Heavy haul train minimum tracking distance prediction method, system, device and medium

Abstract

The application discloses a heavy-haul train minimum tracking distance prediction method, system, device and medium, and relates to the technical field of traffic control and intelligent prediction. The method comprises the following steps: acquiring running state data of a heavy-haul train, wherein the running state data at least comprises a rear vehicle speed and a front vehicle speed; performing feature engineering processing on the running state data to construct derived features, wherein the derived features comprise brake propagation delay features; constructing a neural network model and a fuzzy width learning model based on the running state data and the derived features; weighting and integrating the prediction results of the neural network model and the fuzzy width learning model by using a preset weight to obtain a trained integrated model; and performing real-time prediction on the minimum tracking distance of the heavy-haul train by using the trained integrated model. The application overcomes the defects of the prior art, such as the neglect of brake delay in a physical model and the insufficient performance of a single model, and finally realizes high-precision, high-robustness and real-time prediction of the minimum tracking distance of the heavy-haul train.

Description

Technical Field

[0001] This application relates to the field of traffic control and intelligent prediction technology, and in particular to a method, system, device and medium for predicting the minimum tracking distance of heavy-load trains. Background Technology

[0002] Heavy-haul railway transportation is characterized by long train formations, heavy weight, long braking distances, and complex operating conditions. As my country's heavy-haul railways develop towards higher density, higher speed, and larger capacity, safe interval control for heavy-haul train tracking has become a key technology for improving line transport capacity and ensuring traffic safety. The minimum tracking distance directly determines the operating efficiency and safety redundancy of heavy-haul trains, and its accurate prediction is of great significance for train operation control, automatic driving, and intelligent dispatching.

[0003] Traditional methods for calculating the following distance of heavy-haul trains often employ fixed safety distance methods, empirical formulas, or simplified dynamic models. For example, safety separation models based on physical formulas typically only consider the speeds of the preceding and following trains, braking deceleration, and driver reaction time, failing to adequately account for practical factors such as the significant propagation delay of air braking, the strong nonlinearity of braking characteristics, and variations in train formation length. This results in overly conservative predictions (wasting transport capacity) or insufficient safety margins (potential safety hazards).

[0004] In recent years, some studies have attempted to use single machine learning models (such as neural network models and support vector machines) for tracking distance prediction. However, the operating conditions of heavy-haul trains are complex and variable, and the braking system exhibits strong time-varying and uncertainties. Single models are prone to overfitting, weak generalization ability, and poor robustness, making it difficult to meet the prediction accuracy requirements of practical engineering. In addition, most existing data-driven methods ignore key physical characteristics such as the propagation delay of air brakes in heavy-haul trains and the differences in multi-vehicle cooperative braking, resulting in deviations between the models and real operating scenarios, and failing to provide a reliable basis for safe and efficient train tracking.

[0005] Therefore, how to construct a method for predicting the minimum tracking distance of heavy-haul trains that integrates the advantages of multiple models, considers the physical characteristics of braking propagation, and possesses strong robustness and high accuracy has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a method, system, device and medium for predicting the minimum tracking distance of heavy-haul trains, in order to solve the problems in existing methods for predicting the minimum tracking distance of heavy-haul trains. Traditional physical models do not fully consider key characteristics such as air brake propagation delay, resulting in conservative prediction results or insufficient safety margins. On the other hand, single machine learning models have weak generalization ability, poor robustness and insufficient prediction accuracy under complex working conditions.

[0007] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a method for predicting the minimum tracking distance of heavy-haul trains, including: Acquire the operating status data of heavy-haul trains, wherein the operating status data includes at least the speed of the following train and the speed of the preceding train; The operating status data is subjected to feature engineering to construct derived features, including braking propagation delay features; Based on the operational status data and the derived features, a neural network model and a fuzzy width learning model are constructed respectively. The prediction results of the neural network model and the fuzzy width learning model are weighted and integrated using preset weights to obtain a trained ensemble model; The minimum tracking distance of the heavy-haul train is predicted in real time using a trained ensemble model.

[0008] Secondly, this application provides a minimum tracking distance prediction system for heavy-haul trains, including: The data acquisition module is used to acquire the operating status data of the heavy-haul train, and the operating status data includes at least the speed of the following train and the speed of the preceding train. A derived feature construction module is used to perform feature engineering processing on the running status data and construct derived features, including braking propagation delay features. The model building module is used to build a neural network model and a fuzzy width learning model based on the running state data and the derived features, respectively. An integration module is used to perform weighted integration of the prediction results of the neural network model and the fuzzy width learning model using preset weights to obtain a trained integrated model. The real-time prediction module is used to predict the minimum tracking distance of the heavy-haul train in real time using a trained ensemble model.

[0009] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the heavy-load train minimum tracking distance prediction method described in any one of the above.

[0010] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the heavy-haul train minimum tracking distance prediction method described above.

[0011] Compared with the prior art, this application discloses the following technical effects: This application provides a method, system, device, and medium for predicting the minimum tracking distance of heavy-haul trains. By performing feature engineering on operational status data, derived features are constructed, including brake propagation delay features, thereby transforming the physical constraints of air brake propagation into data-driven input variables. Due to the long formation of heavy-haul trains and the limited propagation speed of air brakes, there is a significant time delay in the propagation of braking commands along the train. This delay directly affects the accuracy of the minimum tracking distance calculation. Existing methods do not include this delay as an explicit feature input, causing the model to be unable to effectively perceive the impact of braking asynchrony. This application solves the problem of prediction bias caused by traditional models ignoring brake propagation delay by introducing brake propagation delay features, achieving the fusion of physical characteristics and data models, and improving the model's adaptability to the braking dynamics of heavy-haul trains.

[0012] Furthermore, heavy-haul train operation data exhibits strong nonlinearity (such as sudden speed changes and braking level switching) and strong uncertainty (such as braking coefficient fluctuations and environmental interference). Single neural network models are prone to overfitting and have poor robustness; while single fuzzy width learning models can handle uncertainty, their nonlinear fitting ability is relatively limited. This application constructs a neural network model and a fuzzy width learning model based on operational status data and derived features, respectively, and uses preset weights to weighted integrate the prediction results of the neural network model and the fuzzy width learning model to obtain a trained ensemble model. By simultaneously constructing two heterogeneous models, utilizing the neural network model to capture nonlinear mappings and the fuzzy width learning model to suppress uncertainty interference, the advantages of nonlinear fitting and uncertainty handling are complemented, solving the problems of weak generalization ability and insufficient prediction accuracy of single models under complex operating conditions, and improving the accuracy and stability of the prediction results.

[0013] After offline training, the ensemble model's forward computation involves only simple matrix operations and activation function calculations, resulting in low computational latency and the ability to quickly output predicted values ​​as operational status data is updated. By utilizing the trained ensemble model to predict the minimum tracking distance of heavy-haul trains in real time, real-time and dynamic prediction of the minimum tracking distance of heavy-haul trains is achieved, providing timely and reliable data support for train operation control and safety interval adjustment.

[0014] In summary, this application overcomes the shortcomings of existing technologies, such as ignoring braking delay in physical models and insufficient performance of single models, by constructing braking propagation delay features, constructing neural network models and fuzzy width learning models, and using preset weighted integrated prediction. Ultimately, it achieves high-precision, high-robust, and real-time prediction of the minimum tracking distance of heavy-haul trains. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for predicting the minimum tracking distance of a heavy-haul train according to an embodiment of this application; Figure 2 A speed-time curve of the braking process of a heavy-haul train provided in an embodiment of this application; Figure 3 A comparison curve of the predicted and actual minimum tracking distance of each model provided in another embodiment of this application for heavy-haul trains; Figure 4 for Figure 3 A magnified view of a portion of the low to medium tracking distance range; Figure 5 for Figure 3 A magnified view of a portion of the medium-to-high tracking distance range; Figure 6 A bar chart comparing the prediction errors of various models provided in an embodiment of this application; Figure 7 An adaptive weight change curve of the integrated model provided in another embodiment of this application.

[0017] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The minimum tracking distance prediction method for heavy-haul trains provided in this application can be applied to heavy-haul trains equipped with speed sensors, braking system status monitoring modules, and wireless communication units. Train operation status data (including following train speed, preceding train speed, emergency braking distance, etc.) are collected via the onboard data bus and transmitted to the onboard intelligent terminal or ground dispatch center server. The onboard intelligent terminal or ground server performs feature engineering processing on the collected data, constructs derived features including braking propagation delay features, and builds neural network models and fuzzy width learning models respectively. The minimum tracking distance prediction value is obtained through weighted integration. The prediction results can be used by the train operation control system for safety interval adjustment, braking warning, and optimized scheduling. In some embodiments, this method can also be completed independently by the onboard terminal or ground server without uploading data to the cloud.

[0021] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting the minimum tracking distance of heavy-haul trains is provided. This method is executed by computer equipment, specifically by an onboard intelligent terminal or a ground dispatch center server alone, or by the onboard terminal and the ground dispatch center server working together. In this embodiment, the method is described using an onboard intelligent terminal as an example, and includes the following steps: Acquire the operating status data of heavy-haul trains, wherein the operating status data includes at least the speed of the following train and the speed of the preceding train; The operating status data is subjected to feature engineering to construct derived features, including braking propagation delay features; Based on the operational status data and the derived features, a neural network model and a fuzzy width learning model are constructed respectively. The prediction results of the neural network model and the fuzzy width learning model are weighted and integrated using preset weights to obtain a trained ensemble model; The minimum tracking distance of the heavy-haul train is predicted in real time using a trained ensemble model.

[0022] The proposed method for predicting the minimum tracking distance of heavy-haul trains addresses the issue that, due to the long formation of heavy-haul trains and the limited propagation speed of air brakes, there is a significant time delay in the propagation of braking commands along the train. This delay directly affects the accuracy of the minimum tracking distance calculation. Existing methods do not include this delay as an explicit feature input, resulting in the model's inability to effectively perceive the impact of braking asynchrony. This application addresses this by performing feature engineering on the operating status data to construct derived features that include the braking propagation delay. This transforms the physical constraints of air brake propagation into data-driven input variables, thereby solving the problem of prediction bias caused by traditional models ignoring the braking propagation delay. It achieves the fusion of physical characteristics and data models, improving the model's adaptability to the braking dynamics of heavy-haul trains. Furthermore, heavy-haul train operation data exhibits strong nonlinearity and uncertainty. Single neural network models are prone to overfitting and have poor robustness, while single fuzzy width learning models, although capable of handling uncertainty, have relatively limited nonlinear fitting capabilities. This application constructs a neural network model and a fuzzy width learning model based on processed feature data, and integrates their prediction results using preset weights, with the weight of the neural network model being less than that of the fuzzy width learning model. This approach utilizes the neural network model to capture nonlinear mappings, while the fuzzy width learning model suppresses uncertainty interference, and a higher weight is given to enhance the latter's contribution. This solves the problem of weak generalization ability and insufficient prediction accuracy of single models under complex conditions, achieving a complementary advantage in nonlinear fitting and uncertainty handling, significantly improving prediction accuracy and stability. Finally, the trained ensemble model is used to predict the minimum tracking distance of heavy-haul trains in real time. The ensemble model, trained offline, has low forward computation latency and can quickly output predicted values ​​as the operation status data updates, achieving high-precision and robust dynamic safety interval prediction, providing a reliable basis for the safe and efficient operation of heavy-haul trains.

[0023] Furthermore, this application can dynamically adjust the ensemble weights through a sliding window mechanism and update the fuzzy width learning model online using an incremental learning approach, thereby adapting to the time-varying nature of heavy-haul train operating conditions, further improving prediction robustness and reducing online computational overhead.

[0024] In another exemplary embodiment, a method for predicting the minimum tracking distance of a heavy-haul train is provided. This method is executed by an onboard intelligent terminal and specifically includes the following steps 101 to 105, wherein: Step 101: Obtain the operating status data of heavy-haul trains.

[0025] In a specific example, based on the actual operating data of a heavy-haul railway line, the sampling period is exemplarily set to 300ms. The collected heavy-haul train operating status data includes at least the speeds of the following and preceding trains. Further, it may also include emergency braking distance, emergency deceleration, normal deceleration, braking coefficient, safety margin, and difference in braking distance. The collected operating status data is used as training samples, totaling... Each training sample contains a set of input features and their corresponding ground truth minimum tracking distance. It should be noted that the sampling period and number of samples mentioned above are merely examples and can be adjusted according to specific network conditions and data volume in actual applications.

[0026] Step 102: Perform feature engineering on the running status data to construct derived features.

[0027] The original data (i.e., the operating status data of heavy-haul trains) is subjected to feature engineering to construct derived features. The derived features include relative distance features, speed ratio features, braking safety factor, braking time features, relative speed features, deceleration ratio features, braking efficiency coefficient, and braking propagation delay features unique to this application.

[0028] In this embodiment, the braking propagation delay characteristic is estimated based on the train formation length and the air brake propagation speed, specifically as follows:

[0029] in, Braking propagation delay time (unit: seconds). The train formation length (in meters) is given by train i. This refers to the propagation speed of air braking. For example, the propagation speed of air braking... The actual value range is usually 180 to 220 m / s, for example, 200 m / s can be taken.

[0030] To assist in constructing braking propagation delay features, the following physical model that considers braking propagation delay can be introduced (this model is not directly used as the prediction output, but only for constructing derived features related to braking safety):

[0031] in, The dynamic safety interval between trains, i.e., the minimum tracking distance, For safety margin, express The difference in kinetic energy between the following vehicle and the preceding vehicle at a given moment, when the speed of the following vehicle... greater than the speed of the vehicle in front When the time is right, the kinetic energy difference term takes a positive value; otherwise, it takes a zero value. Minimum safe deceleration for the following vehicle This indicates the braking delay distance of the vehicle behind.

[0032] It should be noted that the above parameters (such as safety margin) Minimum safe deceleration (etc.) can be set according to the specific line and train type, and this example does not limit it. In this application, the minimum tracking distance refers to the minimum dynamic safety interval that the following train must maintain with the preceding train, ensuring that it can brake safely without a rear-end collision even in the worst-case scenario where the preceding train brakes suddenly and the following train's reaction is delayed. This distance is the core basis for the train operation control system to adjust the safety interval.

[0033] In addition to the braking propagation delay feature mentioned above, this embodiment also constructs a variety of derived features to enhance the model's ability to represent complex operating conditions. The following uses "braking safety factor" and "braking efficiency factor" as examples to illustrate their construction logic.

[0034] (1) Braking safety factor The braking safety factor is used to quantify the safety redundancy of braking capacity under current operating conditions. It is defined as the ratio of the maximum available braking deceleration of the following vehicle to the actual required deceleration. The calculation formula is as follows:

[0035] in, The maximum available braking deceleration of the following vehicle (which can be estimated in real time based on the braking system status, adhesion coefficient, etc.). The minimum deceleration required to avoid a rear-end collision is estimated based on the speed difference between the vehicles in front and behind and the current distance:

[0036] In the above formula, and These are the speeds of the following vehicle and the speed of the vehicle in front, respectively. This represents the current actual tracking distance. This is a preset safety margin. When When, it indicates that the current braking capacity is sufficient; when When the braking force is low, it indicates insufficient braking capacity and a higher risk of rear-end collision. This feature effectively reflects the dynamic changes in braking safety margin.

[0037] (2) Braking efficiency coefficient Braking efficiency coefficient This term describes the degree of braking asynchrony caused by the train's formation length in heavy-haul trains. It is defined as the ratio of the time difference between the time it takes for the brake cylinder pressure of the head locomotive (the foremost locomotive) and the tail car (the rearmost car) to reach the target value to the air brake propagation time. The calculation formula is as follows:

[0038] in, The time it takes for the brake cylinder pressure of the lead locomotive to reach the target value. This refers to the time it takes for the brake cylinder pressure of the tail locomotive to reach the target value. For ideal synchronous braking, Due to the propagation delay of air braking, typically Furthermore, the larger this value, the more severe the braking asynchrony, the greater the longitudinal impulse of the train, and the correspondingly increased the required safe tracking distance. This characteristic can be obtained directly from train operation monitoring devices or braking system monitoring data.

[0039] Other derived features, such as relative distance features ( ), speed ratio characteristics ( ), braking time characteristics (the ratio of emergency braking distance to current speed), relative speed characteristics ( Characteristics such as deceleration ratio can be directly calculated from the original operating state data through similar basic mathematical transformations (difference, ratio, integral, etc.). Those skilled in the art can make conventional selections based on the actual data characteristics, which will not be elaborated here.

[0040] For each of the above original and derived features Normalization is performed to eliminate the influence of different feature dimensions. For example, Z-score normalization can be used.

[0041] in, and Features The mean and standard deviation.

[0042] like Figure 2 The figure shows the speed-time curves of a heavy-haul train braking while following another train. The horizontal axis represents time t (in seconds, 0–10 s), and the vertical axis represents speed v (in m / s, 0–22 m / s, approximately 0–79.2 km / h). Three speed curves are shown in the figure: the thin dashed line represents the speed curve of the preceding train. The vehicle in front decelerates smoothly from its initial speed (approximately 15 m / s) until it approaches zero, serving as the target for the following vehicle. The thick dashed line represents the ideal speed curve of the following vehicle without braking propagation delay. The following vehicle decelerates synchronously with the vehicle in front from moment zero, with a larger deceleration and a faster speed decrease, ultimately coming to a stop before the vehicle in front, corresponding to the ideal safe following state. The thick solid line represents the actual speed curve of the following vehicle with braking propagation delay. The following vehicle maintains a constant initial speed (20 m / s) during the delay phase, only beginning to decelerate at maximum speed after the delay ends, resulting in a speed curve exhibiting a "plateau + steep descent" pattern. The gray shaded area in the figure represents the braking propagation delay interval, the length of which is the braking propagation delay time. (Also known as braking response time). During this phase, the following vehicle does not apply brakes and continues at a speed of... The vehicle travels at a constant speed; after the delay ends, the following vehicle begins to brake at full speed. According to the physical model in step 102 (formula (2)), the additional distance traveled by the following vehicle during the braking delay is: This is called the braking delay distance, which is used to calculate the dynamic safety separation between trains. One of the core aspects. The longer the delay time and the higher the initial speed of the following vehicle, the greater the braking delay distance and the higher the risk of a rear-end collision.

[0043] Figure 2 The impact of quantified braking propagation delay on the safe distance is given by formula (2). The project provides intuitive physical support; it guides the design of braking systems for Automatic Train Operation (ATO), Automatic Train Protection (ATP), and vehicle following models, by shortening the time required for braking. Optimize deceleration to reduce safety distance requirements and improve line traffic efficiency; verify the safety of the car-following control strategy, compare braking curves with and without delay, and assess the impact of delay on rear-end collision risk.

[0044] Step 103: Based on the processed original data and derived features, construct a neural network model and a fuzzy width learning model respectively.

[0045] Step A1) Construction of the neural network model In one example, a five-layer feedforward neural network model is constructed: the input layer receives a feature vector after feature engineering, which includes original features and derived features; the number of neurons in the first hidden layer is 20, the number of neurons in the second hidden layer is 15, the number of neurons in the third hidden layer is 10, and the number of neurons in the output layer is 1 (corresponding to the predicted value of the minimum tracking distance). The activation function of the hidden layers is tansig, and the activation function of the output layer is purelin.

[0046] Furthermore, the aforementioned five-layer feedforward neural network model includes an input layer, a hidden layer, and an output layer.

[0047] For the input layer, it receives an input feature vector composed of all normalized features. in, Each of these represents a feature, and n is the dimension of the input feature. These features include the original data (rear vehicle speed, front vehicle speed, emergency braking distance, emergency deceleration, common deceleration, braking coefficient, safety margin, difference in braking distance, etc.) and the derived features constructed in step 102 (relative distance feature, speed ratio feature, braking safety coefficient, braking time feature, relative speed feature, deceleration ratio feature, braking efficiency coefficient, braking propagation delay feature).

[0048] For the first hidden layer, the computational relationship between the input layer and the first hidden layer is as follows:

[0049] in, For the first hidden layer The output of each neuron; For the first Input features To the first hidden layer The connection weights of each neuron ; For the first hidden layer The bias of each neuron. The tansig activation function is defined as:

[0050] in, The weighted sum of neurons before activation; activation function The output range is (-1, 1), which has zero-center characteristics and helps to accelerate network convergence.

[0051] For the second hidden layer, the calculation relationship between the first hidden layer and the second hidden layer is as follows:

[0052] in, For the second hidden layer The output of each neuron; The connection weights from the first hidden layer to the second hidden layer; For the second hidden layer Bias of each neuron.

[0053] For the third hidden layer, the calculation relationship between the second and third hidden layers is as follows:

[0054] in, For the third hidden layer The output of each neuron; The connection weights from the second hidden layer to the third hidden layer; For the third hidden layer Bias of each neuron.

[0055] For the output layer, the computational relationship between the third hidden layer and the output layer is as follows:

[0056] in, For output; The connection weights from the third hidden layer to the output layer; This is the bias for the output layer.

[0057] The output layer has 1 neuron, corresponding to the predicted minimum tracking distance. The purelin activation function is defined as a linear function:

[0058] Step A2) Training the neural network model The training algorithm for the above five-layer feedforward neural network model adopts the Levenberg-Marquardt (LM) algorithm, and the loss function is the mean squared error (MSE). The training termination condition is exemplarily set as a maximum of 1000 iterations or an error less than 1×10⁻⁶. -5 The values ​​above are for illustrative purposes only and can be adjusted according to the data size in actual applications.

[0059] Specifically, the LM algorithm combines the advantages of gradient descent and Gauss-Newton's method, featuring fast convergence and good stability. The parameter update formula for the LM algorithm is:

[0060] in, This is the weight adjustment amount; is the Jacobian matrix, containing the partial derivatives of the error function with respect to each network parameter; It is the identity matrix; This is the output error vector; It is the damping factor; This is a transpose operation.

[0061] The LM algorithm switches between gradient descent and Gauss-Newton methods. When the loss function value (e.g., mean squared error) decreases compared to the previous iteration, the damping factor μ is divided by the decay factor (0.1 for example), causing the algorithm to tend towards the Gauss-Newton method and speeding up convergence. When the loss function value increases compared to the previous iteration, the damping factor μ is multiplied by the growth factor (10 for example), causing the algorithm to tend towards gradient descent to ensure stability.

[0062] This embodiment uses mean squared error (MSE) as the loss function during the training process:

[0063] in, For the first The true minimum tracking distance for each sample. These are the predicted values ​​from the neural network model. , This represents the number of training samples.

[0064] The output of each layer is calculated using the formula described above to obtain the predicted value of the neural network model. The error between the predicted value of the neural network model and the true value of the minimum tracking distance is calculated. The gradient of the error with respect to each network parameter is calculated using the LM algorithm. The network weights and biases are adjusted according to the parameter update formula (10). When the preset maximum number of iterations is reached (1000 times in this embodiment) or the error is less than the set threshold (1×10 in this embodiment), the error is adjusted. -5 Stop training when ( ).

[0065] Preferably, after the neural network model has been trained, three metrics are used to evaluate the model performance: mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). 2 These metrics are also used to evaluate the performance of subsequent fuzzy width learning models and ensemble models. The specific calculation methods for these metrics are as follows:

[0066] in, for The average of the true values ​​of each training sample.

[0067] Step A3) Construction of the fuzzy width learning model In this embodiment, the input features of the fuzzy width learning model are the same as those of the neural network model, which are the original features after normalization and all the aforementioned derived features.

[0068] The fuzzy width learning model consists of a three-layer core structure: a fuzzification layer, a fuzzy inference layer, and a width learning layer.

[0069] Fuzzification layer: Gaussian membership function is used. The number of fuzzy subsets corresponding to each feature is 5 to 7 for example. The centers are initialized by Fuzzy C-Means Clustering (FCM) algorithm, and the width is calculated based on the sample dispersion after clustering (e.g., the standard deviation of samples of the same type).

[0070] Fuzzy inference layer: The TS (Takagi-Sugeno, hereinafter referred to as TS) fuzzy model is adopted. The number of fuzzy rules is 20 to 40 for example, and the activation intensity is the product of the membership degree of each feature.

[0071] For the wide learning layer: the number of feature nodes is typically 30-50, the number of booster nodes is typically 20-40, the activation function for feature nodes is tanh, and the activation function for booster nodes is sigmoid. The output weights are solved using ridge regression.

[0072] In this embodiment, a fuzzification layer and a fuzzy inference layer are added to optimize the width learning layer during heavy-haul train operation. The following specific steps are followed in the construction of the fuzzy width learning model: Step A31) Calculate the membership degree of each feature relative to different fuzzy subsets using the Gaussian membership function, defining the first... Input features In the The membership degree of each fuzzy subset is:

[0073] in, For the first Input features In the Membership degree in a fuzzy subset; For the first Input features In the The center of a fuzzy subset; For the first Input features In the The width of a fuzzy subset; B is the number of fuzzy subsets for each feature (e.g., 5 to 7).

[0074] This step provides the mathematical form of the membership function, but the central... and width It is not yet determined and needs to be completed by the next step of parameter estimation.

[0075] Step A32) Initialize the center of the fuzzy subset using the FCM algorithm. This allows the fuzzy subset to effectively cover the distribution range of the input features, and then the width is calculated based on the scattering of the clustered samples. :

[0076] in, For belonging to the first A sample set of fuzzy subsets. When input features near the center At that time, membership degree A value approaching 1 indicates that the feature completely belongs to the fuzzy subset; when the input feature... Far from the center At that time, membership degree A rapid decay to 0 indicates that the feature hardly belongs to the fuzzy subset; the smoothness of the Gaussian function ensures the continuity of the membership degree change.

[0077] At this point, all parameters in the membership function defined in step A31 have been determined, thus obtaining the specific membership function, which will be used in the next step to calculate the activation intensity of each fuzzy rule.

[0078] Step A33) Obtain directly through the FCM algorithm Fuzzy rules ( Example values ​​are 20-40). The activation strength of the fuzzy rule is calculated using the TS model and is defined as the activation strength from the input feature to the fuzzy rule. The membership degree of each cluster center (Gaussian function), or the product of the membership degrees of each feature (if a combination method is used).

[0079] Specifically, no. The form of a fuzzy rule is:

[0080] in, The first Input features in fuzzy rules The corresponding fuzzy subset; For the first The consequent constant term of a fuzzy rule; For the first The consequent coefficients of a fuzzy rule.

[0081] For the sake of simplicity, this embodiment directly calculates the first... Activation strength of fuzzy rules for:

[0082] in, For the first The input feature at the th ... Membership degree on a fuzzy subset; For the first The first fuzzy rule The fuzzy subset number corresponding to each input feature; The multiplication operation is equivalent to the "AND" operation in fuzzy logic. Fuzzy rules can only be fully activated when all conditions are met. This represents the total number of fuzzy rules (e.g., 20-40). This embodiment obtains this number using the FCM algorithm. After identifying cluster centers, the projection values ​​of each cluster center on each feature dimension are matched with the existing fuzzy subset centers from step A32, and the nearest fuzzy subset number is taken as the cluster center. After this step, each rule receives an activation strength. The activation intensities of all rules constitute the activation intensity vector. This vector will serve as the input to the feature nodes in the next width learning layer.

[0083] Step A34) Activate the intensity vector Feature nodes are mapped using random weights. :

[0084] in, Output vectors for feature nodes. The number of feature nodes is set to 30-50 in this embodiment; The feature map weight matrix is ​​initialized randomly. It is the bias vector; The hyperbolic tangent activation function is defined as:

[0085] The function outputs in the range (-1, 1), exhibiting a zero-center property, which helps improve learning efficiency. Therefore, the activation intensity vector... Transformed into feature nodes Feature nodes It will be used simultaneously to enhance node generation and the construction of joint feature matrices.

[0086] Step A35) Select feature nodes By mapping random weights to augmented node H:

[0087] in, To enhance the node output vector, Nh represents the number of enhanced nodes, which is set to 20-40 in this embodiment; The augmentation mapping weight matrix is ​​initialized randomly. It is the bias vector; The sigmoid activation function is defined as follows:

[0088] This function outputs in the range (0,1), introducing non-linear features and enhancing the model's ability to fit complex data. At this point, the required feature nodes for the wide learning layer are complete. Both the enhanced node H and the node have been obtained, and they will be spliced ​​together in the next step.

[0089] Step A36) By splicing feature nodes Together with the enhancement node H, they form the joint feature matrix A of the wide learning layer:

[0090] The joint feature matrix A contains all the information of the original activation strength after nonlinear mapping, and will be directly used to solve the output weights of ridge regression.

[0091] Step A37) The output weights w are solved using the ridge regression formula, without the need for iterative training:

[0092] The predicted values ​​of the fuzzy width learning model can be obtained. for:

[0093] in, The minimum tracking distance is the true value; λ is the regularization parameter (its value ranges from 1×10). -5 Up to 1×10 -3 When λ=0, ridge regression degenerates into ordinary least squares; when λ>0, introducing the λI term can solve the matrix... This helps identify potential singularities, prevent overfitting, and improve the model's generalization ability.

[0094] This step depends on the joint feature matrix A output from step A36 and the true values ​​of the training samples. Finally, the output weights of the width learning layer are obtained. The fuzzy width learning model was constructed.

[0095] Step 104: Use preset weights to weight and integrate the prediction results of the two models to obtain the trained ensemble model.

[0096] Specifically, the trained neural network model and the fuzzy width learning model are weighted and integrated according to preset weights to obtain a trained ensemble model, which outputs the minimum tracking distance prediction value:

[0097] in, and These are the preset initial weights for the neural network model and the fuzzy width learning model, respectively.

[0098] It should be noted that the "preset weight" mentioned in this application refers to the initial preset weight; under this initial preset weight, the weight of the neural network model is less than the weight of the fuzzy width learning model.

[0099] In a specific example, the initial preset weights of the neural network model are 0.4, and the initial preset weights of the fuzzy width learning model are 0.6. The ensemble formula is then:

[0100] It should be noted that as long as the weights of the neural network model are less than the weights of the fuzzy width learning model under the initial preset weights, the specific values ​​can be adjusted according to the actual data characteristics. Here, 0.4 and 0.6 are just examples. It should be understood that the preset weights here refer to the static weights used in the model training and initial deployment phases, which provide a stable starting point for subsequent dynamic adjustments.

[0101] Step 105: Use the trained ensemble model to predict the minimum tracking distance of heavy-haul trains in real time.

[0102] During train operation, real-time operating status data (including at least the speeds of the following and preceding trains) is collected. The real-time data undergoes the same feature engineering processing as in step 102 (including constructing braking propagation delay features, etc.). The processed feature data is then input into the trained neural network model and fuzzy width learning model respectively to obtain their respective predicted values. These predicted values ​​are then weighted and integrated according to the preset weights in step 104 to obtain the minimum tracking distance prediction value under the current operating conditions.

[0103] like Figure 3 and its local magnification Figure 4 , Figure 5 The figure shows a comparison curve of the predicted and actual minimum tracking distances for heavy-haul trains by various models. The horizontal axis represents the sample index (sorted in ascending order of the actual minimum tracking distance, totaling 1000 samples), and the vertical axis represents the minimum tracking distance (unit: cm, i.e., 330–360 m). The figure shows four curves: the thin solid line represents the actual value (i.e., the baseline value of the minimum tracking distance obtained through actual calculation or measurement), the dashed line represents the predicted value of the neural network model, the dotted-dash line represents the predicted value of the fuzzy width learning model, and the thick solid line represents the predicted value of the ensemble model (the method proposed in this application).

[0104] Overall, the true value curve shows a monotonically increasing trend (because the samples are arranged in ascending order of true values). The prediction curves of all models fluctuate around the true values, and the overall trend is basically consistent with the true values, indicating that each model can capture the changing pattern of the minimum tracking distance. Among them, the prediction curve of the ensemble model is closest to the true value and has the smallest fluctuation range; the prediction curves of the neural network model and the fuzzy width learning model fluctuate more and have more obvious local biases.

[0105] To show the details, Figure 3The text also includes elliptical annotations of two typical intervals with local magnification. See [link / reference]. Figure 4 ,Right now Figure 3 The magnified area in the upper left corner corresponds to sample indices 200-300 (low to medium tracking distance range). It can be seen that the curve of the ensemble model almost coincides with the true value, while the curves of the neural network model and the fuzzy width learning model show significant upward and downward deviations; see also... Figure 5 ,Right now Figure 3 The magnified area in the lower right corner corresponds to sample indices 700-800 (medium to high tracking distance range), which also verifies that the ensemble model can still accurately follow the true value at higher tracking distances, while other models have significantly larger prediction biases and fluctuations.

[0106] Figure 3 The superiority of the ensemble model proposed in this application in minimum tracking distance prediction is verified. Compared with a single neural network model or a fuzzy width learning model, the ensemble model exhibits the highest goodness of fit, the smallest error, and the strongest robustness to the true value, and maintains stable high-precision predictions across the entire range (low, medium, and high tracking distance segments). This figure can be used to verify the performance of the safe distance model in ATO and ATP, demonstrating that the method in this application can more accurately calculate the minimum safe tracking distance, thereby optimizing the following interval and improving line traffic efficiency while ensuring driving safety (avoiding rear-end collisions).

[0107] like Figure 6 The chart shown is a bar chart comparing the prediction errors of different models. The horizontal axis represents the three models (neural network model, fuzzy width learning model, and ensemble model), and the vertical axis represents the prediction error (unit: cm). The solid black bars represent the mean absolute error (MAE), and the hollow white bars represent the root mean square error (RMSE). The specific values ​​are as follows: the MAE of the neural network model is 79.7 cm, and the RMSE is 99.6 cm; the MAE of the fuzzy width learning model is 50.8 cm, and the RMSE is 66.8 cm; the MAE of the ensemble model is 44.7 cm, and the RMSE is 59.1 cm.

[0108] Depend on Figure 6It can be seen that the ensemble model has the lowest MAE and RMSE, indicating that it has the highest average prediction accuracy, the strongest ability to suppress large errors, and the most stable prediction. The fuzzy width learning model is next; the neural network model has the largest error and the worst prediction accuracy and stability. MAE reflects the average deviation of the model prediction. The ensemble model has an MAE of 44.7 cm, meaning that the average prediction error is less than 50 cm, which is beneficial for accurately planning the following interval and improving the efficiency of line operation. RMSE is more sensitive to extreme large errors. The ensemble model has an RMSE of 59.1 cm, indicating that it has almost no predictions that deviate significantly from the true value, while the neural network model has an RMSE as high as 99.6 cm, which has a large error and may pose a safety hazard in actual train control.

[0109] Figure 6 Yes Figure 3 Quantitative summary of the prediction comparison curve: Figure 3 This visually demonstrates that the prediction curve of the ensemble model best matches the true value. Figure 6 The accuracy advantage of the ensemble model is mathematically demonstrated through the MAE and RMSE metrics. The combined figures fully verify that the ensemble model proposed in this application outperforms both single neural network models and fuzzy width learning models in terms of accuracy and stability in the minimum tracking distance prediction task for heavy-haul trains, providing reliable algorithmic support for ATO and ATP safety distance calculations. Specifically, the neural network model excels at nonlinear mapping, while the fuzzy width learning model is adept at handling uncertainty and can learn incrementally. The two complement each other: the ensemble model's MAE is 44.7 cm, a 12% reduction compared to the single optimal model (50.8 cm for FBLs); the RMSE is 59.1 cm, an 11.5% reduction compared to 66.8 cm for FBLs. (Appendix) Figure 3 , Figure 6 As shown. Furthermore, the fuzzy width learning model has higher weights, which can suppress the risk of overfitting in neural network models under conditions of high uncertainty.

[0110] In yet another exemplary embodiment, an adaptive weighting mechanism and an incremental learning strategy are further introduced to adapt to the time-varying nature of operating conditions.

[0111] Step C1) Adaptive adjustment: A sliding window mechanism is used to dynamically adjust the preset weights based on the prediction error of each model within the window.

[0112] As a further optimization of the aforementioned static weighted ensemble scheme, this embodiment also introduces an adaptive weighting mechanism to adapt to the time-varying nature of train operating conditions. This mechanism dynamically adjusts the ensemble weights based on the recent prediction performance of each model, in addition to preset weights, thereby further improving the robustness and accuracy of the predictions.

[0113] Specifically, defining a window length L = 300 sampling points to capture the time-varying characteristics of train operation, the sliding window data at each moment is:

[0114] in, for The sliding window data at any given moment; This is the original time series signal; In the adaptive weighted ensemble strategy, the ensemble weights are dynamically adjusted based on the predictive performance of each model within the sliding window.

[0115] definition The root mean square errors of the predicted and true values ​​within the sliding window for the time-matter neural network model and the fuzzy width learning model are respectively... and The adaptive weights are calculated as follows:

[0116] in, and They are respectively The ensemble weights of the time-based neural network model and the fuzzy width learning model; It is a very small positive number (e.g., 10). -8 ), used to avoid a denominator of 0.

[0117] The initial window (the first 300 samples) can use fixed weights (such as 0.4 and 0.6), and adaptive adjustments will begin once the window is full.

[0118] like Figure 7 The graph shown is a curve illustrating the adaptive weight changes of the ensemble model. The horizontal axis represents the sample index (a total of 1000 test samples), and the vertical axis represents the ensemble weights (ranging from 0 to 1). The solid black lines in the graph represent the weights of the neural network model. The black dashed lines represent the weights of the fuzzy width learning model. The two gray dashed lines represent the initial baselines (initial weights of 0.4 for the neural network model and 0.6 for the fuzzy width learning model).

[0119] From sample 0 to 300 (the initial stage), the weights of the two models remain fixed (0.4 and 0.6), meaning that predictions are made using preset initial weights. Starting from sample 300, an adaptive adjustment phase begins. The weights of the neural network model increase significantly from 0.4 to the range of 0.65–0.72, then slowly decrease back to approximately 0.64; the weights of the fuzzy width learning model simultaneously decrease from 0.6 to the range of 0.28–0.35, then slowly increase back to approximately 0.36. The sum of the two weight curves remains approximately 1, consistent with the weight constraints of the ensemble model. The slight fluctuations in the curves reflect the model's real-time dynamic optimization of weights based on the prediction error of each sample, allowing the sub-model with higher prediction accuracy to receive a higher contribution percentage.

[0120] The core principle of this adaptive weighting mechanism is to dynamically adjust the weights (i.e., ensemble weights) based on the root mean square error of each model within the sliding window, assigning higher weights to models with smaller errors. Figure 7 It can be seen that after sample 300, the prediction accuracy of the neural network model is significantly better than that of the fuzzy width learning model. Therefore, its weight is greatly increased, making it the main contributor to the ensemble model. This dynamic weighting strategy leverages strengths and avoids weaknesses, enabling the ensemble model to maintain optimal prediction accuracy across the entire sample range. This explains why the ensemble model outperforms the single model in both MAE and RMSE metrics. This adaptive mechanism can adapt to changes in prediction accuracy under different train operating conditions (such as constant speed, braking, and acceleration), providing more robust and accurate algorithmic support for ATO and ATP safety distance calculations.

[0121] In this embodiment, the braking propagation delay feature can also be constructed in a time-series manner, for example, by considering the cumulative effect of historical braking commands. Define the braking propagation delay influence factor. :

[0122] Where u(k) is the braking command at the current moment (a value of 1 indicates braking, and 0 indicates no braking); The number of sampling points corresponding to the braking propagation delay (e.g., if the sampling period is...). In this embodiment, 300ms is used as an example, and the delay time is... ,but ); This is the attenuation factor, with a value ranging from 0.85 to 0.95. This feature can effectively capture the cumulative effect of braking commands during propagation.

[0123] In neural network model training, this embodiment introduces L2 regularization to prevent overfitting and defines a loss function. for:

[0124] in, This is the regularization coefficient, with a value ranging from 0.001 to 0.1; This represents the number of layers in the neural network model. For the first Layer weight matrix The square of the Frobenius norm.

[0125] Step C2) Incremental learning: Update the output weights of the fuzzy width learning model using an incremental formula in the form of recursive least squares.

[0126] The above incremental formula is:

[0127] in, and They are respectively Time and Output the weight vector at each time step; and They are respectively Time and The covariance matrix at time t; for The input feature vector at time step; for The true value of the minimum tracking distance at any given time.

[0128] Preferably, this embodiment also introduces a model performance monitoring mechanism. When the prediction error of the ensemble model exceeds a set threshold, it triggers model retraining or parameter adjustment. An adaptive threshold is defined. :

[0129] in, and These are the mean and standard deviation of the prediction error within the sliding window, respectively. When the current prediction error exceeds this threshold, an incremental model update or retraining is automatically triggered.

[0130] This application also provides an application scenario in which the aforementioned method for predicting the minimum tracking distance of heavy-haul trains is applied. Specifically: a heavy-haul railway line undertakes coal transportation tasks, and the train formation consists of "2 locomotives + 108 C80 type freight cars", with a total formation length of... =1450 meters, maximum operating speed 80 km / h, minimum safe deceleration 0.6 m / s 2 air braking propagation speed The speed is set to 200 m / s. The route includes complex sections such as long downhill slopes, tunnels, and curves, requiring real-time prediction of the minimum tracking distance between the following and preceding vehicles to ensure driving safety and improve the route's throughput.

[0131] Step 1: Data Acquisition and Preprocessing In the train operation control system, the speed of the following train is collected in real time with a sampling period of 300 ms. Speed ​​of the vehicle in front Data such as emergency braking distance, common deceleration, braking coefficient, and safety margin were collected. Data was continuously collected for 10 hours, yielding a total of 120,000 samples (M=120,000). Each sample contains 12 original features and 8 derived features.

[0132] Step 2: Feature Engineering Processing Feature engineering is performed on the original data to calculate the following derived features: Braking propagation delay characteristics: ; Kinetic energy difference term: When the speed of the following vehicle is greater than that of the preceding vehicle, Otherwise, it is 0; Braking delay distance: ; Relative distance characteristics, speed ratio characteristics, braking safety factor, braking time characteristics, relative speed characteristics, deceleration ratio characteristics, and braking efficiency coefficient.

[0133] All features are normalized using Z-score to eliminate the influence of dimensions.

[0134] Step 3: Model Building and Training A five-layer feedforward neural network model was constructed: an input layer with 20 neurons (corresponding to 20 features), a first hidden layer with 20 neurons, a second hidden layer with 15 neurons, a third hidden layer with 10 neurons, and an output layer with 1 neuron (predicting the minimum tracking distance). The hidden layer activation function was tansig, and the output layer activation function was purelin. The Model-Leyer (LM) algorithm was used for training, with a maximum of 1000 iterations and an error threshold of 1×10⁻⁶. -5 The dataset was trained using all 120,000 samples. After training, the MAE was 2.15m and the RMSE was 3.42m on the test set.

[0135] A fuzzy width learning model is constructed, with 5 fuzzy subsets (B=5) for each feature, and centers are initialized using FCM. The number of fuzzy rules is D=30. The width learning layer has 40 feature nodes and 30 augmentation nodes. Ridge regression regularization parameters are used. 1×10 -4The output weights are solved in one step on the same training set, without iteration. After training, the MAE is 1.86m and the RMSE is 2.91m on the test set.

[0136] Step 4: Weighted Integration Preset weights: The preset initial weights for the neural network model are 0.4, and the preset initial weights for the fuzzy width learning model are 0.6. The ensemble formula is:

[0137] The ensemble model achieved a MAE of 1.42m and an RMSE of 2.23m on the test set, both of which were superior to the single model.

[0138] Step 5: Real-time Prediction Application At a certain moment, the speed of the vehicle in front is 65 km / h (approximately 18.06 m / s), and the speed of the vehicle behind is 72 km / h (20.00 m / s). The distance between the two vehicles is 800 m. Ahead is a downhill slope of 6‰, with a braking coefficient of 0.85 and a safety margin of 50 m.

[0139] Real-time prediction process 1) Real-time collection of current operating status data: , The train has a length of 1450 m and a safety margin of 50 m.

[0140] 2) Calculate the braking propagation delay characteristics: .

[0141] 3) Calculate the kinetic energy difference term: .

[0142] 4) Calculate the braking delay distance: .

[0143] 5) After normalizing all features, input them into the neural network model and the fuzzy width learning model respectively: Neural network model outputs predicted values It is 152m; The fuzzy width learning model outputs the predicted value. It is 168m; 6) Weighted integration:

[0144] The ensemble model predicts a minimum tracking distance of 162 m (rounded).

[0145] The predicted value (162 m) is much smaller than the current actual distance (800 m), indicating that the current safety margin of the train is sufficient. The system does not issue a braking warning, and the train can maintain its current speed. Simultaneously, this predicted value is displayed in real time at the driver's cab and the dispatch center for subsequent safety interval adjustments and optimized scheduling. If the speed of the following train further increases or the speed of the preceding train decreases, the integrated model will dynamically update the predicted value. When the predicted value approaches the actual distance, the system will issue a braking suggestion to ensure driving safety.

[0146] This application example verifies the effectiveness of the proposed method in a real heavy-haul railway environment: by introducing braking propagation delay features, the model can perceive the risks caused by air braking delay; through weighted integration of a neural network model and a fuzzy width learning model, the prediction accuracy and robustness are both superior to a single model; the real-time prediction capability meets the timeliness requirements of train operation control.

[0147] Based on the same inventive concept, this application also provides a heavy-haul train minimum tracking distance prediction system for implementing the above-described heavy-haul train minimum tracking distance prediction method. The solution provided by this system is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the heavy-haul train minimum tracking distance prediction system provided below can be found in the limitations of the heavy-haul train minimum tracking distance prediction method described above, and will not be repeated here.

[0148] In one exemplary embodiment, a minimum tracking distance prediction system for heavy-haul trains is provided, comprising: The data acquisition module is used to acquire the operating status data of the heavy-haul train, and the operating status data includes at least the speed of the following train and the speed of the preceding train. A derived feature construction module is used to perform feature engineering processing on the running status data and construct derived features, including braking propagation delay features. The model building module is used to build a neural network model and a fuzzy width learning model based on the running state data and the derived features, respectively. An integration module is used to perform weighted integration of the prediction results of the neural network model and the fuzzy width learning model using preset weights to obtain a trained integrated model. The real-time prediction module is used to predict the minimum tracking distance of the heavy-haul train in real time using a trained ensemble model.

[0149] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores heavy-haul train operating status data, derived features after feature engineering, and model parameters. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for predicting the minimum tracking distance of heavy-haul trains.

[0150] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0151] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0152] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0153] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0154] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0155] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0156] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0157] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0158] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the minimum tracking distance of heavy-haul trains, characterized in that, include: Acquire the operating status data of heavy-haul trains, wherein the operating status data includes at least the speed of the following train and the speed of the preceding train; The operational status data is subjected to feature engineering to construct derived features, including a braking propagation delay feature. This braking propagation delay feature is estimated based on the train formation length and the air brake propagation speed, specifically as follows: in, For braking propagation delay time, For train Group length, The propagation speed of air braking; Based on the operational status data and the derived features, a neural network model and a fuzzy width learning model are constructed respectively; the construction of the fuzzy width learning model specifically includes: Fuzzification layer: Maps input features to fuzzy subsets and uses Gaussian membership functions to calculate the membership degree of each input feature relative to different fuzzy subsets; Fuzzy inference layer: The activation intensity of each rule is calculated based on multiple fuzzy rules, and the fuzzy rules adopt the TS fuzzy model; Width learning layer: The activation intensity is taken as input, and a width learning network is constructed through feature mapping and enhancement mapping. The output weights are solved by ridge regression. The prediction results of the neural network model and the fuzzy width learning model are weighted and integrated using preset weights to obtain a trained integrated model; The minimum tracking distance of the heavy-haul train is predicted in real time using a trained ensemble model; considering the braking propagation delay time, the minimum tracking distance is used in feature construction through the following physical model: in, The dynamic safety interval between trains, i.e., the minimum tracking distance, For safety margin, express The difference in kinetic energy between the following vehicle and the preceding vehicle at a given moment, when the speed of the following vehicle... greater than the speed of the vehicle in front When the time is right, the kinetic energy difference term takes a positive value; otherwise, it takes a zero value. For the minimum safe deceleration of the following vehicle, This indicates the braking delay distance of the vehicle behind.

2. The method for predicting the minimum tracking distance of heavy-haul trains according to claim 1, characterized in that, The preset weights are initial preset weights; under the initial preset weights, the weights of the neural network model are less than the weights of the fuzzy width learning model.

3. The method for predicting the minimum tracking distance of heavy-haul trains according to claim 1, characterized in that, It also includes: employing a sliding window mechanism to dynamically adjust the preset weights based on the prediction errors of each model within the window, wherein... The weights at each time point are calculated as follows: in, and They are respectively The ensemble weights of the time-based neural network model and the fuzzy width learning model; and They are respectively Root mean square error of the time-based neural network model and the fuzzy width learning model within a sliding window; It is a positive number.

4. The method for predicting the minimum tracking distance of heavy-haul trains according to claim 1, characterized in that, Also includes: When new data arrives, the output weights of the fuzzy width learning model are updated using an incremental learning approach. The incremental update formula is as follows: in, and They are respectively Time and Output the weight vector at each time step; and They are respectively Time and The covariance matrix at time t; for The input feature vector at time step; for The true value of the minimum tracking distance at any given time.

5. A system for implementing the minimum tracking distance prediction method for heavy-haul trains according to any one of claims 1-4, characterized in that, include: The data acquisition module is used to acquire the operating status data of the heavy-haul train, and the operating status data includes at least the speed of the following train and the speed of the preceding train. A derived feature construction module is used to perform feature engineering processing on the running status data and construct derived features, including braking propagation delay features. The model building module is used to build a neural network model and a fuzzy width learning model based on the running state data and the derived features, respectively. An integration module is used to perform weighted integration of the prediction results of the neural network model and the fuzzy width learning model using preset weights to obtain a trained integrated model. The real-time prediction module is used to predict the minimum tracking distance of the heavy-haul train in real time using a trained ensemble model.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the method for predicting the minimum tracking distance of a heavy-haul train as described in any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting the minimum tracking distance of heavy-load trains as described in any one of claims 1-4.

Images