Railway locomotive control method and system, electronic device

Abstract

The present application relates to the technical field of rail vehicle control scheme design, and particularly relates to a rail vehicle control method and system and an electronic device. The method comprises: capturing a visual image of a target area by a visual camera arranged in front of the rail vehicle; collecting point cloud information of the target area by a laser radar arranged in front of the rail vehicle; processing the visual image and the point cloud information of the target area respectively, identifying the opening and closing state of a turnout in the target area, and obtaining state information of a traffic light; if the opening and closing state of the turnout and the state information of the traffic light both meet the passing condition, controlling the rail vehicle to pass through the turnout; if the opening and closing state of the turnout or the state information of the traffic light does not meet the passing condition, outputting alarm information about the opening and closing state of the turnout or the state information of the traffic light not meeting the passing condition. The deep learning model combining the features of vision and laser radar realizes higher-precision turnout identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of track locomotive control scheme design, and particularly relates to a track locomotive control method and system and an electronic device. BACKGROUND

[0002] In the track transportation industry, the locomotive passing through the turnout needs to be confirmed by the driver, and in some cases, the confirmation needs to be made by two drivers simultaneously. That is, the driver manually controls the acceleration, deceleration, parking, turnout pulling, air door opening and other actions of the track locomotive from the driver's cab at the front and rear ends of the locomotive, resulting in a large number of workers in the auxiliary transportation system, high labor intensity of the driver and the possibility of unsafe accidents during transportation driving.

[0003] Therefore, the prior art still needs to be further developed. SUMMARY

[0004] The present application aims to overcome the above technical deficiencies and provide a track locomotive control method and system and an electronic device to solve the problems existing in the prior art.

[0005] To achieve the above technical purpose, according to the first aspect of the present application, the present application provides a track locomotive control method, comprising:

[0006] S100, capturing a visual image of a target area by using a visual camera arranged in front of the track locomotive;

[0007] S200, processing the visual image and the point cloud information of the target area respectively, identifying the opening and closing state of the turnout of the target area, and obtaining the state information of the traffic light;

[0008] S300, if the opening and closing state of the turnout and the state information of the traffic light both meet the passing condition, controlling the track locomotive to pass through the turnout; if the opening and closing state of the turnout or the state information of the traffic light does not meet the passing condition, outputting an alarm information about the opening and closing state of the turnout or the state information of the traffic light not meeting the passing condition.

[0009] S300, if the opening and closing state of the turnout and the state information of the traffic light both meet the passing condition, controlling the track locomotive to pass through the turnout; if the opening and closing state of the turnout or the state information of the traffic light does not meet the passing condition, outputting an alarm information about the opening and closing state of the turnout or the state information of the traffic light not meeting the passing condition.

[0010] Specifically, the processing of the image and the point cloud information of the target area respectively, identifying the opening and closing state of the turnout of the target area and the state of the traffic light, comprises:

[0011] applying a 3x3 convolution kernel K to perform convolution operation on the visual image:

[0012] Xconv = Conv(I, K);

[0013] Among them, the convolution kernel K is a weight matrix, and the local features of the turnout in the image are extracted through multi-layer convolution. Xconv represents the result of convolution, Conv represents the convolution function, and I represents the visual image.

[0014] Residual connections are used to solve the gradient vanishing problem and optimize network training:

[0015] Yrcs=Xconv+F(Xconv,W);

[0016] Where F represents the combination of convolution operation and activation function, W is the weight matrix; Xconv represents the result of convolution, and Yrcs represents the result of residual operation;

[0017] The high-level features of the turnout image are extracted through multiple convolutional layers, and the feature map F is finally output. image , represents the spatial and structural information of the turnout in the image:

[0018] F image =ResNet(I);

[0019] Among them, ResNet represents the ResNet network, and I represents the visual image.

[0020] Specifically, the processing of the image and point cloud information of the target area respectively to identify the switch opening and closing status and traffic light status of the target area includes:

[0021] Input point cloud data P={p1,p2,…,p n}∈R n×3 , where n is the number of points in the point cloud, p i =(C i ,y i ,z i ) is the three-dimensional coordinate of the i-th point in the point cloud;

[0022] Each point in the point cloud is converted into a high-dimensional feature vector through a multi-layer perceptron network:

[0023] f i =MLP(p i ,W);

[0024] Among them, MLP is a multi-layer perceptron, W is a weight matrix;

[0025] PointNet uses the maximum pooling operation to aggregate the features of all points to obtain a global feature representation:

[0026] F pointcloud =max(f i );

[0027] This operation performs maximum pooling on the features of all points to obtain the global descriptor F pointcloud , which represents the characteristics of the turnout structure in the point cloud.

[0028] Specifically, the method further includes:

[0029] Using the weighted fusion method, the features and feature maps of the turnout structure in the point cloud are learned by machine learning. image The fusion weights of the turnout features in the model are used to automatically adjust their respective influences.

[0030] Specifically, the method further includes: confidence calculation: mapping the fused feature vector to a probability distribution through a fully connected layer or a support vector machine (SVM) to represent the possibility of different turnout states:

[0031] =Softmax(W fc F fused +b);

[0032] in, is the state probability distribution of the output, W fc and b are the weight and bias of the fully connected layer respectively; the probability of each switch state is obtained through the softmax function, and the switch state includes normal, abnormal, and alarm;

[0033] By confidence ,Dynamically adjust the fusion weight α, the dynamic adjustment formula is as follows:

[0034] a new =a old +β( max -0.5);

[0035] in, max is the maximum confidence value, β is the adjustment step size, a old is the weight before adjustment, a new is the weight after adjustment.

[0036] Specifically, the weighted fusion method is used to learn the features and feature maps of the turnout structure in the machine learning point cloud. image The fusion weights of the turnout features in the model are automatically adjusted to adjust their respective influences, including:

[0037] Let the visual feature be F image And the point cloud feature is F pointcloud ;

[0038] Concatenate these two eigenvectors into a comprehensive eigenvector F fused :

[0039] F fused =αF image +(1-α)F pointcloud ;

[0040] Among them, α∈[0,1] is the fusion weight, which represents the contribution ratio of visual features and point cloud features;

[0041] Based on the vector F after feature fusion fused Make turnout status judgment and calculate the confidence of the recognition result;

[0042] The rule engine makes the final switch status judgment based on the fused features and confidence level. The decision rule formula is as follows:

[0043] State=RuleEngine(F fused , );

[0044] RuleEngine function outputs the switch state according to the features and confidence. State represents the switch state. F fused represents the comprehensive feature vector, is the state probability distribution of the output.

[0045] Specifically, the method further includes:

[0046] The support vector machine (SVM) is used to judge the turnout status. The feature vector F is trained on the model. fused Mapped to the turnout state, the decision formula of SVM is:

[0047] y=sign(w T F fused +b);

[0048] Where W is the weight matrix, b is the bias, sign is the sign function, y represents the final state of the turnout, and T represents the transpose.

[0049] Specifically, the method further includes:

[0050] During system operation, as more turnout status data is collected, we can perform incremental learning. Each time a new data point is obtained, the system will update the model parameters based on this new data point. The incremental update formula is as follows:

[0051] θ new =θ old +η·∇L(θ old ,F fused ,y true );

[0052] Among them, θ oldis the parameter of the updated model, θ new is the parameter of the updated model, η is the learning rate, ∇L is the gradient of the loss function, L is the loss function of the model, F fused represents the comprehensive feature vector, y true represents the calculated state of the turnout.

[0053] According to the second aspect of the present application, a rail locomotive control system is provided, comprising:

[0054] An acquisition module includes a visual camera arranged in front of the rail locomotive for shooting visual images of the target area, and a laser radar arranged in front of the rail locomotive for collecting point cloud information of the target area;

[0055] A control module is used to process the visual images and point cloud information of the target area respectively, identify the opening and closing state of the turnout of the target area, and obtain the state information of the traffic light. If the opening and closing state of the turnout and the state information of the traffic light both meet the passing condition, the rail locomotive is controlled to pass through the turnout. If the opening and closing state of the turnout or the state information of the traffic light does not meet the passing condition, alarm information about the opening and closing state of the turnout or the state information of the traffic light not meeting the passing condition is output.

[0056] According to the third aspect of the present application, an electronic device is provided, comprising a memory and a processor, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the processor to implement the above-mentioned rail locomotive control method.

[0057] Advantages:

[0058] 1. Labor cost is saved:

[0059] The need for driver confirmation is reduced, and the consumption of human resources is reduced, especially in the case of double confirmation.

[0060] 2. Improve operational efficiency:

[0061] The speed of the locomotive passing through the turnout is accelerated, and the overall transportation process is optimized.

[0062] 3. Enhance safety:

[0063] Through automatic judgment, the risk of human error is reduced, and the safety of railway transportation is ensured.

[0064] 4. Support for automated operation:

[0065] Provide a basis for unmanned and higher automated train operation, and improve the modernization level of the railway system.

[0066] 5. The deep learning model combining visual and laser radar feature extraction realizes higher precision of turnout identification.

[0067] 6. Dynamically adjust the fusion strategy based on the confidence of the recognition results to improve the adaptability of the system.

[0068] 7. Continuously optimize the model through real-time feedback and online learning to adapt to different environments and changing conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 is a flow chart of a rail vehicle control method provided in a specific embodiment of the present invention;

[0070] Figure 2 It is a schematic diagram of the system composition of a rail locomotive control system provided in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0071] In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is clearly and completely described below in conjunction with the drawings of the present invention. Based on the embodiments in this application, other similar embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of this application. In addition, the directional words mentioned in the following embodiments, such as "up", "down", "left", "right", etc., are only reference to the directions of the drawings. Therefore, the directional words used are used to illustrate rather than limit the invention.

[0072] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0073] See also Figure 1 The present invention provides a rail vehicle control method, comprising:

[0074] S100: Using a visual camera arranged in front of a rail vehicle to capture a visual image of a target area.

[0075] The laser radar installed in front of the rail locomotive is used to collect point cloud information of the target area.

[0076] S200: Process the visual image and point cloud information of the target area respectively, identify the opening and closing status of the turnout in the target area, and obtain the status information of the traffic light.

[0077] As will be appreciated, for visual feature extraction, the present invention utilizes a convolutional neural network (CNN) architecture, specifically a ResNet (Residual Network) as the underlying network for feature extraction. ResNet excels in deep networks and avoids the vanishing gradient problem, making it well-suited for extracting complex turnout structures.

[0078] Specifically, the processing of the image and point cloud information of the target area respectively to identify the switch opening and closing state and the traffic light state of the target area includes:

[0079] input pre-processed image I ∈ R H×W×3 , where H and W are the height and width of the image, and 3 represents the RGB channel.

[0080] The pre-processing of the image comprises:

[0081] The image is denoised using mean filtering or Gaussian filtering or median filtering or bilateral filtering.

[0082] A 3x3 convolution kernel K is applied to the visual image for convolution operation:

[0083] Xconv = Conv(I, K);

[0084] where the convolution kernel K is a trainable weight matrix, and local features of the turnout in the image are extracted through multi-layer convolution, the local features of the turnout including track edge, connection between turnout and track, turnout opening and closing connection point, track gap and track geometry, etc., Xconv represents the result of convolution, Conv represents the convolution function, and I represents the visual image.

[0085] Preferably, the convolution kernel K .

[0086] The gradient vanishing problem is solved by residual connection to optimize the training of the network:

[0087] Yrcs = Xconv + F(Xconv, W);

[0088] where F represents the combination of convolution operation and activation function (such as Relu function), W is the weight matrix; Xconv represents the result of convolution, and Yrcs represents the result of residual operation;

[0089] High-level features of the turnout image are extracted through multiple convolution layers, and finally a feature map F image is output, representing the spatial and structural information of the turnout in the image:

[0090] F image = ResNet(I);

[0091] where ResNet represents the ResNet network, and I represents the visual image.

[0092] It can be understood that for the laser radar point cloud data, PointNet is used because it can directly process the original point cloud and effectively handle the unordered nature of the point cloud. PointNet processes point cloud data by using a symmetric function.

[0093] The processing flow is as follows:

[0094] Specifically, the processing of the image and point cloud information of the target area respectively to identify the switch opening and closing state and the traffic light state of the target area includes:

[0095] Input point cloud data P={p1,p2,…,p n}∈R n×3 , where n is the number of points in the point cloud, p i =(C i ,y i ,z i ) is the three-dimensional coordinate of the i-th point in the point cloud;

[0096] Each point in the point cloud is converted into a high-dimensional feature vector through a multi-layer perceptron network:

[0097] f i =MLP(p i ,W);

[0098] Among them, MLP is a multi-layer perceptron, W is a weight matrix;

[0099] PointNet uses the maximum pooling operation to aggregate the features of all points to obtain a global feature representation:

[0100] F pointcloud =max(f i );

[0101] This operation performs maximum pooling on the features of all points to obtain the global descriptor F pointcloud , which represents the characteristics of the turnout structure in the point cloud.

[0102] Specifically, the present invention uses visual images to identify the status of traffic lights, including the following steps:

[0103] Image capture: Use a camera or other device to capture images of traffic intersections including traffic lights in real time.

[0104] Preprocessing: Clean the captured image, such as denoising, grayscale processing, and histogram equalization, in order to improve the effect of subsequent processing.

[0105] Object detection: Apply deep learning models such as YOLO (You Only Look Once), SSD (Single Shot MultiBox Detector), or Faster R-CNN to locate traffic lights in images. These models can identify the location and size of traffic lights in images.

[0106] Feature extraction: Use a convolutional neural network (CNN) to extract features such as color, shape, and texture from the detected traffic light areas.

[0107] State recognition: Based on the features, a classification model (such as SVM, decision tree, or again using CNN) is trained to recognize the specific color state of the traffic light (red, green, yellow, or off).

[0108] Specifically, the method further includes:

[0109] Using the weighted fusion method, the features and feature maps of the turnout structure in the point cloud are learned by machine learning. image The fusion weights of the turnout features in the model are used to automatically adjust their respective influences.

[0110] Specifically, the method further includes: confidence calculation: mapping the fused feature vector to a probability distribution through a fully connected layer or a support vector machine (SVM) to represent the possibility of different turnout states:

[0111] =Softmax(W fc F fused +b);

[0112] in, is the state probability distribution of the output, W fc and b are the weight and bias of the fully connected layer respectively; the probability of each switch state is obtained through the softmax function, and the switch state includes normal, abnormal, and alarm;

[0113] By confidence ,Dynamically adjust the fusion weight α, the dynamic adjustment formula is as follows:

[0114] a new =a old +β( max -0.5);

[0115] in, max is the maximum confidence value, β is the adjustment step size, a old is the weight before adjustment, a new is the weight after adjustment.

[0116] Preferably, β=0.1 or 0.05.

[0117] Specifically, the confidence level ,Dynamically adjust the fusion weight α, including:

[0118] When the confidence is low, increase the weight α of the point cloud feature to improve the stability of the system;

[0119] When the confidence is high, the weight α of the point cloud feature is reduced to enhance the influence of the visual feature.

[0120] Specifically, the present invention sets the confidence threshold in the range of [0.6, 0.9]:

[0121] When the confidence level is lower than 0.6, the system may not be able to make reliable judgments and the point cloud weight needs to be increased.

[0122] When the confidence level is higher than 0.9, the visual features are more credible and the point cloud weight is reduced.

[0123] Specifically, the weighted fusion method is used to learn the features and feature maps of the turnout structure in the machine learning point cloud. image The fusion weights of the turnout features in the model are automatically adjusted to adjust their respective influences, including:

[0124] Let the visual feature be F image And the point cloud feature is F pointcloud ;

[0125] Concatenate these two eigenvectors into a comprehensive eigenvector F fused :

[0126] F fused =αF image +(1-α)F pointcloud ;

[0127] Where α∈[0,1] is the fusion weight, which represents the contribution ratio of visual features to point cloud features. This weight is automatically adjusted based on system feedback. The initial value of α is preferably 0.5, indicating that the initial contribution ratio of visual features and point cloud features is equal.

[0128] Based on the vector F after feature fusion fused Make turnout status judgment and calculate the confidence of the recognition result;

[0129] The rule engine makes the final switch status judgment based on the fused features and confidence level. The decision rule formula is as follows:

[0130] State=RuleEngine(F fused , );

[0131] RuleEngine function outputs the switch state according to the features and confidence. State represents the switch state. F fused represents the comprehensive feature vector, is the state probability distribution of the output.

[0132] It is understandable that we use the rule engine to make the final switch status judgment based on the integrated features and confidence. The rules are as follows:

[0133] Normal traffic conditions: If the switch opening and closing status is normal and the traffic light is green, the system allows the locomotive to pass.

[0134] Abnormal status processing: If the switch status is abnormal (for example, the switch opening and closing is wrong or the traffic light does not match), an alarm is triggered.

[0135] Specifically, the method further includes:

[0136] The support vector machine (SVM) is used to judge the turnout status. The feature vector F is trained on the model. fused Mapped to the turnout state, the decision formula of SVM is:

[0137] y=sign(w T F fused +b);

[0138] Where W is the weight matrix, b is the bias, sign is the sign function, y represents the final state of the turnout, and T represents the transpose.

[0139] Specifically, y = +1: the turnout is in normal state (turnout opening and closing is normal), and the traffic light status complies with the traffic rules. y = −1: the turnout is in abnormal state (turnout opening and closing is incorrect), and the traffic light status does not comply with the traffic rules.

[0140] Understandably, for higher accuracy, we use a machine learning model, Support Vector Machine (SVM), to make more complex judgments. The model is trained to map feature vectors to turnout states.

[0141] Specifically, the method further includes:

[0142] During system operation, as more turnout status data is collected, we can perform incremental learning. Each time a new data point is obtained, the system will update the model parameters based on this new data point. The incremental update formula is as follows:

[0143] θ new =θ old +η·∇L(θ old ,F fused ,y true );

[0144] Among them, θ old is the parameter of the model before updating, θ new is the parameter of the updated model, η is the learning rate, ∇L is the gradient of the loss function, L is the loss function of the model, F fused represents the comprehensive feature vector, and ytrue represents the calculated state of the turnout.

[0145] It should be noted that during online learning, weights are automatically adjusted based on the new data distribution. By monitoring the system's performance (confidence and recognition accuracy), the system can adjust the feature fusion weight α to ensure that the system is always in optimal condition.

[0146] S300. If the switch opening and closing status and the traffic light status information both meet the traffic conditions, control the rail locomotive to pass the switch; if the switch opening and closing status or the traffic light status information do not meet the traffic conditions, output an alarm message indicating that the switch opening and closing status or the traffic light status information do not meet the traffic conditions.

[0147] Specifically, in a preferred embodiment, the method further includes:

[0148] If the switch opening and closing status and traffic lights meet the conditions for passage, the control system allows the locomotive to pass safely. If the switch opening and closing status or traffic lights are abnormal, the system will send an alarm to the dispatching center for further processing and instructions.

[0149] It is understandable that the present invention has the following technical effects:

[0150] 1. Save labor costs:

[0151] Reduce the need for driver confirmation and reduce human resource consumption, especially when double confirmation is required.

[0152] 2. Improve operational efficiency:

[0153] Speed ​​up the locomotive's passage through the switch and optimize the overall transportation process.

[0154] 3. Enhanced security:

[0155] Through automated judgment, the risk of human error is reduced and the safety of railway transportation is ensured.

[0156] 4. Support automated operations:

[0157] It provides the basis for unmanned and higher-level automated train operations, and improves the modernization level of the railway system.

[0158] 5. A deep learning model that combines vision and lidar feature extraction to achieve higher-precision turnout recognition.

[0159] 6. Dynamically adjust the fusion strategy based on the confidence of the recognition results to improve the adaptability of the system.

[0160] 7. Continuously optimize the model through real-time feedback and online learning to adapt to different environments and changing conditions.

[0161] See also Figure 2The present application provides another embodiment, which provides a rail locomotive control system, comprising:

[0162] The acquisition module 100 comprises a visual camera arranged in front of the rail locomotive, used for shooting a visual image of the target area; and a laser radar arranged in front of the rail locomotive, used for collecting point cloud information of the target area;

[0163] The control module 200 is used for processing the visual image and the point cloud information of the target area respectively, identifying the opening and closing state of the turnout of the target area, and acquiring the state information of the traffic light; if the opening and closing state of the turnout and the state information of the traffic light both meet the passing condition, the rail locomotive is controlled to pass through the turnout; if the opening and closing state of the turnout or the state information of the traffic light does not meet the passing condition, alarm information about the opening and closing state of the turnout or the state information of the traffic light not meeting the passing condition is output.

[0164] It should be noted that the present application has the following technical effects:

[0165] 1. Saving labor cost:

[0166] Reducing the confirmation requirement for the driver, reducing the consumption of human resources, especially in the case of double confirmation.

[0167] 2. Improving operation efficiency:

[0168] Speeding up the locomotive passing through the turnout, optimizing the overall transportation process.

[0169] 3. Enhancing safety:

[0170] Through automatic judgment, reducing the risk of human error, ensuring the safety of railway transportation.

[0171] 4. Supporting automated operation:

[0172] Providing a basis for unmanned and higher automated train operation, improving the modernization level of the railway system.

[0173] 5. Deep learning model combining visual and laser radar feature extraction, realizing higher precision turnout identification.

[0174] 6. Dynamically adjusting the fusion strategy according to the recognition result confidence, improving the adaptability of the system.

[0175] 7. Continuously optimizing the model through real-time feedback and online learning, adapting to different environments and changing conditions.

[0176] In the preferred embodiment, the present application also provides an electronic device, comprising:

[0177] The computer device can be a server, a terminal, or any other electronic device having necessary computing and / or processing capabilities. In one embodiment, the computer device can include a processor, a memory, a network interface, a communication interface, and the like connected by a system bus. The processor of the computer device can be configured to provide necessary computing, processing, and / or control capabilities. The memory of the computer device can include a non-volatile storage medium and an internal memory. The non-volatile storage medium can store an operating system, a computer program, and the like therein or thereon. The internal memory can provide an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface and the communication interface of the computer device can be configured to connect and communicate with external devices through a network. The computer program, when executed by the processor, performs the steps of the method of the present application.

[0178] The present application can be implemented as a computer readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the method of the embodiments of the present application to be performed. In one embodiment, the computer program is distributed over a plurality of computer devices or processors coupled by a network, such that the computer program is stored, accessed, and executed by one or more computer devices or processors in a distributed manner. A single method step / operation, or two or more method steps / operations, can be performed by a single computer device or processor, or by two or more computer devices or processors. One or more method steps / operations can be performed by one or more computer devices or processors, and one or more other method steps / operations can be performed by one or more other computer devices or processors. One or more computer devices or processors can perform a single method step / operation, or perform two or more method steps / operations.

[0179] It will be appreciated by those skilled in the art that the method steps of the present application can be instructed by a computer program to relevant hardware such as a computer device or a processor, which can be stored in a non-transitory computer readable storage medium, and which, when executed, causes the steps of the present application to be performed. Depending on the circumstances, any reference herein to a memory, storage, database, or other medium can include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tapes, floppy disks, magneto-optical data storage devices, optical data storage devices, hard disks, solid-state disks, and the like. Examples of volatile memory include random access memory (RAM), external cache memory, and the like.

[0180] It is understandable that the present invention has the following technical effects:

[0181] 1. Save labor costs:

[0182] Reduce the need for driver confirmation and reduce human resource consumption, especially when double confirmation is required.

[0183] 2. Improve operational efficiency:

[0184] Speed ​​up the locomotive's passage through the switch and optimize the overall transportation process.

[0185] 3. Enhanced security:

[0186] Through automated judgment, the risk of human error is reduced and the safety of railway transportation is ensured.

[0187] 4. Support automated operations:

[0188] It provides the basis for unmanned and higher-level automated train operations, and improves the modernization level of the railway system.

[0189] 5. A deep learning model that combines vision and lidar feature extraction to achieve higher-precision turnout recognition.

[0190] 6. Dynamically adjust the fusion strategy based on the confidence of the recognition results to improve the adaptability of the system.

[0191] 7. Continuously optimize the model through real-time feedback and online learning to adapt to different environments and changing conditions.

[0192] The various technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification as long as such combination does not conflict.

[0193] The specific embodiments of the present invention described above do not limit the scope of protection of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A railway locomotive control method, characterized in that: The method comprises: S100, using a visual camera disposed in front of the rail vehicle to capture a visual image of the target area; Use the laser radar installed in front of the rail locomotive to collect point cloud information of the target area; S200, processing the visual image and point cloud information of the target area respectively, identifying the opening and closing status of the turnout in the target area, and obtaining the status information of the traffic light; S300: If the switch opening / closing state and the traffic light status information both meet the passage conditions, control the locomotive to pass through the switch; if the switch opening / closing state or the traffic light status information does not meet the passage conditions, output an alarm message indicating that the switch opening / closing state or the traffic light status information does not meet the passage conditions; For visual feature extraction, ResNet is used as the basic network for feature extraction to obtain the feature map F image , for the lidar point cloud data, use PointNet to process it and get the point cloud feature F pointcloud ; Specifically, the method further includes: Using the weighted fusion method, the features and feature maps of the turnout structure in the point cloud are learned by machine learning. image The fusion weights of the turnout features in the ensemble are used to automatically adjust their respective influences; Specifically, the weighted fusion method is used to learn the features and feature maps of the turnout structure in the machine learning point cloud. image The fusion weights of the turnout features in the model are automatically adjusted to adjust their respective influences, including: Let the visual feature be F image And the point cloud feature is F pointcloud ; Concatenate these two eigenvectors into a comprehensive eigenvector F fused : F fused =αF image +(1-α)F pointcloud ; Among them, α∈[0,1] is the fusion weight, which represents the contribution ratio of visual features and point cloud features; Based on the vector F after feature fusion fused Make turnout status judgment and calculate the confidence of the recognition result; The rule engine makes the final switch status judgment based on the fused features and confidence level. The decision rule formula is as follows: State=RuleEngine(F fused , ); RuleEngine function outputs the switch state according to the features and confidence. State represents the switch state. F fused represents the comprehensive feature vector, is the state probability distribution of the output; Specifically, the method further includes confidence calculation: mapping the fused feature vector to a probability distribution through a fully connected layer or a support vector machine (SVM) to represent the likelihood of different turnout states: =Softmax(W fc F fused +b); in, is the state probability distribution of the output, W fc and b are the weight and bias of the fully connected layer respectively; the probability of each switch state is obtained through the softmax function, and the switch state includes normal, abnormal, and alarm; By confidence ,Dynamically adjust the fusion weight α, the dynamic adjustment formula is as follows: a new =a old +β( max -0.5); in, max is the maximum confidence value, β is the adjustment step size, a old is the weight before adjustment, a new is the weight after adjustment.

2. The railway locomotive control method according to claim 1, characterized in that: The processing of the visual image and point cloud information of the target area respectively, identifying the opening and closing status of the turnout in the target area, and obtaining the status information of the traffic light includes: Apply a 3×3 convolution kernel K to perform a convolution operation on the visual image: Xconv =Conv(I,K); Among them, the convolution kernel K is a weight matrix, and the local features of the turnout in the image are extracted through multi-layer convolution. Xconv represents the result of convolution, Conv represents the convolution function, and I represents the visual image. Residual connections are used to solve the gradient vanishing problem and optimize network training: Yrcs=Xconv+F(Xconv,W); Where F represents the combination of convolution operation and activation function, W is the weight matrix; Xconv represents the result of convolution, and Yrcs represents the result of residual operation; The high-level features of the turnout image are extracted through multiple convolutional layers, and the feature map F is finally output. image , represents the spatial and structural information of the turnout in the image: F image =ResNet(I); Among them, ResNet represents the ResNet network, and I represents the visual image.

3. The railway locomotive control method according to claim 2, characterized in that: The processing of the visual image and point cloud information of the target area respectively, identifying the opening and closing status of the turnout in the target area, and obtaining the status information of the traffic light includes: Input point cloud data P={p1,p2,…,p n }∈R n×3 , where n is the number of points in the point cloud, p i =(C i ,y i ,z i ) is the three-dimensional coordinate of the i-th point in the point cloud; Each point in the point cloud is converted into a high-dimensional feature vector through a multi-layer perceptron network: f i =MLP(p i ,W); Among them, MLP is a multi-layer perceptron, W is a weight matrix; PointNet uses the maximum pooling operation to aggregate the features of all points to obtain a global feature representation: F pointcloud =max(f i ); This operation performs maximum pooling on the features of all points to obtain the global descriptor F pointcloud , represents the characteristics of the turnout structure in the point cloud; R represents the set of real numbers, C i 、y i 、z i are the three coordinate values ​​of the three-dimensional coordinates of the i-th point in the point cloud.

4. The railway locomotive control method according to claim 1, characterized in that: The method further comprises: The support vector machine (SVM) is used to judge the turnout status. The feature vector F is trained on the model. fused Mapped to the turnout state, the decision formula of SVM is: y=sign(w T F fused +b); Where W is the weight matrix, b is the bias, sign is the sign function, y represents the final state of the turnout, and T represents the transpose.

5. The railway locomotive control method according to claim 1, characterized in that: The method further comprises: During system operation, as more turnout status data is collected, we can perform incremental learning. Each time a new data point is obtained, the system will update the model parameters based on this new data point. The incremental update formula is as follows: i new =θ old +η·∇L(θ old ,F fused ,y true ); Among them, θ old is the parameter of the model before updating, θ new is the parameter of the updated model, η is the learning rate, ∇L is the gradient of the loss function, L is the loss function of the model, F fused represents the comprehensive feature vector, y true Indicates the calculated status of the turnout.

6. A railway locomotive control system, characterized in that: The railway locomotive control method according to any one of claims 1 to 5 comprises: The acquisition module includes a visual camera arranged in front of the rail locomotive, which is used to capture the visual image of the target area; and also includes a laser radar arranged in front of the rail locomotive, which is used to collect point cloud information of the target area; The control module is used to process the visual image and point cloud information of the target area respectively, identify the opening and closing status of the switch in the target area, and obtain the status information of the traffic light; if the opening and closing status of the switch and the status information of the traffic light both meet the traffic conditions, control the rail locomotive to pass through the switch; if the opening and closing status of the switch or the status information of the traffic light do not meet the traffic conditions, output an alarm message indicating that the opening and closing status of the switch or the status information of the traffic light do not meet the traffic conditions.

7. An electronic device, characterized in that: include: Memory; and a processor, wherein the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the railway locomotive control method according to any one of claims 1 to 5 is implemented.

Images