Methods, devices, equipment, and media for rail vehicle operation control based on onboard edge computing

Abstract

This application discloses a method, device, equipment, and medium for rail vehicle operation control based on onboard edge computing, relating to the field of vehicle operation control technology. The method includes: acquiring pre-collected historical operation-related data of the rail vehicle; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time track surface images; preprocessing the historical operation-related data and dividing it according to track surface conditions to construct a target training dataset; training an initial adhesion coefficient estimation model using the target training dataset to obtain a trained adhesion coefficient estimation model, and then performing trimming and fine-tuning to obtain a target adhesion coefficient estimation model; deploying the target adhesion coefficient estimation model on the onboard edge device of the target rail vehicle, so that the onboard edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the target wheel-rail adhesion coefficient.

Description

Technical Field

[0001] This invention relates to the field of vehicle operation control technology, and in particular to a method, device, equipment and medium for rail vehicle operation control based on onboard edge computing. Background Technology

[0002] During rail vehicle operation, the wheel-rail adhesion coefficient is a key parameter affecting the traction, braking performance, and operational safety of rail vehicles. It provides a reliable basis for the traction / braking control system, thereby improving operational safety and energy efficiency. Therefore, real-time and accurate estimation of the wheel-rail adhesion coefficient is crucial. However, the wheel-rail adhesion coefficient is affected by various dynamic factors, such as rail surface humidity, contamination (oil stains, fallen leaves, etc.), wheel and rail material properties, and operating speed, resulting in strong time-varying and nonlinear characteristics. Traditional adhesion coefficient estimation methods typically rely on simplified models or empirical parameters, resulting in slow response speeds and difficulty in timely tracking rapid changes in the adhesion state, thus lacking real-time performance. While existing intelligent methods such as deep learning possess some adaptive capabilities, they generally suffer from poor model generalization and strong dependence on training data, leading to low prediction accuracy and difficulty meeting the requirements of high-reliability on-board applications. Furthermore, deep learning-based intelligent estimation methods typically have a large number of model parameters and computational complexity, resulting in high inference time and resource consumption, which cannot meet the limitations of on-board edge devices in terms of computing power, storage space, and power consumption. Direct deployment often requires high-performance hardware support, significantly increasing system cost and energy consumption.

[0003] In summary, optimizing the wheel-rail adhesion coefficient estimation method to meet the needs of on-board edge computing while ensuring prediction accuracy, and thus achieving reliable control of wheel-rail vehicles, is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method, apparatus, equipment, and medium for rail vehicle operation control based on onboard edge computing, which can optimize the wheel-rail adhesion coefficient estimation method to meet the requirements of onboard edge computing while ensuring prediction accuracy, thereby achieving reliable control of wheel-rail vehicles. The specific solution is as follows: Firstly, this application provides a rail vehicle operation control method based on onboard edge computing, including: Acquire pre-collected historical operation-related data of the rail vehicle; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time images of the track surface; The historical operation-related data is preprocessed and divided according to the orbital surface status to construct the target training dataset; The initial adhesion coefficient estimation model is trained using the target training dataset to obtain a trained adhesion coefficient estimation model, and the trained adhesion coefficient estimation model is then trimmed and fine-tuned to obtain the target adhesion coefficient estimation model. The target adhesion coefficient estimation model is deployed on the on-board edge device of the target rail vehicle so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

[0005] Optionally, the preprocessing of the historical operation-related data and the partitioning based on the orbital plane state to construct the target training dataset includes: The historical operation-related data is deduplicated to obtain deduplicated historical operation-related data; The deduplicated historical running data is subjected to time-series alignment processing to obtain the corresponding time-series running data. The corresponding track surface state is identified using the real-time track surface image in the time-series operation related data, and the time-series operation related data is divided according to the track surface state to obtain the divided time-series operation related data. The actual adhesion coefficients of the time-series related data after the partitioning are determined respectively, and the labels of the time-series related data after the partitioning are marked by the actual adhesion coefficients to obtain the target training dataset.

[0006] Optionally, the step of performing time-series alignment processing on the deduplicated historical runtime-related data to obtain corresponding time-series runtime-related data includes: The sampling frequency of the deduplicated historical running data is determined, and the sampling frequency is compared with a preset benchmark sampling frequency to obtain the corresponding comparison results; Based on the comparison results, the deduplicated historical operation-related data is subjected to time-series alignment processing to obtain the corresponding time-series operation-related data; Accordingly, the step of performing time-series alignment processing on the deduplicated historical operation-related data based on the comparison results includes: If the comparison result indicates that the sampling frequency is lower than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are interpolated and bound to the corresponding real-time track surface image. If the comparison result indicates that the sampling frequency is higher than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are resampled and bound to the corresponding real-time track surface image.

[0007] Optionally, the initial adhesion coefficient estimation model includes a first convolutional neural network, a second convolutional neural network, and a feature fusion fully connected network; the first convolutional neural network is used to perform image recognition on the real-time image of the rail surface and extract the corresponding rail surface state features; the second convolutional neural network is used to extract the time series features of the vehicle operating parameters and the environmental monitoring parameters; the feature fusion fully connected network is used to fuse the rail surface state features and the time series features to predict and estimate the wheel-rail adhesion coefficient based on the obtained fused features.

[0008] Optionally, training the initial adhesion coefficient estimation model using the target training dataset to obtain a trained adhesion coefficient estimation model includes: The target training dataset is input into the initial adhesion coefficient estimation model. Backpropagation is performed based on the preset objective function and through the stochastic gradient descent algorithm to update the model parameters of the initial adhesion coefficient estimation model, thereby obtaining the trained adhesion coefficient estimation model. The preset objective function is used to calculate the deviation between the actual adhesion coefficient and the predicted adhesion coefficient; the predicted adhesion coefficient is the adhesion coefficient predicted by the initial adhesion coefficient estimation model.

[0009] Optionally, the step of pruning and fine-tuning the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model includes: The target importance index of each channel in each layer of the trained adhesion coefficient estimation model is determined by using a preset importance index calculation formula. The channels are sorted in descending order based on the target importance index to obtain the corresponding sorting results; By using a predetermined channel importance threshold and combining it with the ranking results, the channels in the trained adhesion coefficient estimation model are pruned to obtain the target adhesion coefficient estimation model.

[0010] Optionally, the process of determining the target importance index includes: Determine the feature height and feature width of the target channel, and determine the feature values ​​of each row and column of the target channel; the target channel is the channel in the trained adhesion coefficient estimation model; Based on the feature height, the feature width, and the feature value, and using the preset importance index calculation formula, the target importance index of the target channel is determined.

[0011] Secondly, this application provides a rail vehicle operation control device based on on-board edge computing, comprising: The data acquisition module is used to acquire pre-collected historical operation-related data of the rail vehicle; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time images of the track surface; The data partitioning module is used to preprocess the historical operation-related data and partition it according to the track surface status to construct the target training dataset. The model fine-tuning module is used to train the initial adhesion coefficient estimation model using the target training dataset to obtain the trained adhesion coefficient estimation model, and to trim and fine-tune the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model. The model deployment module is used to deploy the target adhesion coefficient estimation model on the on-board edge device of the target rail vehicle, so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

[0012] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the aforementioned rail vehicle operation control method based on on-board edge computing.

[0013] Fourthly, this application provides a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned rail vehicle operation control method based on on-board edge computing.

[0014] In this application, historical operation-related data of the rail vehicle is acquired in advance; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time images of the track surface; the historical operation-related data is preprocessed and divided according to the track surface state to construct a target training dataset; the initial adhesion coefficient estimation model is trained using the target training dataset to obtain a trained adhesion coefficient estimation model, and the trained adhesion coefficient estimation model is trimmed and fine-tuned to obtain a target adhesion coefficient estimation model; the target adhesion coefficient estimation model is deployed on the on-board edge device of the target rail vehicle so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and use the obtained target wheel-rail adhesion coefficient to perform operation control of the target wheel-rail vehicle. As can be seen from the above, this application first acquires pre-collected historical operational data of rail vehicles, including vehicle operating parameters, environmental monitoring parameters, and real-time track surface images. This data is preprocessed and divided according to track surface conditions to construct a target training dataset. Then, this dataset is used to train the initial adhesion coefficient estimation model. The trained model is then trimmed and fine-tuned to form the target adhesion coefficient estimation model. Finally, this model is deployed to the onboard edge device of the target rail vehicle to achieve real-time estimation of the wheel-rail adhesion coefficient. The obtained target wheel-rail adhesion coefficient is then used to control the operation of the target wheel-rail vehicle. In this way, through the above process of this application, based on multi-source heterogeneous historical vehicle data, training data that fits the actual operating scenario is constructed through standardized preprocessing and classification. After model training and lightweight optimization, the model's operating efficiency is improved while ensuring estimation accuracy, adapting to the deployment requirements of onboard edge devices. This provides accurate real-time wheel-rail adhesion state assessment for rail vehicles, effectively supporting safe vehicle operation and intelligent control. It optimizes the wheel-rail adhesion coefficient estimation method to meet the needs of onboard edge computing while ensuring prediction accuracy, thereby achieving reliable control of wheel-rail vehicles. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of a rail vehicle operation control method based on on-board edge computing disclosed in this application; Figure 2 This is a flowchart illustrating a rail vehicle operation control method based on on-board edge computing disclosed in this application; Figure 3 This is a structural framework diagram of an initial adhesion coefficient estimation model disclosed in this application; Figure 4 This is a schematic diagram of a model channel clipping disclosed in this application; Figure 5 This is a schematic diagram of the structure of a rail vehicle operation control device based on on-board edge computing disclosed in this application; Figure 6 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Traditional adhesion coefficient estimation methods typically rely on simplified models or empirical parameters, resulting in slow response times and difficulty in tracking rapid changes in the adhesion state, thus lacking real-time performance. While existing intelligent methods such as deep learning possess some adaptive capabilities, they generally suffer from poor model generalization and strong dependence on training data, leading to low prediction accuracy and failing to meet the requirements of high-reliability automotive applications. Furthermore, deep learning-based intelligent estimation methods typically have a large number of model parameters and computational complexity, resulting in high inference time and resource consumption, making it difficult to meet the limitations of automotive edge devices in terms of computing power, storage space, and power consumption. Direct deployment often requires high-performance hardware support, significantly increasing system cost and energy consumption.

[0019] To overcome the aforementioned technical problems, this application provides a rail vehicle operation control method based on onboard edge computing, which can optimize the wheel-rail adhesion coefficient estimation method to meet the needs of onboard edge computing while ensuring prediction accuracy, thereby achieving reliable control of wheel-rail vehicles.

[0020] See Figure 1 As shown in the figure, an embodiment of the present invention discloses a rail vehicle operation control method based on onboard edge computing, including: Step S11: Obtain pre-collected historical operation-related data of the rail vehicle; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time images of the track surface.

[0021] In this embodiment, historical operational data, including vehicle operating parameters (such as speed and axle load), environmental monitoring parameters (such as humidity, temperature, and pollutants), and real-time images of the track surface, are acquired in advance through time-series data collection. The historical operational data includes data such as vehicle speed, acceleration, axle load, wheelset speed, ambient temperature, and ambient humidity, which are stored in the onboard system for easy real-time retrieval.

[0022] It should be pointed out that, as Figure 2 The diagram shows a flowchart of a rail vehicle operation control method based on onboard edge computing provided in this application. Specifically, it includes: Step 1, Data Acquisition (S100): The acquired data includes monitoring parameters of key subsystems of the rail vehicle. Step 2, Data Alignment (S200): The acquired real-time data is aligned and preprocessed to form a database. Step 3, Training Set Construction (S300): Based on track surface conditions, and combined with rail vehicle operating parameters and environmental parameters, Steps 1 and 2 are repeated to form a large-scale training dataset for the rail vehicle. Step 4, Multimodal Estimation Model Construction (S400): A multimodal estimation model (intelligent wheel-rail adhesion coefficient prediction model) based on a convolutional neural network is constructed. Step 5, Model Training (S500): The dataset prepared in Step 3 is fed into the model constructed in Step 4, and the model parameters are updated using stochastic gradient descent algorithm for backpropagation. Step 6, Model Pruning (S600): The model trained in Step 5 is dynamically pruned and fine-tuned. Step 7, Model Deployment (S700): Deploy the model from Step 6 to the edge equipment for online real-time wheel-rail adhesion coefficient prediction. In this way, this embodiment integrates multi-dimensional and multi-source historical information and fuses multi-source sensor information, providing fundamental data support for subsequent model training and data analysis, effectively improving the accuracy of wheel-rail state identification and adhesion coefficient estimation.

[0023] Step S12: Preprocess the historical operation-related data and divide it according to the track surface status to construct the target training dataset.

[0024] In this embodiment, the historical operation-related data is preprocessed by cropping, interpolation, alignment, and deduplication, and then classified according to the track surface status to construct the target training dataset.

[0025] Specifically, the historical operation-related data is deduplicated to obtain deduplicated historical operation-related data; the deduplicated historical operation-related data is then time-aligned to obtain corresponding time-series operation-related data; the real-time track surface images in the time-series operation-related data are used to identify the corresponding track surface state, and the time-series operation-related data is divided according to the track surface state to obtain divided time-series operation-related data; the actual adhesion coefficients of the divided time-series operation-related data are determined respectively, and the divided time-series operation-related data are labeled using the actual adhesion coefficients to obtain the target training dataset. That is, the historical operation-related data is sequentially deduplicated and time-series aligned to obtain standardized time-series operation-related data. Specifically, the deduplicated historical operation-related data includes data on rail vehicle speed, acceleration, axle load, wheelset speed, ambient temperature, and ambient humidity. Each collected data point is time-series aligned to obtain six time-series data sequences from the past six minutes. The time-series data sequences of rail vehicle speed, acceleration, axle load, wheelset speed, ambient temperature, and ambient humidity are arranged into a square matrix, where the first row contains six speed time-series data, the second row contains six acceleration time-series data, and so on. Then, the rail surface state is identified through the real-time images of the rail surface in the time-series operation-related data, and the time-series operation-related data is divided into different rail surface states such as dry, rain, snow, ice, oil, and leaves according to the rail surface state, resulting in divided time-series operation-related data. Labels are then generated based on the detected actual adhesion coefficients corresponding to each segment of the divided time-series operation-related data, and finally, a target training dataset of adhesion coefficients for different rail surface states is constructed.

[0026] It should be noted that the processing flow for time-series alignment of the deduplicated historical operation-related data is as follows: The sampling frequency of the deduplicated historical operation-related data is determined, and the sampling frequency is compared with a preset benchmark sampling frequency to obtain a corresponding comparison result; based on the comparison result, time-series alignment processing is performed on the deduplicated historical operation-related data to obtain the corresponding time-series operation-related data. That is, the sampling frequency of the deduplicated historical operation-related data with different sampling frequencies in multimodal processing is first determined, and then compared with a preset benchmark sampling frequency, such as once per minute. Based on the obtained comparison result, time-series alignment processing is performed on the deduplicated historical operation-related data to obtain the time-series operation-related data. It should be further noted that the processing flow for time-series alignment of the deduplicated historical operation-related data based on the comparison result is as follows: If the comparison result indicates that the sampling frequency is lower than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are interpolated and bound to the corresponding real-time track surface image; if the comparison result indicates that the sampling frequency is higher than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are resampled and bound to the corresponding real-time track surface image. That is, when the comparison result indicates that the sampling frequency is lower than the preset benchmark sampling frequency, the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are interpolated; when it is higher than the preset benchmark sampling frequency, resampling is performed, thereby achieving real-time data alignment. Simultaneously, after processing, all data is bound to the corresponding real-time track surface image to obtain image signals and time-series signals. In this way, data quality is improved through data cleaning, standardization, and annotation processes. Combined with track surface state classification and adhesion coefficient annotation based on image recognition, the dataset is made to better reflect the actual operating scenarios of rail vehicles, providing a reliable and high-quality data foundation for subsequent adhesion coefficient estimation model training. By unifying the time series standards of multi-source data, the accuracy of vehicle operating parameters, environmental parameters, and track surface images is ensured to be accurately matched in the time dimension, thereby improving the accuracy and stability of subsequent data processing and model training.

[0027] Step S13: Train the initial adhesion coefficient estimation model using the target training dataset to obtain the trained adhesion coefficient estimation model, and then trim and fine-tune the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model.

[0028] In this embodiment, an initial stickiness coefficient estimation model is trained using the target training dataset to obtain a trained stickiness coefficient estimation model. Then, dynamic channel pruning and fine-tuning are performed on the model, and the network structure is sparsified and parameters are compressed, which significantly reduces the number of model parameters and computational complexity, thereby reducing the model size and finally obtaining a lightweight and high-precision target stickiness coefficient estimation model.

[0029] It should be noted that the initial adhesion coefficient estimation model includes a first convolutional neural network, a second convolutional neural network, and a feature fusion fully connected network. The first convolutional neural network is used to perform image recognition on the real-time track surface image and extract the corresponding track surface state features. The second convolutional neural network is used to extract the time-series features of the vehicle operating parameters and the environmental monitoring parameters. The feature fusion fully connected network is used to fuse the track surface state features and the time-series features to predict and estimate the wheel-rail adhesion coefficient based on the obtained fused features. That is, the initial adhesion coefficient estimation model is a multimodal estimation model based on a convolutional neural network (CNN), consisting of a first convolutional neural network, a second convolutional neural network, and a feature fusion fully connected network, as shown below. Figure 3 The diagram shows the structural framework of an initial adhesion coefficient estimation model provided in this application. The first convolutional neural network (CNN) is an image feature extraction CNN network for image recognition, responsible for recognizing real-time images of the rail surface and extracting rail surface state features to determine the rail surface state. The second convolutional neural network (CNN) is a time-series feature extraction CNN network for sequential data of rail vehicle speed, acceleration, axle load, wheelset speed, ambient temperature, and ambient humidity, responsible for extracting the time-series features of the vehicle operating parameters and the environmental monitoring parameters, achieving feature extraction from multimodal time series. The feature fusion fully connected network is responsible for fusing the two types of features for predicting and estimating the wheel-rail adhesion coefficient. Table 1 shows the structural parameters of the initial adhesion coefficient estimation model provided in this application.

[0030] Table 1 Structural parameters of the initial adhesion coefficient estimation model

[0031] Where Conv2d represents 2D convolution operation, C represents the number of channels, K represents the kernel size, BN is batch normalization, ReLU is the activation function, MP represents max pooling, GAP represents global average pooling, and Linear represents fully connected operation.

[0032] Understandably, the process of training the initial stickiness coefficient estimation model using the target training dataset is as follows: The target training dataset is input into the initial stickiness coefficient estimation model. Backpropagation is performed based on a preset objective function and a stochastic gradient descent algorithm to update the model parameters of the initial stickiness coefficient estimation model, resulting in a trained stickiness coefficient estimation model. The preset objective function is used to calculate the deviation between the actual stickiness coefficient and the predicted stickiness coefficient. The predicted stickiness coefficient is the stickiness coefficient predicted by the initial stickiness coefficient estimation model. That is, the target training dataset is input into the initial stickiness coefficient estimation model. Guided by the preset objective function that calculates the deviation between the actual stickiness coefficient and the predicted stickiness coefficient, the model parameters are updated through backpropagation using a stochastic gradient descent algorithm and preset training parameters, resulting in a trained stickiness coefficient estimation model. The preset training parameters include an optimizer, a preset loss function, a learning rate, a batch size of 96, and 100 epochs. The optimizer can be Adam (Adaptive Moment Estimation, a deep learning optimization algorithm), using a dynamic learning rate, a batch size of 96, and 100 epochs. The specific formula for expressing the preset objective function is as follows: ; in, Indicates the actual adhesion coefficient; Indicates the predicted adhesion coefficient; is a hyperparameter used for weight control; L is the preset objective function.

[0033] It should be noted that the process of pruning and fine-tuning the post-trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model is as follows: The target importance index of each channel in each layer of the post-trained adhesion coefficient estimation model is determined using a preset importance index calculation formula; the channels are sorted in descending order based on the target importance index to obtain the corresponding sorting results; the channels in the post-trained adhesion coefficient estimation model are pruned using a pre-determined channel importance threshold and in combination with the sorting results to obtain the target adhesion coefficient estimation model. The channel importance threshold t is determined based on the channel pruning ratio r, which is determined experimentally. That is, the target importance index (Important index, II) of each channel in each layer of the post-trained adhesion coefficient estimation model is determined using a preset importance index calculation formula; after sorting the channels in descending order according to this index, channels in the post-trained adhesion coefficient estimation model with II lower than t are pruned using a pre-determined channel importance threshold t, without pruning related connection layers, such as... Figure 4 The diagram shown is a schematic of a model channel clipping method provided in this application, which ultimately yields a target adhesion coefficient estimation model.

[0034] It should be further noted that the process of determining the target importance index is as follows: The feature height and feature width of the target channel are determined, and the feature values ​​of each row and column of the target channel are determined; the target channel is the channel in the post-trained adhesion coefficient estimation model; based on the feature height, feature width, and feature values, and using the preset importance index calculation formula, the target importance index of the target channel is determined. That is, first, the feature height, feature width, and feature values ​​of each row and column of the target channel in the post-trained adhesion coefficient estimation model are determined, and then, combined with this dimensional information, the target importance index of the channel is calculated using the preset importance index calculation formula. The specific expression of the preset importance index calculation formula is as follows: ; in, Characterizing the first l The first layer c One channel II; h and w These represent the height and width of the channel feature, respectively; that is, the feature height and the feature width. Indicates the first l Layer c The first channel i Line number j The eigenvalues ​​of the column. Thus, in this embodiment, after training the initial adhesion coefficient estimation model using the target training dataset, it is cropped and fine-tuned to balance model estimation accuracy and operational efficiency, providing reliable support for subsequent vehicle-mounted deployment. The initial adhesion coefficient estimation model extracts features of different dimensions through sub-networks, fully mining the visual information of the track surface image and accurately capturing the temporal patterns of operating and environmental parameters. Feature fusion then achieves the complementarity and integration of multi-dimensional information, improving the comprehensiveness and accuracy of adhesion coefficient estimation. Using the objective function of accurately quantifying prediction deviation as the core, combined with the stochastic gradient descent algorithm, the model parameters are efficiently optimized, enabling rapid... This method rapidly narrows the gap between predicted and actual values, significantly improving the model's prediction accuracy for wheel-rail adhesion coefficient. It can handle low adhesion conditions such as rain, snow, oil, and fallen leaves, overcoming the limitations of traditional fixed lookup table methods or empirical formulas. By comprehensively quantifying the importance of channels from both spatial and numerical dimensions of channel features, core channels are selected, further reducing the number of model parameters and computational load. This ensures the model meets real-time requirements, retains the core prediction capabilities of the model while eliminating redundant parameters, effectively achieving model lightweighting, reducing computational complexity and storage overhead, adapting to the resource constraints of onboard edge devices, and ensuring that the accuracy of adhesion coefficient estimation is not significantly affected.

[0035] Step S14: Deploy the target adhesion coefficient estimation model on the on-board edge device of the target rail vehicle so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

[0036] In this embodiment, the lightweight target adhesion coefficient estimation model is deployed to the onboard edge device of the target rail vehicle. This allows the onboard edge device to estimate the wheel-rail adhesion coefficient of the target rail vehicle online in real time using the model, and to implement operational control, including traction and braking, based on the obtained target wheel-rail adhesion coefficient. It is understood that the target adhesion coefficient estimation model in this embodiment can continuously update input data to achieve self-optimization, adapting to the influence of long-term factors such as wheel-rail material wear and seasonal changes, and possesses online learning capabilities. In this way, the lightweight model of this embodiment can be efficiently deployed on resource-constrained onboard edge devices, fully adapting to the resource characteristics of the onboard edge devices, achieving low-latency, high-response real-time online calculation of the adhesion coefficient, providing timely and reliable status feedback for the traction and braking systems of the rail vehicle, and improving operational safety and control performance.

[0037] As can be seen from the above, the embodiments of this application first acquire historical operation-related data of the rail vehicle, including vehicle operation parameters, environmental monitoring parameters, and real-time images of the rail surface, preprocess the data and divide it according to the rail surface state to construct a target training dataset, and then use the dataset to train the initial adhesion coefficient estimation model. After training, the model is trimmed and fine-tuned to form a target adhesion coefficient estimation model. Finally, the model is deployed to the on-board edge device of the target rail vehicle to realize real-time estimation of the wheel-rail adhesion coefficient, and the target wheel-rail adhesion coefficient is used to control the operation of the target wheel-rail vehicle. In this way, through the above process of the embodiments of this application, based on multi-source heterogeneous vehicle historical data, training data that fits the actual operation scenario is constructed through standardized preprocessing and classification. Then, through model training and lightweight optimization, the model running efficiency is improved while ensuring estimation accuracy, adapting to the deployment requirements of on-board edge devices. It can provide accurate wheel-rail adhesion state assessment for rail vehicles in real time, effectively supporting safe operation and intelligent control of vehicles. It can optimize the wheel-rail adhesion coefficient estimation method to meet the needs of on-board edge computing while ensuring prediction accuracy, thereby achieving reliable control of wheel-rail vehicles.

[0038] Accordingly, see Figure 5 As shown in the figure, this application embodiment also provides a rail vehicle operation control device based on on-board edge computing, including: Data acquisition module 11 is used to acquire pre-collected historical operation-related data of the rail vehicle; the historical operation-related data includes vehicle operation parameters, environmental monitoring parameters, and real-time images of the track surface; The data partitioning module 12 is used to preprocess the historical operation-related data and partition it according to the track surface status to construct the target training dataset. The model fine-tuning module 13 is used to train the initial adhesion coefficient estimation model using the target training dataset to obtain the trained adhesion coefficient estimation model, and to trim and fine-tune the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model. The model deployment module 14 is used to deploy the target adhesion coefficient estimation model on the on-board edge device of the target rail vehicle, so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

[0039] In some specific embodiments, the data partitioning module 12 may specifically include: The data deduplication unit is used to perform deduplication operations on the historical operation-related data to obtain deduplicated historical operation-related data; The timing alignment submodule is used to perform timing alignment processing on the deduplicated historical running data to obtain the corresponding timing running data. The data partitioning unit is used to identify the corresponding track surface state using the real-time track surface image in the time-series operation-related data, and to partition the time-series operation-related data according to the track surface state to obtain partitioned time-series operation-related data. The coefficient determination unit is used to determine the actual adhesion coefficients of the partitioned time-series running related data, so as to label the partitioned time-series running related data using the actual adhesion coefficients to obtain the target training dataset.

[0040] In some specific implementations, the timing alignment submodule may specifically include: The frequency comparison unit is used to determine the sampling frequency of the deduplicated historical running data and compare the sampling frequency with a preset benchmark sampling frequency to obtain the corresponding comparison result. The timing alignment unit is used to perform timing alignment processing on the deduplicated historical running-related data according to the comparison result, so as to obtain the corresponding timing running-related data. Accordingly, the timing alignment unit is specifically used to: if the comparison result indicates that the sampling frequency is lower than the preset benchmark sampling frequency, then interpolate the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data and bind them with the corresponding real-time track surface image; if the comparison result indicates that the sampling frequency is higher than the preset benchmark sampling frequency, then resample the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data and bind them with the corresponding real-time track surface image.

[0041] In some specific embodiments, the initial adhesion coefficient estimation model includes a first convolutional neural network, a second convolutional neural network, and a feature fusion fully connected network; the first convolutional neural network is used to perform image recognition on the real-time image of the rail surface and extract the corresponding rail surface state features; the second convolutional neural network is used to extract the time series features of the vehicle operating parameters and the environmental monitoring parameters; the feature fusion fully connected network is used to fuse the rail surface state features and the time series features to predict and estimate the wheel-rail adhesion coefficient based on the obtained fused features.

[0042] In some specific embodiments, the model fine-tuning module 13 may specifically include: The backpropagation unit is used to input the target training dataset into the initial adhesion coefficient estimation model, and perform backpropagation based on a preset objective function and through a stochastic gradient descent algorithm to update the model parameters of the initial adhesion coefficient estimation model, thereby obtaining the trained adhesion coefficient estimation model. The preset objective function is used to calculate the deviation between the actual adhesion coefficient and the predicted adhesion coefficient; the predicted adhesion coefficient is the adhesion coefficient predicted by the initial adhesion coefficient estimation model.

[0043] In some specific embodiments, the model fine-tuning module 13 may specifically include: The index determination submodule is used to determine the target importance index of each channel in each layer of the trained adhesion coefficient estimation model by using a preset importance index calculation formula. The channel sorting unit is used to sort each of the channels in descending order based on the target importance index to obtain the corresponding sorting result; The channel pruning unit is used to prune the channels in the trained adhesion coefficient estimation model by using a pre-determined channel importance threshold and combining the ranking results, so as to obtain the target adhesion coefficient estimation model.

[0044] In some specific implementations, the index determination submodule may specifically include: The feature value determination unit is used to determine the feature height and feature width of the target channel, and to determine the feature value of each row and column of the target channel; the target channel is the channel in the trained adhesion coefficient estimation model; An index determination unit is used to determine the target importance index of the target channel based on the feature height, the feature width, and the feature value, and using the preset importance index calculation formula.

[0045] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application. The electronic device 20 may specifically include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the rail vehicle operation control method based on on-board edge computing disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.

[0046] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0047] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0048] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the on-board edge computing-based rail vehicle operation control method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0049] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned disclosed rail vehicle operation control method based on on-board edge computing. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0050] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0051] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0052] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0053] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] The technical solutions provided in this application have been described in detail above. Specific examples have been used 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. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for controlling the operation of rail vehicles based on onboard edge computing, characterized in that, include: Acquire pre-collected historical operational data of rail vehicles; The historical operational data includes vehicle operating parameters, environmental monitoring parameters, and real-time images of the track surface. The historical operation-related data is preprocessed and divided according to the orbital surface status to construct the target training dataset; The initial adhesion coefficient estimation model is trained using the target training dataset to obtain a trained adhesion coefficient estimation model, and the trained adhesion coefficient estimation model is then trimmed and fine-tuned to obtain the target adhesion coefficient estimation model. The target adhesion coefficient estimation model is deployed on the on-board edge device of the target rail vehicle so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

2. The rail vehicle operation control method based on on-board edge computing according to claim 1, characterized in that, The preprocessing of the historical operational data and its partitioning based on orbital plane status to construct the target training dataset includes: The historical operation-related data is deduplicated to obtain deduplicated historical operation-related data; The deduplicated historical running data is subjected to time-series alignment processing to obtain the corresponding time-series running data. The corresponding track surface state is identified using the real-time track surface image in the time-series operation related data, and the time-series operation related data is divided according to the track surface state to obtain the divided time-series operation related data. The actual adhesion coefficients of the time-series related data after the partitioning are determined respectively, and the labels of the time-series related data after the partitioning are marked by the actual adhesion coefficients to obtain the target training dataset.

3. The rail vehicle operation control method based on on-board edge computing according to claim 2, characterized in that, The step of performing time-series alignment processing on the deduplicated historical runtime-related data to obtain corresponding time-series runtime-related data includes: The sampling frequency of the deduplicated historical running data is determined, and the sampling frequency is compared with a preset benchmark sampling frequency to obtain the corresponding comparison results; Based on the comparison results, the deduplicated historical operation-related data is subjected to time-series alignment processing to obtain the corresponding time-series operation-related data; Accordingly, the step of performing time-series alignment processing on the deduplicated historical operation-related data based on the comparison results includes: If the comparison result indicates that the sampling frequency is lower than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are interpolated and bound to the corresponding real-time track surface image. If the comparison result indicates that the sampling frequency is higher than the preset benchmark sampling frequency, then the corresponding vehicle operation parameters and environmental monitoring parameters in the deduplicated historical operation-related data are resampled and bound to the corresponding real-time track surface image.

4. The rail vehicle operation control method based on on-board edge computing according to claim 1, characterized in that, The initial adhesion coefficient estimation model includes a first convolutional neural network, a second convolutional neural network, and a feature fusion fully connected network. The first convolutional neural network is used to perform image recognition on the real-time track surface image and extract the corresponding track surface state features. The second convolutional neural network is used to extract the time series features of the vehicle operating parameters and the environmental monitoring parameters. The feature fusion fully connected network is used to fuse the track surface state features and the time series features to predict and estimate the wheel-rail adhesion coefficient based on the obtained fused features.

5. The rail vehicle operation control method based on on-board edge computing according to claim 1, characterized in that, The step of training the initial adhesion coefficient estimation model using the target training dataset to obtain the trained adhesion coefficient estimation model includes: The target training dataset is input into the initial adhesion coefficient estimation model. Backpropagation is performed based on the preset objective function and through the stochastic gradient descent algorithm to update the model parameters of the initial adhesion coefficient estimation model, thereby obtaining the trained adhesion coefficient estimation model. The preset objective function is used to calculate the deviation between the actual adhesion coefficient and the predicted adhesion coefficient; the predicted adhesion coefficient is the adhesion coefficient predicted by the initial adhesion coefficient estimation model.

6. The rail vehicle operation control method based on on-board edge computing according to any one of claims 1 to 5, characterized in that, The step of pruning and fine-tuning the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model includes: The target importance index of each channel in each layer of the trained adhesion coefficient estimation model is determined by using a preset importance index calculation formula. The channels are sorted in descending order based on the target importance index to obtain the corresponding sorting results; By using a predetermined channel importance threshold and combining it with the ranking results, the channels in the trained adhesion coefficient estimation model are pruned to obtain the target adhesion coefficient estimation model.

7. The rail vehicle operation control method based on on-board edge computing according to claim 6, characterized in that, The process of determining the target importance index includes: Determine the feature height and feature width of the target channel, and determine the feature values ​​of each row and column of the target channel; the target channel is the channel in the trained adhesion coefficient estimation model; Based on the feature height, the feature width, and the feature value, and using the preset importance index calculation formula, the target importance index of the target channel is determined.

8. A rail vehicle operation control device based on onboard edge computing, characterized in that, include: The data acquisition module is used to acquire pre-collected historical operation-related data of rail vehicles; The historical operational data includes vehicle operating parameters, environmental monitoring parameters, and real-time images of the track surface. The data partitioning module is used to preprocess the historical operation-related data and partition it according to the track surface status to construct the target training dataset. The model fine-tuning module is used to train the initial adhesion coefficient estimation model using the target training dataset to obtain the trained adhesion coefficient estimation model, and to trim and fine-tune the trained adhesion coefficient estimation model to obtain the target adhesion coefficient estimation model. The model deployment module is used to deploy the target adhesion coefficient estimation model on the on-board edge device of the target rail vehicle, so that the on-board edge device can use the target adhesion coefficient estimation model to estimate the wheel-rail adhesion coefficient of the target rail vehicle, and perform operation control of the target wheel-rail vehicle based on the obtained target wheel-rail adhesion coefficient.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the rail vehicle operation control method based on on-board edge computing as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store computer programs; wherein, when the computer programs are executed by a processor, they implement the rail vehicle operation control method based on on-board edge computing as described in any one of claims 1 to 7.

Images