A high-speed train idling and coasting detection method, device, medium and product

Abstract

The application discloses a high-speed train idling and coasting detection method, device, medium and product, relates to the high-speed train operation safety and control technical field, and the method comprises the following steps: obtaining the current absolute position of the train and the running distance time sequence, track image time sequence and driving speed time sequence in the past interval; determining the control state of the train according to the driving speed time sequence, and determining the traction error and braking error of the train in combination with the running distance time sequence and the current absolute position; according to the track image time sequence, the track surface dry and wet state of the track is predicted by using a machine learning model; the adhesion coefficient of the track is determined according to the average speed corresponding to the driving speed time sequence; and whether the train idles and coasts is predicted by using a logistic regression classification model according to the traction error, braking error of the train and the track surface dry and wet state and adhesion coefficient of the track. The application can improve the accuracy and adaptability of train idling and coasting detection.

Description

Technical Field

[0001] This application relates to the field of high-speed train operation safety and control technology, and in particular to a method, equipment, medium and product for detecting high-speed train idling and coasting. Background Technology

[0002] With the increasing density of high-speed rail networks and the continuous shortening of train intervals, extremely high demands are placed on the traction and braking performance of trains. However, during long-term operation, the wear and tear on the wheel-rail treads and the effects of extreme weather conditions such as rain and snow significantly reduce the adhesion coefficient between the wheels and rails, leading to slippage and coasting. Slippage and coasting not only reduce the train's power transmission efficiency but also prolong braking distance and can even cause major safety accidents such as rear-end collisions and derailments, seriously threatening operational safety. Therefore, detecting train slippage and coasting is an urgent problem to be solved.

[0003] Currently, traditional methods mainly employ fixed threshold methods. When the difference between the wheel circumferential velocity and the actual train speed exceeds a pre-set threshold, it is judged as slippage or coasting. These methods include speed difference threshold methods, acceleration threshold methods, and creep rate threshold methods. However, traditional methods often suffer from problems such as difficulty in obtaining the actual train speed and the need for filtering of acceleration values ​​containing high-frequency noise. Furthermore, the slippage or coasting thresholds vary significantly under different operating conditions. To ensure operational safety, a smaller threshold is usually chosen, often leading to overly sensitive detection and a high false positive rate. Summary of the Invention

[0004] The purpose of this application is to provide a method, equipment, medium, and product for detecting idling and coasting of high-speed trains, which can improve the accuracy and adaptability of train idling and coasting detection.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for detecting idling and coasting of high-speed trains, including: The system acquires the train's current absolute position, as well as the time series of its travel distance, track image, and speed over a past interval. For the past interval, the starting position is the location of the first transponder, and the ending position is the location of the second transponder. The current absolute position is provided by the second transponder. The first and second transponders are adjacent to each other. The control state of the train is determined based on the time sequence of its travel speed; the control state includes traction state, braking state, and constant speed state. The traction and braking errors of the train are determined based on the travel distance time series, the current absolute position, and the train's control status. Based on the time series of track images, a machine learning model is used to predict the dry and wet conditions of the track surface; the dry and wet conditions of the track surface include dry track, wet track, and snow track. The adhesion coefficient of the track is determined based on the average speed corresponding to the travel speed time series. Based on the train's traction error, braking error, and the track surface's wet / dry condition and adhesion coefficient, a logistic regression classification model is used to predict the train's operating status over a past interval; the operating status includes normal, idling, and coasting.

[0006] Secondly, 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 high-speed train idle and coasting detection method described above.

[0007] Thirdly, 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 high-speed train idle and coasting detection method described above.

[0008] Fourthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the high-speed train idle and coasting detection method described above.

[0009] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, device, medium, and product for detecting slippage and coasting of high-speed trains. First, it acquires the train's current absolute position and time series data of its travel distance, track image, and speed over a past interval. Then, it determines the train's control state based on the speed time series and, combined with the travel distance time series and current absolute position, determines the train's traction and braking errors. Based on the track image time series, it uses a machine learning model to predict the track surface's wet / dry condition. Based on the average speed corresponding to the speed time series, it determines the track's adhesion coefficient. Finally, based on the train's traction and braking errors, the track surface's wet / dry condition, and the adhesion coefficient, it uses a logistic regression classification model to predict whether the train has experienced slippage or coasting over a past interval. Compared to existing technologies, this application, based on multi-source data (including positioning data and visual information) combined with multi-dimensional feature extraction and collaborative analysis using artificial intelligence, can significantly improve the accuracy of identifying train slippage and coasting, reduce the false detection rate, lower the risk of accidents, and ensure the safe and efficient operation of trains under different weather conditions. Attached Figure Description

[0010] 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.

[0011] Figure 1 This is an application environment diagram of a high-speed train idle and coasting detection method according to an embodiment of this application; Figure 2 A flowchart illustrating a high-speed train idle and coasting detection method provided in an embodiment of this application; Figure 3 A schematic diagram illustrating error correction for time-travel distance-speed provided in an embodiment of this application; Figure 4 This is a schematic diagram of the change in train travel distance provided in one embodiment of this application; Figure 5 A schematic diagram of traction error provided for an embodiment of this application; Figure 6 A schematic diagram of braking error provided in an embodiment of this application; Figure 7 This is a schematic diagram of the YOLOv5 model framework provided in an embodiment of this application; Figure 8 A schematic diagram of three track surface dry and wet state detection results obtained based on the YOLOv5 model, provided in an embodiment of this application; Figure 9 A flowchart illustrating another high-speed train idle and coasting detection method provided in an embodiment of this application; Figure 10 This is a schematic diagram illustrating the collection of positioning and visual data provided in an embodiment of this application; Figure 11 This is a schematic diagram of partial positioning data from the EOAS data of a certain high-speed railway provided in an embodiment of this application. Figure 12 A schematic diagram illustrating the comparative experimental results between the proposed method and a conventional method provided in an embodiment of this application; Figure 13 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0012] As a critical infrastructure, the operational safety of high-speed railways is of paramount importance. Train operation on the tracks relies on the complex adhesion relationship between the wheels and rails. The adhesion coefficient defines the maximum traction or braking force that can be transmitted between the wheels and rails. However, under extreme weather conditions such as rain, snow, and melting snow, water films, ice layers, or oil contamination on the track surface drastically reduce the effective adhesion coefficient between the wheels and rails. In such cases, if the applied traction force exceeds the available adhesion limit, the drive wheelset will spin, meaning the wheel's circumferential velocity is higher than the train's overall actual operating speed. Conversely, if the braking force exceeds the adhesion limit during braking, the wheels will slip, meaning they tend to "lock up," and their circumferential velocity will be lower than the train's actual operating speed.

[0013] The hazards of wheel spin and coasting are extremely serious. Firstly, there is a loss of efficiency; wheel spin prevents the effective transmission of traction, wasting energy, while coasting significantly increases braking distance. Secondly, there is mechanical damage; severe wheel spin and coasting can cause abrasions on the wheel treads and rail surfaces, forming "flat scars" or "rail surface peeling," leading to continuous vibration and noise, and significantly increasing maintenance costs. Thirdly, there are serious safety risks; the increased braking distance caused by severe coasting is one of the direct causes of major safety accidents such as rear-end collisions, overstepping of warning signals, and even derailments. Therefore, real-time and accurate detection of wheel spin and coasting is crucial to ensuring the safe, efficient, and stable operation of high-speed trains.

[0014] Traditional methods based on fixed thresholds are currently the most common implementation in vehicle control systems. Their core principle is to monitor one or more physical quantities related to wheel spin and coasting; when their values ​​exceed a preset fixed threshold, an anomaly is determined. Specific methods include: the speed difference threshold method, which compares the difference between the wheel axle speed and the actual train speed; the acceleration threshold method, which monitors the acceleration of the wheels themselves, and judges wheel spin or the start of coasting when acceleration changes abruptly; and the creep rate threshold method, which calculates and monitors the creep rate based on creep theory. The advantages of these methods are their simplicity and fast response. However, their disadvantages are particularly prominent: the train's actual "ground speed" is difficult to measure directly and accurately, and relying on estimation introduces errors; the threshold settings are rigid, usually based on safety margins under the worst operating conditions, tending to be conservative to ensure no missed detections, resulting in a very high false detection rate under normal operating conditions or slight disturbances, frequently triggering unnecessary protective actions, affecting operational efficiency and passenger comfort.

[0015] Furthermore, with the development of artificial intelligence technology, especially the powerful nonlinear fitting and feature learning capabilities of deep learning and neural networks, it has shown potential in detecting wheel spin and coasting. For example, feedforward neural networks can be used to learn wheel-rail contact characteristics at different speeds to indirectly estimate the friction coefficient; adaptive adhesion models can be constructed by combining improved physical models (such as the Polach model) and AI algorithms; sensor data (such as acceleration and sound) can be input into one-dimensional convolutional neural networks or recurrent neural networks for end-to-end fault diagnosis. The advantage of such methods is that they can theoretically model more complex nonlinear relationships. However, the following problems also exist: poor model interpretability, AI models are often regarded as "black boxes," and their decision-making process lacks clear physical meaning. In the field of rail transit, where safety requirements are extremely high, it is difficult to gain the complete trust of operators and regulatory agencies; data dependence and generalization ability, the performance of the model depends heavily on the quality and coverage of the training data. For extreme weather or special operating conditions that are not fully reflected in the training data, the detection performance of the model will drop significantly, i.e., insufficient generalization ability; computational resource requirements, complex deep learning models have high computational requirements for onboard computing units.

[0016] In summary, existing technologies are either too simplistic and rigid, leading to high false positives, or too complex and "black box" structures, resulting in insufficient reliability and adaptability. This application aims to break through this technological deadlock by proposing an innovative solution that integrates multi-source information and combines physical interpretability with intelligent learning capabilities.

[0017] This application aims to address the problems of low accuracy, high false detection rate, and insufficient robustness under extreme weather conditions in existing train slippage detection methods. It provides an intelligent detection method that integrates multi-source positioning data and visual information to achieve accurate and adaptive slippage detection. This method significantly improves the accuracy of train slippage detection in extreme weather conditions, reduces the false detection rate compared to traditional methods, lowers the risk of accidents, and thus ensures the safe and efficient operation of trains under different weather conditions.

[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 high-speed train idle and coasting detection method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send the train's current absolute position, as well as the time series of travel distance, track image, and speed over the past interval to server 104. Upon receiving the data, server 104 determines the train's control status based on the speed time series; determines the train's traction and braking errors based on the travel distance time series, current absolute position, and train control status; predicts the track surface dryness / wetness using a machine learning model based on the track image time series; determines the track adhesion coefficient based on the average speed corresponding to the speed time series; and predicts the train's operating status over the past interval using a logistic regression classification model based on the train's traction and braking errors, track surface dryness / wetness, and adhesion coefficient. The operating status includes normal, idling, and coasting. Server 104 can then feed back the obtained train operating status over the past interval to terminal 102. In addition, in some embodiments, the high-speed train idling and coasting detection method can also be implemented by the server 104 or the terminal 102 separately. For example, the terminal 102 can directly process the train's current absolute position and the time series of travel distance, track image, and speed in the past interval. Alternatively, the server 104 can obtain the train's current absolute position and the time series of travel distance, track image, and speed in the past interval from the data storage system and process them.

[0021] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. The server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers, or it can be a cloud server.

[0022] In one exemplary embodiment, such as Figure 2 As shown, a method for detecting slippage and coasting of high-speed trains is provided. This method is executed by computer equipment, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 206. Wherein: Step 201: Obtain the train's current absolute position and the time series of travel distance, track image, and travel speed within the past interval; for the past interval, the starting position is the location of the first transponder, and the ending position is the location of the second transponder; the current absolute position is provided by the second transponder; the first transponder and the second transponder are adjacent.

[0023] Step 202: Determine the control state of the train based on the travel speed time sequence; the control state includes traction state, braking state and constant speed state.

[0024] Step 203: Determine the traction error and braking error of the train based on the travel distance time series, the current absolute position, and the control status of the train.

[0025] Step 204: Based on the time series of track images, use a machine learning model to predict the dry and wet state of the track surface; the dry and wet state of the track surface includes dry track, wet track, and snow track.

[0026] Step 205: Determine the adhesion coefficient of the track based on the average speed corresponding to the travel speed time series.

[0027] Step 206: Based on the train's traction error, braking error, and the track surface dryness / wetness and adhesion coefficient, a logistic regression classification model is used to predict the train's operating status over a past interval; the operating status includes normal, idling, and coasting.

[0028] As an optional implementation, step 203, "determining the train's traction and braking errors based on the travel distance time series, current absolute position, and train control status," specifically includes: Step 203.1: When the train is in traction control mode, calculate the traction error of the train based on the travel distance time series, and set the braking error of the train to 0.

[0029] Step 203.2: When the train is in braking mode, calculate the braking error of the train based on the travel distance time series, and set the traction error of the train to 0.

[0030] Step 203.3: When the train is in a constant speed control state, set both the traction error and braking error of the train to 0.

[0031] As an optional implementation method, the calculation process for traction error and braking error is as follows: 1-1) Calculate the difference between any two adjacent items in the travel distance time series, excluding the last item.

[0032] 1-2) Average the differences to obtain the average difference.

[0033] 1-3) Calculate the difference between the current absolute position and the second-to-last term in the travel distance time series to obtain the correction difference.

[0034] 1-4) When the train is in traction control mode, the average difference is subtracted from the correction difference to obtain the traction error.

[0035] 1-5) When the train is in braking control mode, the braking error is obtained by subtracting the average difference from the correction difference.

[0036] As an optional implementation, step 204, "predicting the wet / dry state of the track surface using a machine learning model based on the track image time series," specifically includes: Step 204.1: For each track image in the track image time series, use a machine learning model to predict the corresponding track surface wet / dry state.

[0037] Step 204.2: Select the prediction result with the largest number of occurrences from all the prediction results corresponding to the orbit images.

[0038] Step 204.3: When the proportion of the number of the prediction results exceeds a preset threshold, the dry and wet state of the track surface is determined based on the prediction results.

[0039] As an optional implementation, the machine learning model is a YOLOv5 convolutional neural network model (referred to as the YOLOv5 model); the frame rate of the orbital image time series is 2 frames per second.

[0040] As an optional implementation, the logistic regression classification model includes a linear transformation part, a logistic mapping part, and a final prediction output part; wherein, The mathematical expression for the linear transformation part is: ; in, Indicates category The corresponding linear output; , These represent normal, idle, and coasting, respectively. For the input feature vector, , Indicates traction error. Indicates braking error. Indicates the dryness or wetness of the track surface. Indicates the adhesion coefficient of the track; the superscript T indicates transpose; Indicates category The corresponding weight vector, , respectively with One-to-one correspondence; Indicates category The corresponding bias term.

[0041] The mathematical expression for the logical mapping part is: ; in, Represents the input feature vector The corresponding sample belongs to the category The conditional probability.

[0042] The mathematical expression for the final prediction output is: ; in, This indicates the category of the final predicted output.

[0043] As an optional implementation, the formula for calculating the adhesion coefficient is: ; in, The viscosity coefficient is 1. This represents the average speed corresponding to the time series of driving speeds.

[0044] To aid understanding by those skilled in the art, further explanation is provided below.

[0045] The technical solution of this application is based on the combination of multi-source positioning data fusion and track visual information, and constructs a multi-input multi-output classification model to identify the detection method of train slippage and coasting. Specifically, the technical solution achieves the patent objective through the following three core parts.

[0046] (1) Multi-source positioning data acquisition and error feature extraction.

[0047] First, autonomous positioning is achieved through a combination of odometer and transponder polling. Specifically, the onboard odometer, based on a speed sensor, records the cumulative number of wheel revolutions and then multiplies it by the wheel circumference to calculate the distance traveled. The transponder is precisely positioned on the track with zero error in its own position. When the train passes a transponder, the absolute position provided by the transponder is used to correct the cumulative error of the odometer.

[0048] Figure 3 A sample of field data recorded by transponders and odometers is presented.

[0049] Specifically, the horizontal axis (time) comes from the train positioning system, and the vertical axis (distance and speed) comes from the speed sensor. The blue curve represents the distance traveled, the red curve represents the speed, and the black dashed line indicates the moment the train passes the transponder. The train's traction and braking status can be determined from the speed curve. In particular, after transponder error correction at 09:22:38, the distance curve shows a slight but sharp increase; after transponder correction at 09:22:52, the distance curve rate slows down slightly. Between 09:22:45 and 09:22:52, the train passes two adjacent transponders consecutively; during this period, the train is in the traction phase, and its speed continuously increases. The error corrected during this phase is called the traction error. Between 09:22:31 and 09:22:38, the train passes two adjacent transponders consecutively; during this period, the train is in the braking phase, and its speed decreases. The error corrected during this phase is called the braking error.

[0050] After data collection, the data undergoes preprocessing and error analysis, such as... Figure 4 As shown. Preprocessing of sensor data is as follows: Figure 4 As shown in the figure, this graph illustrates the change in the distance traveled between adjacent transponders (RD) over time, where the train... Time passing transponder ,exist Time passing transponder The horizontal axis represents the time within the train positioning system, using... The vertical axis represents the distance traveled by the train as recorded by the odometer (Running Distance, RD), expressed in seconds. Indicated, the unit is meters. Calculation. and The change in the distance traveled between them is denoted as As shown in the following formula: ; It is important to note that Indicates at time through The change after position correction. Calculation. arrive The average change between them is denoted as It reflects the train in arrive The average operating status during the period is shown in the following formula: ; use , , D j+1Let's analyze the RD error. When the train is in traction, all variables have the subscript T; when the train is in braking, all variables have the subscript B. Assume (assuming variables have...) The train is Never through the transponder And without position correction, such as Figure 5 and Figure 6 The red and blue dashed lines in the figure show the curve after position correction.

[0051] like Figure 5 As shown, the train is in traction mode. Assume that... arrive Between, the change in RD is equal to arrive Average change between As shown in the following formula: ; Generally speaking, because the train is in the traction acceleration phase, > However, it was actually discovered that: > Analysis revealed that when the train was... arrive When the axle spins freely, its rotational speed increases rapidly, causing the odometer to record a rapid increase in distance traveled. Therefore, the average change in travel distance... This will also increase accordingly. At any given moment, due to the accumulation of errors, the value recorded by the odometer exceeds the value recorded at the specified time. The actual position value at that moment, and close to it. It is always close to the true value. Time via transponder After the train's position was corrected arrive The change in RD between It will become smaller, thus appearing > Therefore, due to the accumulation of errors, the red dashed line is above the black curve, and these errors are mainly caused by wheel spin during traction. Therefore, and The difference between them is considered as traction error, denoted as As shown in the following formula: ; To avoid frequent detections, a threshold percentage θ can be set, as shown in the following formula: .

[0052] Threshold triggering: To avoid noise interference, feature flags are generated only for significant errors that exceed a set threshold (e.g., θ=25%).

[0053] That is, during the train's traction acceleration phase, if wheel spin occurs, the wheel speed will abnormally increase, causing the odometer to record a "falsely high" distance traveled in the first half of the section. When the train is in... When the point is "pulled back" to its true position, the "corrected travel distance" in the latter half of the interval will appear abnormally small. This method accurately calculates the traction error by comparing the difference between the "average change assuming no idling" and the "actual corrected change." The larger this value, the more severe the idling.

[0054] During the train's braking and deceleration phase, if skidding occurs, the wheels tend to lock up, and the rotational speed decreases or even drops to zero, causing the odometer to record a falsely low distance traveled in the first half of the section. When the train is in When correction is applied, the "corrected travel distance" in the latter half of the interval will appear abnormally large. This method accurately calculates the braking error by comparing the difference between the "actual corrected change" and the "average change assuming no slippage". The larger this value, the more severe the slippage.

[0055] Similarly, such as Figure 6 As shown, the train is in a braking state. Assume that... arrive Between, the change in RD is equal to arrive Average change between As shown in the following formula: ; When the train is arrive When slippage occurs, the wheel axle nearly locks up. During this period, the distance traveled recorded by the odometer changes very little or not at all, resulting in... The value is very small. At any given moment, the value recorded by the odometer will be less than the actual value at that location, and even more so, less than... The actual position value at any given moment. Therefore, in Time passing transponder After position correction, and Change in travel distance between The value will be very large. Therefore, the following will occur: > Therefore, the blue dashed line will be below the black curve.

[0056] This phenomenon is mainly caused by wheel-rail lock-up during braking, leading to wheel slippage. We will... and The difference between them is called braking error (skid error), as shown in the following formula: ; Through the above error analysis algorithm, two important characteristics were obtained: traction error and braking error.

[0057] Generally, the number of transponders in a transponder group of a CTCS-3 level train control system should not exceed 3; the distance between adjacent transponders within a transponder group should be 5m ± 0.5m; the interval between transponder groups should not be less than 200m; when the distance between two adjacent transponder groups in a CTCS-3 level section is greater than 1500m, a positioning transponder group should be added between the two transponder groups for train positioning. (That is, the maximum interval between transponder groups should not exceed 750m) Note: In this application, adjacent transponders are not from the same transponder group, but rather the last transponder in the previous transponder group and the first transponder in the next transponder group.

[0058] (2) Visual information acquisition and track surface status recognition.

[0059] First, visual information is acquired using onboard cameras installed in the train's cab, which continuously capture images of the track surface ahead of the train at a certain frame rate (e.g., 24fps), collecting data on different weather conditions (sunny, rainy, snowy). Then, an algorithm is designed to identify the dryness or wetness of the track surface (dry rail, wet rail, snowy rail). Finally, track surface features and adhesion coefficients are extracted.

[0060] The cockpit camera acquires data at a rate of 24 frames per second, processing one image every 0.0417 seconds. This is a waste of resources for detecting slow-changing static features like the track's wetness or dryness. Therefore, the data is downsampled. Reducing the frame rate to 2 frames per second decreases the latency to 0.5 seconds. Assuming the train travels at 300 km / h, each frame covers 41.67 meters of track length, allowing for rapid detection of the overall track surface wetness or dryness while minimizing the chance of missing localized conditions.

[0061] After preprocessing the camera data, predictions are made based on the YOLOv5 convolutional neural network model. This model adopts a modular design, consisting of a backbone network, a neck network, and a head network. The modular design facilitates subsequent targeted adjustments and upgrades to each network module according to different detection needs, thereby better adapting to target detection tasks in railway scenarios and providing more reliable technical support for the safe operation of railways.

[0062] See Figure 7 , Figure 7Each module is a built-in module of the YOLOv5 model.

[0063] The following section describes the various modules in the Backbone network.

[0064] The Focus module is a downsampling module used to sample the input orbital image (640×640×3) into four 320×320×3 feature maps by alternating rows and columns, and then stitch them together to form a 320×320×12 high-channel feature map, achieving downsampling with "no information loss" and reducing the amount of computation.

[0065] The CBS module is the basic convolutional unit (convolutional layer + batch normalization layer + SiLU activation function). Convolution extracts local features, batch normalization stabilizes the training distribution, and SiLU activation introduces non-linear expression. It is the basic component of the entire network.

[0066] Bottleneck = True: This is a residual unit with residual connections. First, dimensionality is reduced using 1×1 CBS, then orbital features are extracted using 3×3 CBS depth, and finally, residual connections preserve the original orbital information.

[0067] The `add` module is used for residual connections. It directly adds the output of Bottleneck to the input features (the parts that pass through Bottleneck and the parts that do not pass through Bottleneck), preserving key features of the orbital region and avoiding the loss of important information during training.

[0068] Bottleneck = False: Residual unit without residual connections. The structure is the same as Bottleneck=True, but without residual connections. It is used in scenarios where the number of channels does not match or where residuals need to be weakened.

[0069] The 1×1conv module uses 1×1 convolutions to adjust the number of channels and focus on track-related feature channels (such as enhancing the channel weights of wet track reflections and snow track textures, and suppressing interference from irrelevant backgrounds).

[0070] CSP1: The CSP (Cross Stage Partial) backbone module processes track features in two ways. One way is to extract track features through two Bottleneck depths; the other way is to quickly adjust the channels using only 1×1 conv. Finally, the two features are concatenated, which reduces the amount of computation while retaining rich feature information about the track's dry and wet states.

[0071] The concat module: Feature stitching: merges track features from different branches in the channel dimension (e.g., fusing multi-dimensional information such as wet track reflection, snow track texture, and dry track rust) to improve feature richness.

[0072] The following describes the various modules in the Neck network.

[0073] CBS module: Re-extracts the fused orbital features to enhance key features.

[0074] The subsampled module (downsampling) reduces the size of high-resolution low-level detail features (such as the edges of water stains and the texture of snow) to match the size of low-resolution high-level semantic features (such as the overall orientation of the track and its wetness or dryness), allowing low-level features to be passed to higher levels and enhancing the semantic understanding of the track region.

[0075] CSP2 module: The neck module of the CSP structure, which is simpler than CSP1. It internally stacks a bottleneck without residuals. Its function is to reduce the amount of computation during feature fusion while maintaining the integrity of the features.

[0076] The upsampled module amplifies low-resolution high-level semantic features to match the size of high-resolution low-level detail features, allowing the high-level features to acquire more detailed information.

[0077] The Backbone outputs feature maps at three scales: P3 (red channel, 80×80, low-level detail features, responsible for small water stains, snow textures, etc.), P4 (yellow channel, 40×40, mid-level balanced features, responsible for medium-sized snow accumulation and localized wet spots), and P5 (blue channel, 20×20, high-level semantic features, responsible for the overall track area and large areas of wet track / snow accumulation). The Neck network achieves repeated interaction between low-level detail features and high-level semantic features through multiple rounds of upsampling, downsampling, and concat-CSP2-concat processing.

[0078] Taking P3 combined with P4 as an example: P3 is subsampled, compressing the feature map size from 80×80 to 40×40, aligning it spatially with the mid-layer P4 feature. Then, concat is used to stitch the downsampled P3 feature with the original mid-layer P4 feature, transferring the subtle details of the low-layer to the mid-layer scale. CSP2 is then used to recombine the concat-stitched features. Upsampled is then used to restore the recombinated 40×40 mid-layer fused feature to 80×80, realigning it with the low-layer P3 high-resolution feature. Finally, concat is used for a second feature stitching, completing a round of subsampled → concat → CSP2 → upsampled → concat, which allows the high-resolution P3 feature to fully retain the details of water stains and snow textures while fusing mid-layer features.

[0079] In the Neck network, features at each scale undergo multiple rounds of fusion, including concat → CSP2 → concat, which enhances the interaction between low-level detail features (textures of water stains and snow) and high-level semantic features. Multiple upsampled and subsampled processes ensure more complete feature transfer across different scales, improving the detection sensitivity for subtle local targets (water stains and snow).

[0080] The following section describes the various modules within the Head network.

[0081] The 1×1 conv module transforms the feature maps at three scales: P3 (80×80, red), P4 (40×40, yellow), and P5 (20×20, blue). It outputs a predictive tensor for each scale, containing the position offsets of all anchor boxes, confidence scores, and class probabilities. Its function is to convert the number of channels in the feature map into the dimension required for detection: 3×(5+nc), where 3 represents 3 anchor boxes per pixel, 5 = center coordinates x / y + width / height w / h + confidence score, and nc = number of classes. In this embodiment, nc = 3, corresponding to dry tracks, wet tracks, and snow tracks.

[0082] The 1×1 conv module allows for early verification of detection results during dimensionality transformation and tensor prediction. For example, in the image, the 1×1 conv module connected to the P3 path output will identify that the texture of the gravel area outside the track does not conform to the water stain feature. In the tensor prediction, the confidence level of the anchor box in this area will be set far below the threshold, and it will be marked as "invalid prediction" in advance.

[0083] The Merge Detections module merges the prediction tensors at the P3, P4, and P5 scales into a single global prediction tensor, integrating anchor box prediction information from all scales.

[0084] Bounding box + NMS (Target Box + Non-maximum Suppression): From the merged prediction tensor, combined with the preset anchor boxes, the coordinates of the true detection box corresponding to each anchor box are calculated, and the confidence and category are obtained. The candidate boxes with low confidence are filtered out by using a confidence threshold, and the boxes with the highest confidence are kept to remove redundancy.

[0085] Bounding boxes combined with NMS (non-maximum suppression) can be used in conjunction with validation logic to filter the initial detection and classification results. The detection results are as follows: Figure 8As shown, dry rail represents a dry rail, wet rail represents a wet rail, and snow rail represents a snow rail; 0.61, 0.50, and 0.47 represent the corresponding confidence levels. The validation logic is implemented based on prior knowledge, such as the specular reflection characteristics of wet rails and the texture uniformity of snow rails. Wet rails also have characteristics such as dispersed bright reflective points and large areas of continuous bright areas. Note that after the model is trained, it essentially contains the prior knowledge and validation logic.

[0086] The track adhesion coefficient is affected by many factors, including not only the track's wet / dry condition but also the train speed. A speed sensor is used to obtain the train speed. Based on the currently widely used practical formula for the adhesion coefficient of wet rails in China: ; This serves as a unified formula for calculating the adhesion coefficient (traction and braking) for high-speed trains and a formula for calculating the adhesion coefficient (braking) for conventional trains. During calculation... The input is the average speed corresponding to the time series of driving speeds.

[0087] The above analysis reveals two important characteristics: the track's wet and dry state and the adhesion coefficient.

[0088] (3) Construction of multi-input multi-output detection model.

[0089] A multiple-input multiple-output (MIMO-LR) logistic regression classification model is constructed. Compared with other deep neural networks, the logistic regression model has low computational cost, is suitable for real-time in-vehicle inference, and is highly efficient. The model's weight coefficients intuitively reflect the contribution of each input feature to the final decision, meeting the safety interpretability requirements of rail transit and demonstrating strong interpretability. It can output the probability of belonging to each category, providing support for subsequent risk assessment and graded early warning.

[0090] The input consists of four key features (traction error, braking error, dry and wet states of the rail surface (DWS), and adhesion coefficient), and the output consists of three classification results (normal, idling, and coasting).

[0091] The model is trained using a large amount of historical data on normal operation, artificially induced spinning or gliding, for supervised learning; the loss function uses cross-entropy loss plus L2 regularization to prevent overfitting and improve the model's generalization ability; and the model parameters are optimized using gradient descent.

[0092] Four evaluation metrics were used to measure model performance: precision, recall, F1 score, and area under the curve (AUC).

[0093] The MIMO-LR model mainly consists of two parts: linear transformation and logistic mapping.

[0094] ; in, , representing the three predicted categories (normal, idling, gliding). A one-vs-Rest multi-class classification strategy is employed, where the model learns an independent set of parameters for each category M. n represents the number of training samples (n). N is 264, meaning 264 training samples are used. Represents the input feature vector . This represents the transpose of the weight vector corresponding to category M. . This represents the bias term corresponding to category M, used to adjust the intercept of the linear output. Let represent the linear output of class M, which is the result of a linear transformation and is subsequently mapped to a probability using the sigmoid function, as shown in the following equation: ; in, Indicates the given input When the condition is met, the conditional probability that a sample belongs to the Mth class (denoted by y) is given by the Sigmoid function. Compress it to the (0, 1) interval so that it falls within the range of probability values.

[0095] loss function The weight vector will be adjusted. The m-th element (1) m 4) to reduce The value of L2 regularization is used to improve prediction accuracy. It consists of two parts: cross-entropy loss (first half) and L2 regularization (second half), as shown in the following equation: ; in, It is the one-hot encoding of the true label of sample n; if sample n belongs to class M, then =1; otherwise =0.

[0096] Cross-entropy loss reflects the predicted probability With real labels The degree of matching. The L2 regularization term penalizes larger matches by squaring the weights. Values ​​are used to avoid overfitting. This represents the strength of regularization.

[0097] and The gradient update is shown in the following equation: ; in, and Represent the loss function respectively right and The gradient is used to represent and The direction and extent of the impact on the loss; It is an L2 regularization pair The gradient, which is obtained by penalizing larger values. The value is used to constrain the weight to a small range. This represents the learning rate in this model. =0.01.

[0098] Finally, the model iterates through all values ​​of M to find the one that makes... The maximized value of M is used as the prediction result, as shown in the following formula: ; in, This represents the final predicted category. Finally, the model iterates through all M values ​​to find the one that makes the prediction correct. The maximum value of M is used as the prediction result.

[0099] Feature values ​​acquired from multiple sensors are fused and then input into the MIMO-LR model for training. The final prediction results are then reviewed by experts. The flowchart is as follows. Figure 9 As shown.

[0100] This problem involves feature extraction through train travel distance error analysis and video information processing. The model takes four key features as input and three judgment categories as output, including the time, location, and frequency of wheel spin. Specifically, the model determines whether wheel spin has occurred in the area between two adjacent transponders, so the time and location of wheel spin are based on this area. The frequency is a statistical result of each detection completed by the model.

[0101] A key feature of this technical solution is the fusion of multi-source sensor data with visual information, employing a machine learning model to address specific safety issues in high-speed train operation. This approach significantly improves the efficiency and accuracy of slippage detection, reduces false detection rates, and ensures operational safety.

[0102] In addition to the basic technical solution described above, this application can also be improved or extended through the following alternative solutions: (1) Introduce more sensor modalities: For example, acoustic sensors, adding microphone arrays, and collecting the sound emitted when the wheel and rail are in contact. Specific howling or friction noise will be generated when the wheel is spinning or sliding, and the audio characteristics can serve as an effective supplement to visual and positioning data; infrared thermal imagers, when the wheel is spinning or sliding severely, the contact surface will generate local high temperature due to frictional heat. Thermal imaging can intuitively capture this "hot spot" and provide strong evidence for judgment in extreme situations; 3D lidar or depth cameras: used to more accurately detect the depth of water accumulation, snow thickness or foreign object intrusion on the track surface, providing richer three-dimensional environmental information than 2D images.

[0103] (2) Model-level substitution and optimization: In addition to a single MIMO-LR, multiple different types of models (such as XGBoost, SVM, neural networks) can be trained, and a meta-classifier can be designed to dynamically select the most reliable model results based on the current operating conditions (such as speed range, weather) to form a more powerful integrated system.

[0104] (3) System Architecture Upgrade: From a system architecture perspective, edge-cloud collaborative computing can be considered, placing computationally intensive tasks such as visual inspection on the vehicle's edge computing unit for real-time execution to ensure low-latency response. Simultaneously, a large amount of runtime data and model update tasks can be synchronized to the cloud platform for large-scale data analysis, model retraining, and fleet-level knowledge sharing. (4) Expanding application scenarios: Extending this technology to urban rail transit, heavy-haul railways, and even trolleybuses. These scenarios also face adhesion control problems, and the operating environment is more complex, with equally urgent needs.

[0105] Beneficial effects: This application effectively overcomes the shortcomings of traditional methods (such as the acceleration threshold method) that rely on a single signal and ignore environmental factors, resulting in a high false detection rate, by integrating multi-sensor data and visual information. Its core principle is to extract traction and braking errors with clear physical meaning using transponder and speed sensor data to directly quantify slip and coasting phenomena. At the same time, the onboard camera identifies the dry and wet state of the track surface and dynamically estimates the adhesion coefficient, thereby jointly representing the internal and external states of wheel-rail contact as a multi-dimensional feature vector. Finally, the vector is input into a multi-input multi-output logistic regression model for collaborative analysis and state classification, achieving accurate and reliable identification of slip and coasting in complex operating environments. Compared with traditional methods, it significantly reduces the false detection rate by 35.29%, improving the safety and adaptability of high-speed train operation.

[0106] Here is an example.

[0107] One implementation approach for this example is to deploy the detection system on a high-speed railway, utilizing speed sensors, transponders, and onboard cameras to collect multi-source information, including positioning data, speed data, and video data of the train's movement ahead. Figure 10 As shown.

[0108] First, the positioning data is preprocessed to calculate the change in running distance between adjacent transponders and its average value, thereby extracting traction and braking errors. Visual data is used to identify the dry and wet states of the track surface through the YOLOv5 model, and the identification results are combined with real-time speed data to estimate the adhesion coefficient using the wet track formula. Subsequently, the above features are input into a pre-trained MIMO-LR model, which outputs the current train status as idling, coasting, or normal. The system has been verified by experts in actual operation, and the detection results are reliable.

[0109] (1) Scene setup and system deployment 1) Test subject: A train traveling south on a certain high-speed railway line.

[0110] 2) Test section: Downstream section between a certain West Station and a certain East Station.

[0111] 3) System Hardware Configuration: An odometer is used. Standard axle-end speed sensors are installed on the bogies of the lead car. The train's travel distance is obtained by accumulating the number of wheel revolutions and multiplying them by the wheel circumference. This data includes accumulated errors and will produce significant deviations during idle or coasting. A transponder antenna for receiving positioning signals is installed under the car. An industrial-grade camera with an IP67 protection rating (1920×1080 resolution, 24fps) is installed inside the windshield at the front of the cab. All sensor data is connected to an industrial computer (equipped with an Intel i7 processor and 16GB RAM) in the cab for real-time processing.

[0112] Transponders are laid on the track, providing precise absolute geographic coordinates. As a train passes, the transponder information is used to correct for accumulated errors in the odometer. Its positional error can be considered zero.

[0113] (2) Data acquisition and preprocessing 1) Location data: At 10:15:17 EOAS system time, the train passed transponder 069-3-23-009-2, and at 10:15:28 passed the adjacent transponder 069-3-23-011. The odometer readings increased from 220986.7 meters to 221916.4 meters. During this section, the train speed increased from 281 km / h to 288 km / h. The weather that day was light rain, and the track surface was wet.

[0114] 2) Visual data: Within the adjacent transponder interval, the camera acquires track images at a rate of 2 frames per second. The acquired images are fed into the YOLOv5 model (which has been pre-trained on thousands of dry, wet, and snow-covered track images) for inference in real time.

[0115] After aligning the timestamps and organizing them, the train's operation data for the (j+1) time periods can be obtained as shown in Table 1: Table 1 Train Operation Data Table

[0116] Among them, transponder and transponder It is an adjacent transponder.

[0117] (3) Feature extraction process: 1) Calculation of travel distance error: The computer reads the train's travel distance error... Passing the transponder at 10:15:17 (Number 069-3-23-009-2), in Passing the transponder at 10:15:28 Intermediate odometer (ODO) records (10 records in total, j=9) between (number 069-3-23-011). Calculate the change in distance traveled at adjacent time points. and calculate arrive Change in average travel distance between =85.95 m (the train is in the traction phase in this section), at arrive ( Calculate between ) =39.40 m. Threshold Taking 25%, the calculation yields: ; ; Exceeding the threshold This triggered a potential "idle" characteristic.

[0118] 2) Track state identification and adhesion coefficient estimation: The YOLOv5 model analyzes the acquired image sequences and detects the "wet rail" feature in more than 90% of the frames.

[0119] Therefore, the track dry / wet condition (DWS) feature is quantified as wet track = 1 (dry track = 0, snow cover = 2).

[0120] At the same time, the system reads from the speed sensor that the train has departed from... to The average velocity corresponding to the velocity sequence at time t is 284 km / h. The estimated viscosity coefficient is obtained using the viscosity coefficient calculation formula. The value is 0.0739. The MIMO-LR model predicts the probability of each class as follows: P(idle spin| = 0.88; P (gliding| = 0.05; P(Normal| = 0.07; According to the arg max rule, the model will eventually output the current state as "idling".

[0121] In some embodiments, the method further includes an expert knowledge review and online model learning closed loop: the obtained classification results and their confidence scores, along with the original feature data, are pushed to the human-computer interaction interface for review by drivers or ground experts; when manual correction feedback is received, the data and its corrected label are stored in an incremental learning database; periodically or when the data accumulates to a certain scale, the database is used to incrementally update and train the multi-input multi-output classification model to achieve continuous optimization and adaptation of the model.

[0122] To evaluate the performance of the high-speed train slippage and coasting detection method based on the fusion of multi-source positioning data and visual information features, this embodiment conducted three sets of experiments. The first set involved a threshold comparison experiment, exploring the impact of different threshold values ​​on the travel distance error analysis algorithm and the traditional acceleration thresholding method. The second set involved a multi-model comparison experiment, studying the advantages of this method compared to other popular classification models under both weather-considered and weather-independent conditions. The final set involved a comparison experiment between this method and traditional methods, evaluating the advantages of this method over the traditional acceleration thresholding method.

[0123] (1) Threshold comparison test: The threshold value is crucial to the detection sensitivity. In the analysis of travel distance error, the threshold... The magnitude of the threshold determines the amount of traction error and braking error; in the acceleration threshold method, the threshold... This also determines the sensitivity of the detection. Therefore, we conducted a threshold comparison experiment based on 4784 observation data points from a specific high-speed train on a specific day and month within a 407.32-kilometer operating section. Some positioning data are as follows: Figure 11 As shown in Table 2, the experimental comparison results are presented.

[0124] Table 2 Threshold Comparison Results

[0125] Through comparative analysis, it was found that in the acceleration threshold method: threshold The magnitude of the threshold has a significant impact on the fluctuation of the detection results; while the threshold in the walking distance error analysis algorithm... The size has little impact on the detection results. Final selection =1.0 , =25%.

[0126] (2) Multi-model comparison experiment: Precision, recall, F1 score, and area under the curve (AUC) were used as evaluation metrics. ; ; ; EOAS data from a specific high-speed railway line was used during a specific date and time period. This data covers 25 trains, generating 158,369 observation records with a cumulative mileage of 25,603.93 kilometers. Using this data, the MIMO-LR model was compared with K-NN (K-Nearest Neighbors), XGBoost, Support Vector Machines (SVM), Naive Bayes, and RandomForest in two scenarios: one considering weather conditions (W scenario), and the other not considering weather conditions (NW scenario). Table 3 shows the experimental results for each model.

[0127] Table 3 Results of the multi-model comparison experiment

[0128] The MIMO-LR model is built using the Scikit-learn library and preprocessed and analyzed using the Pandas library. Comparison reveals that in the W scenario, the Naive Bayes model has the worst overall performance, but its AUC metric is good, meaning it can distinguish between positive and negative samples, but cannot clearly rank them. K-NN, XGBoost, SVM, and Random Forest have similar performance. In the NW scenario, SVM performs best, while Random Forest performs even better. When weather conditions are considered, the performance of all models improves, with the MIMO-LR model performing best.

[0129] (3) Comparison experiment with traditional methods: The MIMO-LR method was compared with the acceleration threshold method (ATM) in the traditional method on high-speed rail data. The experiment used the operation data of a certain train on a certain day of a certain month of a certain year, covering a distance of 407.32 kilometers and 6706 observation records. Compared with ATM, MIMO-LR reduced the number of idle coasting detections from 17 to 11, reduced the number of false detection samples by 6, and reduced the false detection rate by 35.29%. The comparison results are as follows. Figure 12 As shown.

[0130] This application also provides an application scenario in which the above-mentioned high-speed train slippage and coasting detection method based on the fusion of multi-source positioning data and visual information features is applied. Specifically, the slippage and coasting detection method provided in this embodiment can be applied to a high-speed train active safety protection system. Slippage and coasting detection includes a state perception stage and a control decision stage; the state perception stage is used to identify the train's operating state based on multi-source sensor data; the control decision stage is used to trigger corresponding traction / braking control strategies based on the detection results; the slippage and coasting detection method based on the fusion of multi-source positioning data and visual information features in this embodiment belongs to the state perception stage.

[0131] This application detects train slippage and coasting by fusing multi-source positioning data and combining it with track visual information. Through multi-source positioning data fusion, the advantages of multiple sensors can be complemented, train positioning errors can be corrected, and preliminary practical judgment of the detection results can be provided, solving the problem of poor model interpretability commonly found in artificial intelligence detection methods. Furthermore, the dryness and wetness of the track surface can be intuitively obtained through visual information, and the adhesion coefficient can be estimated. Based on this, core features are extracted through error analysis and visual information analysis, and these features are used as input to the constructed multi-input multi-output classification model, ultimately achieving the prediction of slippage and coasting.

[0132] 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 13As shown, this 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 operation of the operating system and computer programs stored in the non-volatile storage media. The database stores train operation data. The I / O interfaces are used for exchanging information 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 detecting idling and coasting of a high-speed train.

[0133] Those skilled in the art will understand that Figure 13 The structures shown are merely block diagrams of some structures related to the present application and do 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 shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0134] 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.

[0135] 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.

[0136] 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. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0137] 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.

[0138] 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 detecting idling and coasting of high-speed trains, characterized in that, include: The system acquires the train's current absolute position, as well as the time series of its travel distance, track image, and speed over a past interval. For the past interval, the starting position is the location of the first transponder, and the ending position is the location of the second transponder. The current absolute position is provided by the second transponder. The first and second transponders are adjacent to each other. The control state of the train is determined based on the time sequence of its travel speed; the control state includes traction state, braking state, and constant speed state. The traction and braking errors of the train are determined based on the travel distance time series, the current absolute position, and the train's control status. Based on the time series of track images, a machine learning model is used to predict the dry and wet conditions of the track surface; the dry and wet conditions of the track surface include dry track, wet track, and snow track. The adhesion coefficient of the track is determined based on the average speed corresponding to the travel speed time series. Based on the train's traction error, braking error, and the track surface's wet / dry condition and adhesion coefficient, a logistic regression classification model is used to predict the train's operating status over a past interval; the operating status includes normal, idling, and coasting.

2. The high-speed train idle and coasting detection method according to claim 1, characterized in that, The traction and braking errors of the train are determined based on the travel distance time series, the current absolute position, and the train's control status, specifically including: When the train is in traction control mode, the traction error of the train is calculated based on the travel distance time series, and the braking error of the train is set to 0. When the train is in braking mode, the braking error of the train is calculated based on the travel distance time series, and the traction error of the train is set to 0. When the train is in a constant speed control state, both the traction error and braking error of the train are set to 0.

3. The high-speed train idle and coasting detection method according to claim 2, characterized in that, The calculation process for traction error and braking error is as follows: Calculate the difference between any two adjacent items in the travel distance time series, excluding the last item. The average difference is obtained by averaging the individual differences. Calculate the difference between the current absolute position and the second-to-last term in the travel distance time series to obtain the correction difference; When the train is in traction control mode, the traction error is obtained by subtracting the correction error from the average difference. When the train is in braking mode, the braking error is obtained by subtracting the average difference from the correction difference.

4. The high-speed train idle and coasting detection method according to claim 1, characterized in that, Based on the time series of orbital images, a machine learning model is used to predict the wet and dry conditions of the orbital surface, specifically including: For each track image in the time series of track images, a machine learning model is used to predict the corresponding dry and wet state of the track surface. Select the prediction result with the largest number of occurrences from all the prediction results corresponding to the orbit images; When the proportion of the predicted results exceeds a preset threshold, the track surface wetness / dryness state is determined based on the predicted results.

5. The high-speed train idle and coasting detection method according to claim 1, characterized in that, The machine learning model is a YOLOv5 convolutional neural network model; the frame rate of the track image time series is 2 frames per second.

6. The high-speed train idle and coasting detection method according to claim 1, characterized in that, The logistic regression classification model includes a linear transformation part, a logistic mapping part, and a final prediction output part; wherein... The mathematical expression for the linear transformation part is: ； in, Indicates category The corresponding linear output; , These represent normal, idle, and coasting, respectively. For the input feature vector, , Indicates traction error. Indicates braking error. Indicates the dryness or wetness of the track surface. Indicates the adhesion coefficient of the track; the superscript T indicates transpose; Indicates category The corresponding weight vector, , respectively with One-to-one correspondence; Indicates category The corresponding bias term; The mathematical expression for the logical mapping part is: ； in, Represents the input feature vector The corresponding sample belongs to the category The conditional probability; The mathematical expression for the final prediction output is: ； in, This indicates the category of the final predicted output.

7. The high-speed train idle and coasting detection method according to claim 1, characterized in that, The formula for calculating the adhesion coefficient is: ； in, The viscosity coefficient is 1. This represents the average speed corresponding to the time series of driving speeds.

8. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the high-speed train idle and coasting detection method according to any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the high-speed train idle and coasting detection method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the high-speed train idle and coasting detection method as described in any one of claims 1-7.

Images