A dynamic verification system for shunting safety based on multi-source sensing and interlocking logic

Abstract

This invention relates to the field of locomotive shunting safety technology, and discloses a dynamic verification system for shunting safety based on multi-source sensing and interlocking logic. The system includes: a positioning subsystem that collects real-time location information of personnel and locomotives via positioning terminals and modules and transmits it to the central control system; an equipment status sensing subsystem that deploys status acquisition units at key nodes to acquire and upload the physical status of anti-runaway devices, derailment devices, switches, and level crossings in real time; a central control system that stores shunting coupler plans and the expected status of track-related equipment and vehicle sets, and a built-in interlocking comparison module that verifies the consistency between the received physical status and the expected status to generate a comparison result; and a spatial relationship model established based on the personnel and vehicle positions, combined with the comparison result to generate linkage control commands, which are then sent to the ground signal execution unit to perform safety protection operations. This invention can improve the overall safety of the shunting safety protection system.

Description

Technical Field

[0001] This invention belongs to the field of locomotive shunting safety technology, and particularly relates to a dynamic verification system for shunting safety based on multi-source sensing and interlocking logic. Background Technology

[0002] With the rapid development of the railway and urban rail transit industries, shunting operations at depots, parking lots, and marshalling yards are becoming increasingly frequent. As a core component of the rail transit system, shunting operations involve numerous safety factors, including personnel, vehicles, tracks, and equipment. The working environment is complex, and the reliability of safety protection is directly related to the safety of personnel and national property.

[0003] Currently, in the existing depot shunting operation system, safety protection mainly relies on traditional manual operation, localized automated equipment, and decentralized video surveillance. Typical solutions in the existing technology include: for turnout control, local operation or centralized interlocking control of electric switch machines is commonly used; for anti-runaway measures, ordinary anti-runaway wheel chocks are widely used, and their installation and removal rely entirely on manual execution and confirmation; for derailment device management, manual or semi-automatic control is used, relying on mechanical limit switches or manual inspection to determine their "raised / lowered" status; for level crossing protection, although some crossings are equipped with barriers or audible and visual alarms, they are often controlled by local buttons or timers; for personnel and vehicle positioning, the location of on-site personnel is mainly determined by walkie-talkie communication or visual judgment, while the locomotive position is roughly estimated based on the dispatcher's experience or track circuit occupancy information.

[0004] However, in practical applications, the aforementioned existing technologies lack systematic data integration and real-time linkage capabilities, exhibiting the following obvious objective shortcomings: First, the status of anti-runaway devices cannot be sensed in real time, posing a risk of derailment due to missed removal or misplacement. Existing technology commonly uses ordinary anti-runaway trackers without sensor feedback devices; their status confirmation relies entirely on manual verification of paper plans and visual inspection. If the trackers are not removed in time due to personnel negligence before shunting operations begin, a "track-pulling derailment" accident is highly likely once the locomotive begins to pull the train. Furthermore, if the anti-runaway device is incorrectly placed in an undesignated location, it may also cause vehicle collisions.

[0005] Second, the reliability of key equipment status assessment is low, and there is a lack of dynamic consistency verification with shunting routes. Traditional derailers mainly rely on mechanical limit switches to determine their status, which are easily affected by environmental factors such as dust, corrosion, and mechanical jamming, resulting in false alarms. Furthermore, they cannot meet the real-time requirements of dynamic shunting route management. At the same time, the actual indication information of turnout switch machines is often distorted due to communication interruptions or equipment failures. If dispatchers arrange routes based on incorrect indication information, and the system lacks a verification mechanism for the dynamic matching of equipment mechanical status with shunting route plans, locomotives will enter the wrong track, causing lateral collisions or derailment accidents.

[0006] Third, the protection measures at level crossings are passive and isolated, failing to balance safety and traffic efficiency. Existing crossing barriers are not logically integrated or data-interlocked with the shunting route system. If the barriers fail to close in time when a locomotive is about to pass, pedestrians or vehicles within the yard could easily run over, causing a collision. Conversely, if the barriers are kept closed for safety reasons, it severely disrupts normal traffic flow within the yard, creating a safety hazard of "barriers not closed when they should, and vehicles entering when barriers are only partially closed."

[0007] Fourth, the positioning accuracy is severely insufficient, lacking high-precision warning capabilities for intrusion into work areas. Current technology has extremely low resolution in perceiving the positions of personnel and locomotives (track circuits often only provide a rough assessment of sections ranging from tens to hundreds of meters), making it impossible to construct a precise three-dimensional spatial relationship between "person-vehicle-track." When workers are operating between tracks or locomotives approach high-risk areas such as end-point barriers, the system cannot accurately warn of dangerous distances, causing personal injury protection and collision protection to heavily rely on the individual experience of workers and drivers.

[0008] In conclusion, the overall safety of existing shunting safety protection technologies is relatively low. Summary of the Invention

[0009] This invention provides a dynamic verification system for shunting safety based on multi-source sensing and interlocking logic, which can improve the overall safety of the shunting safety protection system.

[0010] To achieve the above objectives, the present invention provides a shunting safety dynamic verification system based on multi-source perception and interlocking logic, comprising: a central control system, and a positioning subsystem, an equipment status perception subsystem, a ground signal execution unit, and a locomotive on-board terminal that are communicatively connected to the central control system. The positioning subsystem includes a positioning terminal and a positioning module. The positioning terminal is used to collect the location information of the operators in real time and send the location information of the operators to the central control system. The positioning module is used to collect the location information of the locomotive in real time and send the location information of the locomotive to the central control system. The equipment status perception subsystem includes status acquisition units deployed at multiple key nodes within the target operating area. These units acquire real-time physical status information of shunting anti-runaway equipment, derailment devices, turnout equipment, and level crossings, and transmit this information to the central control system. The equipment status perception subsystem includes a derailment monitoring unit, which comprises a high-definition camera, physical sensors, and an image analysis unit. The high-definition camera is positioned at the derailment installation point to capture track surface images. The image analysis unit extracts and classifies features from the track surface images based on a pre-loaded computer vision detection model, outputting the visual recognition position status of the derailment. The physical sensors acquire the mechanical limit status of the derailment. The central control system stores a shunting coupler plan, which includes the expected state set and expected vehicle set of each track-related equipment. The central control system is equipped with an interlocking comparison module. The interlocking comparison module extracts the expected state set and performs consistency condition verification between the received physical state information and the expected state set to generate a state comparison result. The process of performing consistency condition verification between the received physical state information and the expected state set to generate a state comparison result includes: the interlocking comparison module receives the visual recognition position state and the mechanical limit state and performs cross-validation to determine the actual state, and performs condition verification between the actual state and the corresponding state in the expected state set to generate a state comparison result. The central control system is used to establish a spatial relationship model based on the location information of the operators and the locomotive, generate linkage control commands based on the state comparison results, and send the linkage control commands to the ground signal execution unit to perform safety protection operations. The ground signal execution unit is connected to the locomotive on-board terminal and the central control system respectively.

[0011] This invention utilizes a positioning subsystem comprising a positioning terminal and a positioning module. The positioning terminal collects real-time operator location information and transmits it to the central control system. The positioning module collects real-time locomotive location information and transmits it to the central control system, enabling real-time acquisition of the physical status of key train operation equipment. Furthermore, an equipment status perception subsystem includes status acquisition units deployed at multiple key nodes within the target operation area. These units acquire real-time physical status information of shunting anti-runaway equipment, derailment devices, turnout equipment, and level crossings, transmitting it to the central control system. This eliminates reliance on manual verification and visual inspection. Additionally, the central control system is equipped with an interlocking comparison module, which extracts pre-locking data. The system generates a state comparison result by verifying the consistency between the received physical state information and the expected state set. The central control system is used to establish a spatial relationship model based on the location information of the operators and the locomotive. It generates linkage control commands based on the state comparison results and sends the linkage control commands to the ground signal execution unit to perform safety protection operations. The ground signal execution unit is connected to the locomotive on-board terminal and the central control system. It can realize the dynamic consistency verification between the real-time received physical state of the equipment and the preset expected state set in the shunting coupler plan. Based on the high-precision personnel and locomotive positions, it establishes a dynamic three-dimensional spatial relationship model and automatically generates targeted control commands based on the state comparison results, which can improve the overall safety of the shunting safety protection system. Attached Figure Description

[0012] Figure 1 This is a system structure diagram of a shunting safety dynamic verification system based on multi-source sensing and interlocking logic, provided as an embodiment of the present invention.

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0015] This application provides a dynamic verification system for shunting safety based on multi-source sensing and interlocking logic. The executing entity of this dynamic verification system includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the dynamic verification system for shunting safety based on multi-source sensing and interlocking logic can be executed by software or hardware installed on a terminal device or a server device. The software can be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0016] Reference Figure 1 The diagram shown is a system structure diagram of a shunting safety dynamic verification system based on multi-source sensing and interlocking logic according to an embodiment of the present invention. In this embodiment, the shunting safety dynamic verification system based on multi-source sensing and interlocking logic includes: a central control system, and a positioning subsystem, an equipment status sensing subsystem, a ground signal execution unit, and a locomotive onboard terminal that are communicatively connected to the central control system.

[0017] The positioning subsystem includes a positioning terminal and a positioning module. The positioning terminal is used to collect the location information of the operators in real time and send the location information of the operators to the central control system. The positioning module is used to collect the location information of the locomotive in real time and send the location information of the locomotive to the central control system.

[0018] Understandably, the positioning subsystem refers to the real-time spatial perception unit that constructs a three-dimensional spatial relationship model of "people-vehicle-line". Through two-way communication with the central control system, the positioning subsystem transforms discrete physical location information into spatial topology data with business logic, thereby providing an indispensable spatial dimension reference for dynamic verification and proactive intervention of shunting operations.

[0019] The equipment status perception subsystem includes status acquisition units deployed at multiple key nodes within the target operating area. These units acquire real-time physical status information of shunting anti-runaway equipment, derailment devices, turnout equipment, and level crossings, and transmit this information to the central control system. The equipment status perception subsystem also includes a derailment monitoring unit, which comprises a high-definition camera, physical sensors, and an image analysis unit. The high-definition camera is positioned at the derailment installation point to capture track surface images. The image analysis unit performs feature extraction and classification on the track surface images based on a pre-loaded computer vision detection model, outputting the visual recognition position status of the derailment. The physical sensors acquire the mechanical limit status of the derailment.

[0020] Understandably, the equipment status perception subsystem refers to a distributed Internet of Things (IoT) monitoring network composed of various front-end sensing devices deployed in key locations such as working tracks, throat areas, and level crossings within a depot, marshalling yard, or parking lot.

[0021] Understandably, the status acquisition unit refers to the front-end hardware and communication nodes in the equipment status perception subsystem that specifically perform environmental data perception, edge computing processing, and status reporting. Depending on the target equipment being monitored, its specific form and the technologies employed vary. For example, for a smart track shoe unit, the status acquisition unit is a built-in pressure / displacement sensor and a wireless IoT transmission module; for derailment equipment and level crossings, the status acquisition unit is a high-definition network camera, physical limit sensors, and a supporting edge AI visual analysis host; for turnouts, the status acquisition unit is an electric switch machine data interface module accessed via RS485, CAN bus, or Ethernet.

[0022] Understandably, shunting anti-runaway equipment refers to a special safety braking tool installed between the wheel tread and the rail during shunting, marshalling, or maintenance operations in rail transit to prevent stationary or parked unpowered vehicles from moving unexpectedly (i.e., "runaway") due to their own weight, wind force, or track gradient.

[0023] Understandably, turnout equipment refers to key track turning infrastructure installed at the branches or intersections of railway lines to guide locomotives and rolling stock from one track to another or across another track.

[0024] Understandably, a level crossing refers to an area within a depot, parking lot, or marshalling yard where railway lines (tracks) intersect and overlap with roads, motor vehicle lanes, or pedestrian walkways on the same plane. Level crossings are typical high-risk nodes where locomotives and rolling stock intersect with motor vehicles and pedestrians within the yard.

[0025] Specifically, the equipment status perception subsystem includes an intelligent iron shoe unit, which includes at least one pressure displacement sensor. The pressure displacement sensor is used to collect iron shoe status information and send the iron shoe status information to the central control system. The iron shoe status information includes track clamping status information and box return status information. The interlocking comparison module retrieves the expected anti-runaway status of the target track in the shunting coupler plan. When the expected anti-runaway status is determined to be a removal condition and the received iron shoe status information is in the on-rail clamping state, the linkage control command is configured to prohibit the ground signal execution unit from opening the shunting signal.

[0026] Understandably, the on-rail clamping status information refers to the working status data sensed and generated by the pressure displacement sensor integrated inside the smart wheel shoe, indicating that the wheel shoe has been placed on the rail surface of the railway line, and is in close contact with the rail and wheel tread, bearing physical pressure. In the shunting safety protection logic, the on-rail clamping status information clearly indicates that the vehicle on the target track is in a physical anti-runaway locking state; if the expected action of the shunting plan at this time is for the locomotive to pull or advance the vehicle on that track, the on-rail clamping status information is a serious "track obstruction", which will trigger the system's "prohibit vehicle movement" logic lockout to prevent a "wheel shoe derailment" accident.

[0027] Understandably, the "return-to-box status information" refers to the physical status data indicating that the smart track shoes have been completely removed from the rails and placed in the designated storage box or rack location on site. In the system's dynamic safety verification mechanism, this information is the sole core electronic credential confirming that anti-slip measures have been lifted on the target track and that there are no longer any physical obstacles from track shoes on the track surface. Only when the central control system receives the "return-to-box status information" of all relevant track shoes for the planned departure track is the necessary safety prerequisite for allowing the locomotive to enter or commence operations met.

[0028] Understandably, the interlocking comparison module refers to the core logic operation and dynamic verification software unit built into the central control system.

[0029] Understandably, the ground signal execution unit refers to the basic terminal equipment system within a railway station used to transmit train operation permits to locomotive drivers and is responsible for route locking and signal opening at the physical control level.

[0030] Furthermore, the equipment status perception subsystem includes a derailment monitoring unit, which includes a high-definition camera, physical sensors, and an image analysis unit. The high-definition camera is positioned at the derailment installation point to capture images of the track surface. The image analysis unit performs feature extraction and classification on the track surface images based on a pre-loaded computer vision detection model, and outputs the visual recognition position status of the derailment. The physical sensor collects the mechanical limit status of the derailer;

[0031] Understandably, the derailment monitoring unit refers to a composite state sensing and edge computing terminal specifically deployed at the derailment equipment site. It aims to overcome the shortcomings of traditional derailment devices that rely solely on a single mechanical limit switch, which are susceptible to interference from the outdoor environment (such as rust, jamming, and dust accumulation), leading to distorted state feedback.

[0032] Understandably, a high-definition camera refers to visual image acquisition hardware that is fixedly deployed above or to the side of the derailer's installation location.

[0033] Understandably, the mechanical limit state refers to the electrical signal captured and output by physical sensors (such as travel limit switches, angle sensors, etc.) installed at the mechanical transmission components, shafts, or track joints of the derailer.

[0034] Understandably, feature extraction and classification refer to the high-dimensional mathematical abstraction process performed by the deep learning object detection model built into the image analysis unit on the input original two-dimensional track surface pixel image. The deep learning object detection model can be an improved YOLO model based on the ResNet50 backbone network and feature pyramid network. This model is a single-stage end-to-end object detection algorithm. Compared with two-stage models such as Faster R-CNN, it can achieve a detection speed of over 200fps while ensuring detection accuracy (≥98%), which fully matches the 25fps acquisition frame rate requirement of track surface images. It is also suitable for complex lighting and background interference scenarios in railway sites.

[0035] For example, a deep learning object detection model can be implemented as follows: Step 1: (1) Model input; Basic input data: Color images of the track surface captured by a high-definition network camera, with a uniform resolution of 1920×1080 pixels, image format of JPG / PNG, and frame rate of 25fps consistent with the acquisition end; Preprocessed input: The original image is processed by MBLLEN technology for illumination enhancement, noise reduction, size normalization, and normalization of pixel values ​​(to the 0-1 range) to generate fixed-dimensional tensor data, which serves as the direct input for model inference. MBLLEN technology is a Multi-branch Low-light Enhancement Network, whose core function is to transform low-quality dark light images into high-quality standard images; Additional input information: Optional input metadata such as image acquisition time, track segment number, and camera device ID, which are used for subsequent state tracing and logical comparison.

[0036] (2) Model output: The output is a structured tensor result, which contains the core information of derailer state recognition after parsing. It fully matches the system logic comparison and anomaly feedback requirements. Specific fields: Target detection box information: pixel coordinates of the derailer in the image (x1, y1, x2, y2), detection box confidence (0-1), confidence threshold is set to 0.85 (below the threshold is judged as no derailer detected); State classification result: the specific state of the derailer, which is a three-class discrete value, corresponding to: 0=lower rail of derailer, 1=upper rail of derailer, 2=foreign object occlusion; Classification confidence: the probability of the model to classify the above states (0-1), which must be ≥0.9 to be considered as a valid recognition result, otherwise image resampling and secondary inference are triggered; Inference time: the model inference time of a single frame image (ms level), used to monitor the real-time performance of the model. Output data encapsulation: The inference results can be directly parsed into JSON format, which is compatible with the error message format pushed by the system via TCP / IP protocol. TCP / IP protocol (Transmission Control Protocol / Internet Protocol) is the basic communication architecture of the Internet. An example of JSON format is as follows: { "track_id": "G012", "camera_id": "CAM008", "detect_box": [356,289,412,345], "box_conf": 0.96, "device_state": 1, "state_conf": 0.98, "infer_time": 4 } Step 2: Network structure hierarchy of the model; The model adopts the classic three-level architecture of backbone network - neck network - head network. Each level is customized and optimized for railway scenarios, with clear functional division and collaborative work, to achieve accurate extraction, fusion and classification detection of derailment features; (1) Backbone network: ResNet50 deep residual network 1. Core function: Extract multi-level feature information from the preprocessed track surface image, including low-level texture / edge features (derailer shape) and high-level semantic features (relative position of derailer and track surface), while avoiding the gradient vanishing problem in deep network training; 2. Structural composition: It consists of an input layer, a 4-stage residual block, and an output layer. Each residual block contains 2 convolutional layers. It performs feature dimensionality reduction / upgrading through 1×1 convolution and feature extraction through 3×3 convolution. It also introduces shortcut connections so that the input features can skip some layers and ensure that the derailer features are not submerged by background noise such as track and tunnel walls; 3. Scene optimization: For trackside scenes, redundant convolutional layers are removed to reduce the amount of computation and adapt to the real-time inference requirements of the central server.

[0037] (2) Neck Network: Recurrent Feature Pyramid Network (RFPN) 1. Core Function: Multi-scale fusion and enhancement of features extracted by the backbone network to solve the detection accuracy problem of derailers as small targets, and at the same time enhance the feature extraction capability in foreign object occlusion scenarios; 2. Structural Composition: Composed of three parts: bottom-up path, top-down path, and lateral connection. A 4x downsampled feature map is added to retain the detailed information of the derailer as a small target; Bottom-up path: The backbone network features are downsampled step by step to generate feature maps of different scales; Top-down path: The resolution of high-level feature maps is restored by upsampling and matched with low-level feature maps; Lateral connection: The feature maps of different scales are fused, combining low-level detailed features with high-level semantic features; (3) Head Network: IoU-Aware Detection Head; 1. Core Function: Based on the feature map fused by the neck network, it completes the target localization (detection box regression) and state classification of the derailer, and is the final output layer of the model inference; 2. Structural Composition: It consists of a convolutional layer (3×3), a sigmoid activation layer, a classifier, and a regressor, integrating an IoU-Aware intersection-union perception structure; 3. Core Mechanism: It uses the CIoU loss function to optimize the detection box regression, which not only considers the degree of overlap between the predicted box and the real box, but also quantifies the distance between the center point and the aspect ratio, thereby improving the localization accuracy of the derailer; The classifier uses the Softmax activation function to achieve the three classifications of the derailer: "upper rail / lower rail / object occlusion". The CIoU (Complete Intersection over Union) loss function is a loss function used for bounding box regression in object detection; Step 3: For the derailment state recognition model, the following standardized training process is established, based entirely on supervised deep learning training, and consists of 6 core steps. All parameters and operations are quantified and reproducible: Step 1: Construction and Labeling of Training Sample Set: Sample Collection: Under different conditions (sunny / rainy / night / fog) across the country, track surface images of derailers in three states—"upper rail, lower rail, and obstruction by foreign objects (weeds / plastic bags / metal objects)"—were collected. The total sample size was ≥100,000 images, with the sample ratio for each state being 4:4:2, covering different track sections, camera angles, and lighting conditions. Sample Labeling: Manual labeling was performed using the LabelImg labeling tool. The labeling content included: the detection box coordinates of the derailer (x1, y1, x2, y2); state classification labels (0=lower rail, 1=upper rail, 2=obstruction by foreign objects). After labeling, VOC / COCO format label files were generated, corresponding one-to-one with the images. Sample Division: The sample set was divided into a training set (70,000 images), a validation set (20,000 images), and a test set (10,000 images) in a 7:2:1 ratio to ensure that the state distribution of each subset was consistent with the total sample set and to avoid data bias.

[0038] Step 2: Sample Preprocessing and Data Augmentation: To improve the model's generalization and anti-interference capabilities, non-destructive preprocessing and data augmentation are performed on the training set, while the validation and test sets are only preprocessed without data augmentation: Basic preprocessing: Unify image resolution to 1920×1080, normalize pixel values ​​(divided by 255 to the 0-1 range), randomly flip horizontally (probability 0.5), and adjust image brightness / contrast (±20%); Targeted augmentation: For railway trackside scenes, Gaussian noise is added (σ=0.01), fog effect simulation, and shadow simulation to simulate complex outdoor working conditions; For foreign object occlusion scenes, different types of foreign object occlusion layers are randomly added to improve the model's occlusion recognition capability; Data format conversion: The preprocessed images and annotation files are converted into tensor formats that the model can read and packaged in batches (batch size=32).

[0039] Step 3: Model Initialization and Hyperparameter Setting: Model Initialization: A transfer learning strategy is adopted, loading ResNet50 weights pre-trained on the ImageNet dataset as the initial weights for the backbone network. The neck and head networks are initialized using Xavier, a classic method for initializing neural network weights (parameters) to avoid slow training convergence caused by random initialization; Core Hyperparameters: All trainable parameters are explicitly set and quantized to specific values, including: Optimizer: AdamW optimizer, learning rate = 0.001, weight decay = 0.0005, β1 = 0.9, β2 = 0.999; Training Batch: batch size = 32, epoch = 100; Learning Rate Scheduling: Cosine annealing strategy is adopted, with the learning rate decaying by 0.5 every 10 epochs, minimum learning rate = 0.00001; Loss Function Weights: Classification loss weights = 1.0, detection box regression loss weight = 2.0; Model training and process monitoring: Training mode: End-to-end batch training is adopted, the packaged training set batch data is input into the model, the prediction results are obtained through forward propagation, and the weighted total loss of CIoU regression loss and cross-entropy classification loss is calculated; Backpropagation: The gradient of the loss function with respect to all model parameters is calculated through the gradient descent algorithm, the parameters are updated using the AdamW optimizer, and gradient clipping (max_norm=5.0) is used to prevent gradient explosion; Training monitoring: After each epoch (training round), the model is validated on the validation set, and the validation set precision, recall, F1 score (harmonic mean of validation set precision and recall), and average detection precision (mAP@0.5) are recorded; Early stopping mechanism: If the validation set mAP does not improve for 10 consecutive epochs (training rounds), training is stopped immediately to avoid overfitting.

[0040] Step 5: Model Tuning and Weight Saving: Hyperparameter Tuning: For scenarios with poor performance on the validation set (such as foreign object occlusion at night), fine-tune the learning rate (reducing it to 0.0001) and add 50 epochs (training rounds) of fine-tuning training; Model Pruning: Perform lightweight pruning on the trained model, removing redundant convolutional layers and neurons, retaining the core feature extraction layer, reducing the number of model parameters by ≥60%, while ensuring that the detection accuracy loss is ≤0.2%; Optimal Weight Saving: Save the model weight file (.pth format) with the highest mAP on the validation set, recording all training parameters, training steps, and loss curves corresponding to that weight, as pre-training weights for model inference.

[0041] Step 6: Model Performance Verification and Result Disclosure Core Accuracy Indicators: mAP@0.5 ≥ 98.5%; Classification Accuracy for Each State: Upper Rail ≥ 99%, Lower Rail ≥ 99%, Foreign Object Occlusion ≥ 97%; Real-time Indicators: Single-frame image inference time ≤ 5ms, frame rate ≥ 200fps, meeting the real-time processing requirements of a 25fps acquisition end; Robustness Indicators: Detection accuracy decreases by ≤ 1% under rain / night / fog / haze conditions; When the foreign object occlusion rate is ≤ 50%, the basic state of the derailment device can still be accurately identified; Disclosed Content: The working condition distribution of the test set, the specific test results for each indicator, the confusion matrix of the model inference, and the deployment environment and inference code framework of the model are also disclosed.

[0042] Step 7: Model Deployment and Pre-training Weight Update: Model Deployment: Convert the trained pre-trained model to ONNX format (Open Neural Network Exchange), deploy it to the central server, and perform joint debugging with the image acquisition terminal, logic comparison module, and anomaly feedback module; Incremental Training: In practical applications, acquire new derailment state images (especially those of uncovered working conditions), add them to the sample set for incremental training, update the pre-trained weights, and improve the model's ability to adapt to real-world scenarios.

[0043] Furthermore, the equipment status sensing subsystem includes a turnout contact monitoring unit, which includes a high-precision position sensor installed in the turnout switch rail area. The high-precision position sensor collects real-time information on the mechanical contact and positioning status of the switch rail and sends it to the central control system. The interlocking comparison module obtains the route opening logic command issued by the central control system and dynamically compares the mechanical tight-fitting status information of the switch rail with the route opening logic command. When there is a difference between the actual mechanical contact position and the position required by the activation command in the comparison results, the linkage control command is set to block the associated path.

[0044] Understandably, the switch point area refers to the specific physical space within a railway switch structure where the laterally movable section of track at the front end (i.e., the switch rail) is in contact with the fixed track (i.e., the stock rail) it relies on. During railway traffic and shunting operations, the relative position of the switch point area directly determines the direction of wheel guidance (e.g., opening the main line or the siding).

[0045] Understandably, a high-precision position sensor refers to the underlying sensing hardware that is directly installed in the switch point rail area mentioned above to precisely measure the minute physical gap between the switch point rail and the main rail.

[0046] Understandably, blocking associated routes refers to the highest level of proactive safety intervention and defense mechanism triggered when the interlocking comparison module of the central control system detects that the "physical measured tightness status" of the turnout is inconsistent with the "planned opening direction" issued by the dispatcher (i.e., there is a risk of mishandling, non-tightness, or turnout derailment).

[0047] Furthermore, the equipment status perception subsystem includes a level crossing linkage unit, which includes a level crossing barrier controller and a video confirmation unit; The central control system acquires the locomotive's location information, calculates the time node for the locomotive to arrive at the target level crossing based on the operation path set in the shunting coupler plan, and sends an automatic closing command to the level crossing barrier controller based on the time node. The video confirmation unit is used to capture and analyze the action of the railing. When it detects that the railing is actually closed in place, it sends a railing positioning command to the central control system. The central control system monitors the barrier positioning command and, upon receiving the command, responds by opening the travel path ahead of the train.

[0048] Understandably, a level crossing linkage unit refers to an intelligent set of equipment systems integrated and deployed at the intersection of railway lines and on-site roads to achieve physical isolation of the crossing and deep interlocking control of locomotive shunting routes.

[0049] Understandably, a level crossing barrier controller is an industrial-grade logic control and power execution hub deployed at the end of a level crossing within the site and driven by remote commands from a central control system. It is responsible for driving the level crossing barrier to perform raising or lowering actions.

[0050] Furthermore, the equipment status perception subsystem also includes a real-time vehicle feature recognition unit, and the central control system is also equipped with a vehicle consistency verification module. The expected vehicle set within the shunting coupler plan includes a list of planned vehicle identifiers for the vehicles to be dispatched. The vehicle feature real-time identification unit acquires the actual vehicle set identifier in the target track that is in a physically coupled state, and sends the actual vehicle set identifier to the central control system; The vehicle consistency verification module extracts the actual vehicle set identifiers and the planned vehicle identifier list, performs set difference and set union operations, and extracts the missing vehicle set features and redundant vehicle set features. When the result of merging the missing vehicle set features and the redundant vehicle set features is not empty, the vehicle consistency verification module outputs a consistency Boolean flag indicating that there is a difference. The central control system, based on the consistent Boolean flag with discrepancies, sets the linkage control command to block the opening of subsequent signals and issues a fault-prevention forced locking command to the locomotive on-board terminal.

[0051] Understandably, the consistency Boolean flag refers to the binary logic judgment result output by the vehicle consistency verification module (based on the set matching model) after performing the underlying set difference and union operations. It is used to accurately characterize whether the "expected vehicles in the shunting plan at the software level" and the "actual coupled vehicles on site at the physical level" match 100% in terms of identity and quantity.

[0052] Understandably, the error prevention and control mandatory locking command refers to the highest priority active defense safety message generated by the central control system and simultaneously sent to the locomotive on-board terminal and the underlying interlocking equipment when the system determines that there is an anomaly in the on-site vehicle formation (i.e., the above-mentioned consistency Boolean flag error).

[0053] For example, the device status perception subsystem further includes a real-time vehicle feature recognition unit, and the central control system is also equipped with a vehicle consistency verification module. The vehicle consistency verification module uses a vehicle consistency verification algorithm to perform consistency verification, and the implementation steps are as follows: ; The formula parameters for consistency verification using the vehicle consistency verification algorithm are shown in Table 1 below:

[0054] Table 1. Formula parameters for consistency verification using the vehicle consistency verification algorithm

[0055] Furthermore, the equipment status perception subsystem also includes a maintenance warehouse door monitoring unit, which includes a door monitoring camera and an access control controller; The door monitoring camera acquires a video frame stream sequence of the warehouse door area and delineates the region of interest for the door's running trajectory in the video frame stream sequence; The maintenance depot gate monitoring unit extracts the gate outline coordinate features within the region of interest of the gate's running trajectory using edge detection and key point detection algorithms. It then compares the gate outline coordinate features with a preset limit threshold to generate a gate fully open status flag, a gate fully closed status flag, or an abnormal half-open status flag. When the maintenance depot door monitoring unit detects that the door is in motion, it detects moving pixel blocks in the region of interest of the door's running trajectory based on the background subtraction algorithm, and outputs an obstruction signal based on the moving pixel blocks. When the central control system receives an abnormal half-open status flag or an obstruction signal, it triggers the linkage intervention logic to send a stop command to the access controller to stop the current door action, and pushes a prohibition command to the locomotive on-board terminal.

[0056] Understandably, edge detection algorithms refer to a low-level computer vision and image processing technology designed to identify boundary points in digital images where pixel brightness changes drastically.

[0057] Understandably, keypoint detection algorithms refer to a technique used to automatically locate and continuously track specific coordinate points with significant local features in an image. These specific coordinate points with significant local features can be the bottom corner of a door, the gap between two closed doors, or a movable hinge, etc.

[0058] Understandably, background subtraction algorithm refers to a motion detection technique specifically used for segmenting and extracting dynamic targets in video frame sequences.

[0059] The central control system stores shunting plans, which include the expected state set and expected vehicle set of each track-related equipment.

[0060] Understandably, the central control system refers to the core data processing hub of the shunting safety protection system, which is responsible for receiving real-time physical data from underlying multi-source sensing devices.

[0061] Understandably, a shunting coupler plan refers to a set of specific instructions for shunting operations that are pre-compiled and issued or automatically generated by the dispatching and command system (such as the AI ​​intelligent coupler plan automatic generation subsystem). It specifies the standardized process for a single or series of shunting operations, and specifically covers the operation train number, the starting and target tracks, the route taken, the details of the vehicles to be operated, and the safety interlocking prerequisites that must be met for the current operation.

[0062] Understandably, track-related equipment refers to every type of critical infrastructure and safety protection tool that is physically deployed on a specific railway line (track) and its operating route, and whose current physical state directly affects train operation permits and personnel safety. For example, track-related equipment can be derailment devices, intelligent track shoe units, etc. Among them, intelligent track shoe units refer to anti-derailment devices in railways.

[0063] Understandably, the expected state set refers to the list of compliant and safe settings that the aforementioned track-related equipment must be in at specific operational times or stages, based on the currently executing shunting plan.

[0064] Understandably, the expected vehicle set refers to the list of target vehicles that are pre-designated in the current shunting coupler plan and that need to be coupled, uncoupled, towed, or pushed in this operation.

[0065] The central control system is equipped with an interlocking comparison module. The interlocking comparison module extracts a set of expected states and performs consistency condition verification between the received physical state information and the set of expected states to generate a state comparison result. The process of performing consistency condition verification between the received physical state information and the set of expected states to generate a state comparison result includes: the interlocking comparison module receives the visually recognized position state and the mechanical limit state and performs cross-validation to determine the actual state, and performs condition verification between the actual state and the corresponding state in the set of expected states to generate a state comparison result.

[0066] Specifically, the central control system also includes a comprehensive risk level determination module, which is used to receive multiple abnormal indicator variables in the status comparison results output by the interlocking comparison module, and to receive personnel intrusion indicator variables output by the spatial relationship model. The comprehensive risk level determination module internally stores a weight coefficient matrix. The comprehensive risk level assigns a corresponding weight coefficient value to each indicator variable according to the weight coefficient matrix and performs a weighted calculation to obtain a comprehensive risk quantification score. The linkage control command has a hierarchical handling strategy package. The central control system triggers human-machine interface prompts, shunting plan suspension commands, or forced braking commands based on the data range of the comprehensive risk quantification score.

[0067] Understandably, the comprehensive risk level determination module is the central hub for decision-making and safety assessment. It receives in parallel abnormal variables of equipment status from the interlocking comparison module (such as switch not properly engaged, derailment device erroneously raised, track shoes not removed, vehicle identification mismatch, etc.) and personnel intrusion variables output from the spatial relationship model. Its core mechanism lies in using an internally preset risk weight coefficient matrix (e.g., setting personnel intrusion weight to 0.95, switch position error to 0.8, etc.) to mathematically weight and sum the aforementioned concurrent indicator variables, quantifying and integrating multi-dimensional, fragmented physical anomaly data into a dimensionless comprehensive risk quantification score.

[0068] Understandably, the tiered response strategy package refers to a set of step-by-step, structured, automated emergency response mechanisms that are pre-set and stored in the control model to deal with shunting safety threats of different levels of urgency.

[0069] For example, the comprehensive risk level determination module can use a comprehensive risk level determination function to make a determination, which is as follows: ; The parameters in the comprehensive risk level determination function are explained in Table 2 below:

[0070] Table 2. Explanation of parameters in the comprehensive risk level determination function

[0071] The central control system is used to establish a spatial relationship model based on the location information of the operators and the locomotive, generate linkage control commands based on the state comparison results, and send the linkage control commands to the ground signal execution unit to perform safety protection operations. The ground signal execution unit is connected to the locomotive on-board terminal and the central control system respectively.

[0072] Specifically, the spatial relationship model is a three-dimensional spatial coordinate model that includes operators, locomotives, and track trajectories; The central control system is equipped with a distance measuring and early warning module, which calculates the dynamic straight-line distance between the moving locomotive and each operator in real time based on a three-dimensional spatial coordinate model. When the dynamic straight-line distance of the corresponding operator is detected to be less than the preset danger threshold and the coordinates of the positioning terminal are within the closed track area where the shunting operation is about to be carried out, the linkage control command is set to trigger the audible and visual alarm signal of the locomotive on-board terminal and issue a command to prohibit the locomotive from starting to the locomotive on-board terminal.

[0073] This invention utilizes a positioning subsystem comprising a positioning terminal and a positioning module. The positioning terminal collects real-time operator location information and transmits it to the central control system. The positioning module collects real-time locomotive location information and transmits it to the central control system, enabling real-time acquisition of the physical status of key train operation equipment. Furthermore, an equipment status perception subsystem includes status acquisition units deployed at multiple key nodes within the target operation area. These units acquire real-time physical status information of shunting anti-runaway equipment, derailment devices, turnout equipment, and level crossings, transmitting it to the central control system. This eliminates reliance on manual verification and visual inspection. Additionally, the central control system is equipped with an interlocking comparison module, which extracts pre-locking data. The system generates a state comparison result by verifying the consistency between the received physical state information and the expected state set. The central control system is used to establish a spatial relationship model based on the location information of the operators and the locomotive. It generates linkage control commands based on the state comparison results and sends the linkage control commands to the ground signal execution unit to perform safety protection operations. The ground signal execution unit is connected to the locomotive on-board terminal and the central control system. It can realize the dynamic consistency verification between the real-time received physical state of the equipment and the preset expected state set in the shunting coupler plan. Based on the high-precision personnel and locomotive positions, it establishes a dynamic three-dimensional spatial relationship model and automatically generates targeted control commands based on the state comparison results, which can improve the overall safety of the shunting safety protection system.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0075] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0076] The blockchain referred to in this invention is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0077] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0078] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a system claim may also be implemented by a single unit or device through software or hardware. The terms "first," "second," etc., are used to indicate names and do not indicate any specific order.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A shunting safety dynamic verification system based on multi-source sensing and interlocking logic, characterized in that, It includes a central control system, as well as a positioning subsystem, an equipment status perception subsystem, a ground signal execution unit, and a locomotive on-board terminal that are communicatively connected to the central control system. The positioning subsystem includes a positioning terminal and a positioning module. The positioning terminal is used to collect the location information of the operators in real time and send the location information of the operators to the central control system. The positioning module is used to collect the location information of the locomotive in real time and send the location information of the locomotive to the central control system. The equipment status perception subsystem includes status acquisition units deployed at multiple key nodes within the target operating area. These units acquire real-time physical status information of shunting anti-runaway equipment, derailment devices, turnout equipment, and level crossings, and transmit this information to the central control system. The equipment status perception subsystem includes a derailment monitoring unit, which comprises a high-definition camera, physical sensors, and an image analysis unit. The high-definition camera is positioned at the derailment installation point to capture track surface images. The image analysis unit extracts and classifies features from the track surface images based on a pre-loaded computer vision detection model, outputting the visual recognition position status of the derailment. The physical sensors acquire the mechanical limit status of the derailment. The central control system stores a shunting coupler plan, which includes the expected state set and expected vehicle set of each track-related equipment. The central control system is equipped with an interlocking comparison module. The interlocking comparison module extracts the expected state set and performs consistency condition verification between the received physical state information and the expected state set to generate a state comparison result. The process of performing consistency condition verification between the received physical state information and the expected state set to generate a state comparison result includes: the interlocking comparison module receives the visual recognition position state and the mechanical limit state and performs cross-validation to determine the actual state, and performs condition verification between the actual state and the corresponding state in the expected state set to generate a state comparison result. The central control system is used to establish a spatial relationship model based on the location information of the operators and the locomotive, generate linkage control commands based on the state comparison results, and send the linkage control commands to the ground signal execution unit to perform safety protection operations. The ground signal execution unit is connected to the locomotive on-board terminal and the central control system respectively.

2. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The equipment status perception subsystem includes an intelligent iron shoe unit, which includes at least one pressure displacement sensor. The pressure displacement sensor is used to collect iron shoe status information and send the iron shoe status information to the central control system. The iron shoe status information includes track clamping status information and box return status information. The interlocking comparison module retrieves the expected anti-runaway status of the target track in the shunting coupler plan. When the expected anti-runaway status is determined to be a removal condition and the received iron shoe status information is in the on-rail clamping state, the linkage control command is configured to prohibit the ground signal execution unit from opening the shunting signal.

3. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The equipment status sensing subsystem includes a turnout contact monitoring unit, which includes a high-precision position sensor installed in the turnout switch rail area. The high-precision position sensor collects real-time information on the mechanical contact and positioning status of the switch rail and sends it to the central control system. The interlocking comparison module obtains the route opening logic command issued by the central control system and dynamically compares the mechanical tight-fitting status information of the switch rail with the route opening logic command. When there is a difference between the actual mechanical contact position and the position required by the activation command in the comparison results, the linkage control command is set to block the associated path.

4. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The equipment status perception subsystem includes a level crossing linkage unit, which includes a level crossing barrier controller and a video confirmation unit. The central control system acquires the locomotive's location information, calculates the time node for the locomotive to arrive at the target level crossing based on the operation path set in the shunting coupler plan, and sends an automatic closing command to the level crossing barrier controller based on the time node. The video confirmation unit is used to capture and analyze the action of the railing. When it detects that the railing is actually closed in place, it sends a railing positioning command to the central control system. The central control system monitors the barrier positioning command and, upon receiving the command, responds by opening the travel path ahead of the train.

5. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The spatial relationship model is a three-dimensional spatial coordinate model that includes workers, locomotives, and track trajectories; The central control system is equipped with a distance measuring and early warning module, which calculates the dynamic straight-line distance between the moving locomotive and each operator in real time based on a three-dimensional spatial coordinate model. When the dynamic straight-line distance of the corresponding operator is detected to be less than the preset danger threshold and the coordinates of the positioning terminal are within the closed track area where the shunting operation is about to be carried out, the linkage control command is set to trigger the audible and visual alarm signal of the locomotive on-board terminal and issue a command to prohibit the locomotive from starting to the locomotive on-board terminal.

6. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The central control system also includes a comprehensive risk level determination module, which is used to receive multiple abnormal indicator variables from the status comparison results output by the interlocking comparison module, and to receive personnel intrusion indicator variables output by the spatial relationship model. The comprehensive risk level determination module internally stores a weight coefficient matrix. The comprehensive risk level assigns a corresponding weight coefficient value to each indicator variable according to the weight coefficient matrix, and performs a weighted calculation to obtain a comprehensive risk quantification score. The linkage control command has a hierarchical handling strategy package. The central control system triggers human-machine interface prompts, shunting plan suspension commands, or forced braking commands based on the data range of the comprehensive risk quantification score.

7. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The equipment status perception subsystem also includes a real-time vehicle feature recognition unit, and the central control system is also equipped with a vehicle consistency verification module. The expected vehicle set within the shunting coupler plan includes a list of planned vehicle identifiers for the vehicles to be dispatched. The vehicle feature real-time identification unit acquires the actual vehicle set identifier in the target track that is in a physically coupled state, and sends the actual vehicle set identifier to the central control system; The vehicle consistency verification module extracts the actual vehicle set identifiers and the planned vehicle identifier list, performs set difference and set union operations, and extracts the missing vehicle set features and redundant vehicle set features. When the result of merging the missing vehicle set features and the redundant vehicle set features is not empty, the vehicle consistency verification module outputs a consistency Boolean flag indicating that there is a difference. The central control system, based on the consistent Boolean flag with discrepancies, sets the linkage control command to block the opening of subsequent signals and issues a fault-prevention forced locking command to the locomotive on-board terminal.

8. The shunting safety dynamic verification system based on multi-source sensing and interlocking logic as described in claim 1, characterized in that, The equipment status perception subsystem also includes a maintenance warehouse door monitoring unit, which includes a door monitoring camera and an access control controller. The door monitoring camera acquires a video frame stream sequence of the warehouse door area and delineates the region of interest for the door's running trajectory in the video frame stream sequence; The maintenance depot gate monitoring unit extracts the gate outline coordinate features within the region of interest of the gate's running trajectory using edge detection and key point detection algorithms. It then compares the gate outline coordinate features with a preset limit threshold to generate a gate fully open status flag, a gate fully closed status flag, or an abnormal half-open status flag. When the maintenance depot door monitoring unit detects that the door is in motion, it detects moving pixel blocks in the region of interest of the door's running trajectory based on the background subtraction algorithm, and outputs an obstruction signal based on the moving pixel blocks. When the central control system receives an abnormal half-open status flag or an obstruction signal, it triggers the linkage intervention logic to send a stop command to the access controller to stop the current door action, and pushes a prohibition command to the locomotive on-board terminal.

Images