. When the rail surface friction coefficient is high in dry weather, the dynamic threshold can be appropriately decreased to improve the detection sensitivity of slight slippage.

[0123] The dynamic threshold mechanism solves the problem that traditional fixed thresholds cannot adapt to different track environments and operating states, making slippage judgment more accurate and effectively reducing the risk of misjudgment and missed judgment. The adaptive iterative prediction algorithm can correct the speed estimation value of the slippage area in real time, and obtain a reliable speed reference even in the case of continuous slippage, avoiding the accumulation of positioning errors. Clear entry and exit conditions make the judgment of slippage state more rigorous, ensuring that the train can respond in time when slippage occurs, activate the warning and speed correction, and quickly resume normal detection after slippage ends, effectively improving the safety of train operation and providing more reliable status feedback for the Automatic Train Protection (ATP) system.

[0124] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A real-time detection method for train wheel slippage based on a time series prediction model, characterized in that, Includes the following steps: S1. Collect multi-dimensional sensor data during train operation and construct a labeled dataset containing normal operation data and slippage event data; S2. Construct an LSTM-ARIMA combined prediction model that integrates deep learning and statistical analysis. The model integrates an LSTM neural network module, an ARIMA statistical model module, and an adaptive weight synthesis module. S3. Based on the labeled dataset, complete the training optimization of the LSTM model and the parameter calibration of the ARIMA model; S4. Output the predicted speed of the train for the next operating cycle through the combined prediction model; S5. Compare the predicted speed with the real-time collected wheel speed, determine whether the train is in a slipping state based on the deviation characteristics, and estimate the real speed using an adaptive iterative prediction method after entering the slipping state, and then determine whether to exit the slipping state according to preset conditions.

2. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: The multi-dimensional sensing data in step S1 includes wheel speed, train displacement, and vehicle acceleration. The data is collected from the onboard operation logs of actual subway operation projects. After data collection, obvious abnormal records with missing or incorrect transponder data are first cleared, followed by time sequence alignment. Then, a sliding window sampling strategy is used to segment the data sequence, and normal operation data and slippage event data are archived separately. The window length of the sliding window is preset to 5-20 data points, and the time step is preset to 1-5 operating cycles. Both can be adaptively adjusted according to the characteristics of the train operation line and the data sampling frequency.

3. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: Step S1 further includes normalizing the standardized dataset, using an adaptive extreme value calibration algorithm to update the data extreme value range in real time, and achieving dimensional uniformity of sensor data under different line and climate conditions. The normalization formula is: Where x is the data to be processed, x min x is the minimum value among all data. max x is the maximum value among all data, and x' is the data after normalization.

4. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: In step S1, the standardized labeled dataset is divided using stratified sampling: normal operation data is divided into training set, validation set, and test set at a ratio of 70%, 15%, and 15%, respectively; all slippage event data are assigned to the test set, and neither the validation set nor the test set participates in the model training process, ensuring the objectivity and generalization ability of the evaluation results.

5. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: In step S2, the LSTM neural network module consists of a moving average layer, at least two cascaded LSTM layers, and a linear output layer. The specific propagation process is as follows: the original input sequence is smoothed by one-dimensional convolution in the time dimension through a sliding window of fixed size, and the smoothed sequence is restored to the original time sequence format as the input of the LSTM network; the first LSTM layer receives data step by step, updates the "memory unit" and "hidden state" according to the current input and the state of the previous time step, and outputs the sequence to the subsequent LSTM layer; the subsequent LSTM layer receives the output of the preceding layer and extracts high-order time series features, and its output dimension can be the same as that of the preceding layer or adaptively adjusted according to a ratio of 1:0.5 to 2; the final hidden state of the target layer is input into the linear mapping layer and converted into the velocity prediction result of the target output space.

6. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: In step S2, the adaptive weight synthesis module introduces an online learning mechanism based on prediction performance, and the weight update cycle is consistent with the train operation cycle. The weight ratio of the two models is dynamically adjusted according to the difference between the predicted values ​​and the actual values ​​of the LSTM model and the ARIMA model in history. The model with the smaller prediction error is assigned a higher weight, and finally the predicted speed of the next operation cycle is obtained by fusion.

7. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: The training process of the LSTM model in step S3 is as follows: Predict the data for the (N+1)th period by inputting the data from the first N periods; evaluate the prediction bias using the mean squared error loss function, the formula of which is: Among them, y i These are the values ​​after normalizing the training data. For the model, use data y i The reconstructed value after reconstruction; After each training round, the error value is predicted and calculated using the validation set, and the learning rate is dynamically adjusted based on the error. A gradient descent optimization strategy is used to stabilize the error between the training set and the validation set. Training is stopped when the predicted value of the validation set meets the conditions set by the early stopping method. Regularization constraints are introduced during training to suppress overfitting and ensure the model's generalization ability.

8. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: The parameter calibration process of the ARIMA model in step S3 includes: The ADF stationarity test is used to determine whether the original time series data is stationary. If it is not stationary, it is successively differencing until the test is passed, and the difference order is determined. The relevant formula is: Where: Δy t y is the difference value at the current time point. t-1 The original value at the previous time point, t is the time trend term, and δ i These are the lag difference coefficients, used to control higher-order autocorrelation, ε t This is the error term; The optimal parameter combination (p, d, q) is selected from the candidate space using a parametric grid search combined with the AIC criterion. The relevant formula is as follows: AIC = 2k - 2 ln(L) Perform residual testing on the optimal model to ensure that the model fit meets the preset requirements.

9. The real-time detection method for train wheel slippage based on a time series prediction model according to claim 1, characterized in that: In step S5, the slippage determination adopts a dynamic threshold mechanism. The threshold value is adaptively adjusted according to the train's operating line environment and current operating status. When the deviation between the predicted speed and the actual speed exceeds the dynamic threshold, the train is determined to have entered a slippage state. After entering the slippage state, the speed estimate of the slippage area is corrected by an adaptive iterative prediction algorithm; when the deviation returns to the threshold range and the preset stability condition is met, the train is determined to exit the slippage state.

10. A real-time detection device for train wheel slippage based on a time series prediction model, used in accordance with the real-time detection method for train wheel slippage based on a time series prediction model as described in claims 1 to 9, characterized in that: It includes a data extraction module, a data preprocessing module, a model storage module, a speed prediction module, a slippage detection module, and a result output module; The data extraction module is connected to the vehicle driving log system to extract sensor data such as speed, displacement, and acceleration related to skidding. The data preprocessing module cleans, segments, and normalizes the raw data to ensure that the data quality is consistent with the model input. The model storage module stores the trained LSTM-ARIMA combined model for subsequent prediction and judgment. The speed prediction module uses LSTM and ARIMA models to predict train speed respectively, and fuses the prediction results through an adaptive weighting module. The slippage detection module compares the predicted speed with the actual speed to determine whether slippage has occurred, and performs iterative prediction and exit judgment. The result output module will send the detection results and prediction data back to the vehicle system or upload them to the cloud platform for subsequent analysis and alarm.

Images